Medical Imaging and Devices: A Practical Reference

An X-ray tube fires photons through the patient onto a detector, and bone attenuates far more than soft tissue because photoelectric absorption rises steeply with atomic number, so the transmitted pattern is a shadow map of what the beam went through. Radiography is one exposure of a few milliseconds at 60–125 kV, read off a flat-panel detector in seconds. Fluoroscopy is the same physics run continuously at a much lower dose rate, usually pulsed at 3–30 frames per second, so an operator can watch a swallow, a joint reduction, or a catheter moving in real time. A two-view chest radiograph delivers roughly 0.1 mSv of effective dose, about 10 days' worth of the 3 mSv a year of natural background radiation an average person in the US receives. Fluoroscopy dose is a rate rather than a number: US rules cap the entrance air kerma rate at about 88 mGy/min in normal mode and 176 mGy/min in high-level control, and a long interventional case can put several gray onto one patch of skin.

This is the cheapest, fastest and most available imaging in medicine, and it accounts for roughly half of all imaging studies performed. The weakness is geometric: a projection image superimposes everything along the ray, so a lung nodule behind the heart or a fracture hidden by overlapping bone can be invisible, and chest radiography misses nodules that CT finds routinely. Fluoroscopy's characteristic failure mode is deterministic skin injury, because dose accumulates on whichever patch of skin the beam enters. Transient erythema starts around 2 Gy of peak skin dose and necrosis becomes possible above 5–10 Gy, and the injury appears weeks after the patient goes home, so the operator gets no feedback during the case. Scattered radiation off the patient is also the main source of the operator's own occupational dose, which is why interventionalists wear lead and carry dosimeters.

Order a radiograph first for anything skeletal, for chest complaints, and for foreign bodies: if the question is "is it broken" or "is there a pneumothorax", this answers it in minutes for tens of dollars. Move to CT or ultrasound when the question is soft tissue detail, a small lesion, or anything the projection will hide behind overlying structures. Use fluoroscopy when you need to see motion or place something, including swallow studies, line and catheter placement, orthopedic reduction and GI contrast studies. Avoid fluoroscopy for any question a single radiograph answers, because dose accumulates by the second rather than by the exposure. If a case is likely to pass roughly 5 Gy of peak skin dose, record the dose and arrange patient follow-up, since that is the level at which professional societies expect it.

Chest radiograph roughly 0.1 mSv against 3 mSv a year of natural background · tube potential 60–125 kV, fluoroscopy pulsed at 3–30 frames per second · entrance air kerma rate capped near 88 mGy/min normal and 176 mGy/min high-level · skin erythema from about 2 Gy, necrosis possible above 5–10 Gy · radiographic room $100–250k, mobile C-arm $50–200k, fixed angiography suite $0.6–2.5M · 40–80 exams a day per radiography room · Medicare pays roughly $30–70 for a two-view chest radiograph depending on setting.

GE Definium and Siemens Ysio radiographic rooms; GE OEC and Ziehm mobile C-arms; Philips Azurion and Siemens Artis fixed angiography suites; the FDA's fluoroscopic entrance-exposure limits in 21 CFR 1020.32; the peak-skin-dose recording thresholds used by the Society of Interventional Radiology and the ACR.

A radiography room costs $100–250k installed, and the replaceable flat-panel detector inside it is $30–90k of that, so a hospital replacing panels every few years is effectively on a subscription. Service contracts run 5–10% of capital a year, cheaper than CT or MRI because there is no tube working anywhere near its thermal limit. Revenue per study is small, roughly $30–70 from Medicare for a two-view chest radiograph across settings, of which under $10 is the radiologist's professional fee, so a room only pays for itself on volume: 40–80 exams a day, five days a week. Fluoroscopy inverts the arithmetic. A fixed angiography suite is $600k–2.5M plus a construction project, and the fluoroscopy code itself pays very little. What justifies the suite is the procedures it enables, since an interventional case bills in the thousands and the hospital captures the supplies, the room time and the professional fee together. The vendors are the same four that sell CT and MRI, and they price the imaging chain as part of a service relationship rather than as a box.

6 minHow do X-Rays Work?Concerning Reality · 100k+ views

13 minHow to set-up and correctly use a C-arm: Fluoroscopy safety and proceduresClover Learning · 10k+ views

Radiation Dose from X-Ray and CT Exams (RadiologyInfo.org) · Fluoroscopy (FDA)

Every X-ray system in this class is a tube and a detector with a patient in between. The tube boils electrons off a heated tungsten filament and accelerates them across 40–150 kV into a tungsten-rhenium anode, where under 1% of the beam energy becomes X-rays and the rest becomes heat. Heat is the whole engineering problem: the anode is a disc spinning at 3,000–10,000 rpm so the electron beam paints a track rather than a spot, and CT tubes store 3–8 million heat units and are cooled through the anode directly in some designs. The focal spot is 0.3–1.2 mm, and making it smaller sharpens the image while concentrating the same heat into less metal. On the other side, a modern flat-panel detector is either indirect (a columnar cesium iodide scintillator on an amorphous silicon thin-film transistor array, 100–200 µm pixels) or direct (an amorphous selenium photoconductor, 70–100 µm pixels, used where resolution matters most). Detective quantum efficiency is the figure of merit, and cesium iodide panels reach roughly 60–70% at low spatial frequency against roughly 25–35% for the storage-phosphor plates they replaced.

A tube is still the only practical way to make a bright, energy-tunable X-ray beam from wall power, and it is cheap relative to what it enables. Its weakness is that it is a consumable. A hard-working CT tube lasts roughly one to two years, and replacement runs $50–200k, which is usually the largest single line item inside a scanner's service contract; the failure modes are filament evaporation, bearing wear in vacuum where there is no ordinary lubricant, and anode cracking from thermal cycling. Detectors last much longer but a flat panel is fragile in portable use, and a replacement is $30–90k. Direct-conversion selenium gives excellent resolution and almost no light spreading, but selenium stops high-energy photons poorly, which is why it dominates mammography at 25–35 kV and is absent from general radiography at 120 kV.

This is the component layer under every other entry in this class, so the decisions here are purchasing decisions. If you are specifying a general radiography room, buy indirect cesium iodide panels and expect detective quantum efficiency around 60–70%; if you are specifying mammography, you will get amorphous selenium whether you ask for it or not. If you still run computed radiography plates, convert now, because Medicare has cut the technical payment for computed radiography studies by 7% since 2018 and 10% since 2023, and film by 20% since 2017, so the payment system is funding the conversion for you. For CT, negotiate tubes into the service contract rather than buying them separately, since a scanner running 16 hours a day will consume several over its life. If you are sourcing components rather than systems, expect a short supplier list and long lead times.

40–150 kV tube potential, under 1% of input energy converted to X-rays · anode 3,000–10,000 rpm, focal spot 0.3–1.2 mm · CT anode heat storage 3–8 million heat units · flat-panel pixels 100–200 µm indirect, 70–100 µm direct · detective quantum efficiency roughly 60–70% for cesium iodide panels against 25–35% for storage phosphor · CT tube replacement $50–200k, flat panel $30–90k · Medicare technical payment cut 7% for computed radiography from 2018 and 10% from 2023, 20% for film from 2017.

Varex Imaging, the largest merchant supplier of tubes and detectors at roughly $800M of annual revenue; Canon Electron Tubes & Devices; Dunlee, the Philips tube business; the Siemens Straton rotating-envelope CT tube, which cools the anode directly instead of radiating through vacuum; Trixell, the Thales, Siemens and Philips flat-panel joint venture; amorphous selenium mammography detectors from Analogic Canada and Hologic.

The tube and the detector are the two highest-value parts of any X-ray system and together often account for 30–50% of the bill of materials. Tubes are a consumable business: a CT tube sells for $50–200k and a busy scanner needs one every year or two, so across a 10–15 year service life the aftermarket is worth more than the original tube sale. That is a large part of why manufacturers write full-service contracts at 8–12% of capital a year with the tube included, and why the independent service market stays small. Detector panels work the other way. They usually last most of the system's life, so the margin is in the original sale and in the yield of large-area thin-film transistor manufacturing, which is a display-industry process run at tiny volume, and yield on a 43 cm panel is what sets the price. Varex is the one large independent supplier and sells into a market where its OEM customers keep bringing components in-house, so its volume tracks how much of the market the big four choose not to make themselves. For a startup, the practical consequence is that the components are available to buy but the cost curve belongs to someone else.

9 minOverview of the X-Ray Tube and ComponentsClover Learning · 100k+ views

8 minThe Anode | X-ray Machine | X-ray physics #3 | Radiology Physics Course #10Radiology Tutorials · 50k+ views

X-Ray Tube Heating and Cooling (Sprawls Resources) · Diagnostic Radiology Physics: A Handbook for Teachers and Students (IAEA)

A CT scanner spins an X-ray tube and an arc of detectors around the patient on a slip-ring gantry, typically one rotation every 0.25–0.5 seconds, while the table moves through the bore. Thousands of projections are reconstructed into cross-sections, historically by filtered back projection and now usually by iterative or deep-learning reconstruction, giving 0.4–0.6 mm in-plane resolution and slices down to 0.5 mm. Modern scanners carry 64–320 detector rows covering 40–160 mm per rotation, so a chest scan takes a few seconds and a whole-body trauma scan under a minute. Most body imaging uses intravenous iodinated contrast, which is why the invasiveness here runs from external to injected. Dose is measured on the machine as CTDIvol in mGy and dose-length product in mGy·cm, then converted to effective dose with a body-region factor; a chest CT comes out around 6.1 mSv and an abdomen-and-pelvis CT around 7.7 mSv, against roughly 3 mSv a year of natural background and about 0.1 mSv for a chest radiograph.

CT is fast, available around the clock, and answers most acute clinical questions in one pass, which is why it became the backbone of emergency medicine. The cost is radiation. Roughly 93 million CT examinations were performed in the US in 2023, and a 2025 modeling study projected on the order of 100,000 future cancers from that single year of scanning; the estimate is contested, but the scale of population exposure is not. CT also separates tissue only by electron density, so its soft-tissue contrast is far worse than MRI: it cannot distinguish gray from white matter usefully, or characterize a soft-tissue tumor the way MRI can. Dense metal and heavy calcification cause beam-hardening and streak artifacts that can make an implant or a calcified coronary segment unreadable.

Make CT the default for trauma, acute abdominal pain, suspected pulmonary embolism, stroke triage, and staging most solid tumors, because speed and availability matter more than tissue contrast in all of those. Pick MRI when the question is soft tissue, bone marrow, or the brain beyond ruling out hemorrhage, and pick ultrasound first in children and in pregnancy, where the dose argument is strongest. Before ordering a CT in a young patient, ask whether the same study will be repeated ten times over the next decade, because the cumulative dose is the number that matters rather than the single exam. For screening, only low-dose CT in high-risk smokers has both outcome evidence and a reimbursement pathway, and that protocol runs about 1.5 mSv rather than the 6.1 mSv of a diagnostic chest CT.

0.25–0.5 s gantry rotation, 64–320 detector rows, 40–160 mm coverage per rotation · 0.4–0.6 mm in-plane resolution, 0.5 mm slices · chest CT 6.1 mSv, abdomen and pelvis 7.7 mSv, low-dose lung screening 1.5 mSv, chest radiograph 0.1 mSv, natural background 3 mSv a year · roughly 93 million US CT examinations in 2023 · $80–150k refurbished 16-slice to above $3M list for a premium system, plus $200–500k of siting · service contract 8–12% of capital a year · 20–40 studies a day.

GE Revolution, Siemens Somatom Force and X.cite, Philips Incisive CT, Canon Aquilion ONE; the National Lung Screening Trial and the low-dose CT screening program built on it; the ACR Dose Index Registry, which lets a hospital compare its own CTDIvol against national distributions.

The price range is wide enough that "a CT scanner" means very little on its own: a refurbished 16-slice sells for $80–150k, a new mid-range 64-slice for roughly $400–800k, and a premium wide-detector or dual-source system lists above $3M before the discount every buyer negotiates. Siting adds $200–500k, since the room needs lead shielding, three-phase power, chilled air and a floor rated for the gantry. Service contracts run 8–12% of capital a year and mostly buy tubes. Revenue arrives per study: Medicare pays roughly $250–350 for a contrast-enhanced abdomen-and-pelvis CT in a hospital outpatient department, of which roughly $60–80 is the radiologist's professional component, and a freestanding imaging center is paid less for the same scan. Run the arithmetic and the business model is obvious: 25 studies a day, 300 days a year, at $300 a study is about $2.25M of technical revenue, so utilization drives the purchase far more than image quality does. That is also why the second scanner in a hospital is a much harder sale than the first, and why a new entrant selling image quality alone has no argument against an installed base of tens of thousands of adequate scanners.

1 minHow Does a CT Scan Work?NIBIB · 1m+ views

20 minCT physics overview | Computed Tomography Physics Course | Radiology Physics Course Lesson #1Radiology Tutorials · 50k+ views

The Measurement, Reporting, and Management of Radiation Dose in CT (AAPM Report No. 96) · Projected Lifetime Cancer Risks From Current Computed Tomography Imaging (JAMA Internal Medicine)

A conventional CT detector integrates energy: a ceramic scintillator turns X-rays into light, a photodiode turns the light into current, and every photon arriving during a reading is summed together with the high-energy ones counted more heavily than they should be. A photon-counting detector replaces that chain with a single thick semiconductor, cadmium telluride or cadmium zinc telluride roughly 1.6–2 mm deep under a bias field of several hundred volts per millimeter, which converts each X-ray directly into a charge pulse. Fast pulse-processing electronics count each pulse individually and sort it into 2–8 energy bins. Three things follow from that. Electronic noise falls below the counting threshold and drops out of the image entirely, which is what helps most in low-dose and large-patient scanning. The detector needs no reflective septa between elements, so pixels can be subdivided far more finely, reaching about 0.15 mm at isocenter against roughly 0.5 mm on a conventional scanner. And every scan carries spectral information without a second tube, a second rotation, or a decision made before the patient arrives.

What that buys clinically is specific rather than general: ultra-high-resolution scanning that resolves the inside of a coronary stent and fine lung interstitium, iodine contrast-to-noise improvement of roughly 20–40% that a site can spend either as less contrast agent or as less dose, and virtual monoenergetic and material-decomposition images from a routine acquisition, which help most where calcium blooming or metal artifact currently ruins the study. The limits are physical and industrial. At clinical flux a pixel sees up to about 10⁹ photons per mm² per second, and overlapping pulses distort the energy measurement, so pixels have to be both small and fast, and charge shared between neighboring pixels degrades exactly the energy resolution the design exists to provide. Large-area, defect-free cadmium telluride and cadmium zinc telluride crystals grow slowly at low yield, and that crystal supply is the main reason the scanner costs what it does.

Buy one if you are a high-volume cardiac or thoracic center where resolution and spectral data change decisions: heavily calcified coronaries, stent-in-stent restenosis, interstitial lung disease, and patients in whom contrast volume or dose has to be minimized. Do not buy one as a general-purpose replacement in a community hospital, because the same CPT code pays the same amount for a photon-counting scan as for a conventional one and the extra $1M does not come back on volume. If your constraint is throughput, two conventional scanners beat one photon-counting scanner every time. If your constraint is one specific spectral question, a dual-energy conventional scanner does much of the same work for a fraction of the premium. Treat any purchase now as an early-adopter position: the technology works, and the evidence that it changes outcomes rather than images is still being collected.

Cadmium telluride or cadmium zinc telluride 1.6–2 mm thick, several hundred volts per millimeter bias · 2–8 energy bins, no electronic noise above threshold · about 0.15 mm resolution at isocenter against roughly 0.5 mm conventional · iodine contrast-to-noise up roughly 20–40% · flux up to about 10⁹ photons per mm² per second before pile-up · list price roughly $3–4M, a premium of about 30–50% over a premium conventional scanner · installed base in the hundreds worldwide against tens of thousands of conventional scanners · no separate CPT code.

Siemens NAEOTOM Alpha, the first cleared system, which went through the FDA as a 510(k) in September 2021 (K211591) under the same product code as any other CT scanner, followed by the Alpha.Peak, .Pro and .Prime clearances in February 2025 (K243523); Canon's 2021 acquisition of Redlen Technologies, a cadmium zinc telluride crystal house; GE's deep-silicon detector program built on its purchase of Prismatic Sensors; Kromek and Acrorad as the other significant crystal suppliers.

The regulatory story here is the good news and the problem at once. Photon-counting CT cleared as a Class II device through 510(k) against conventional CT as the predicate, which cost months rather than the years a premarket approval would have taken, and no new CPT code came with it. A photon-counting scan bills exactly like any other CT scan, so the whole premium has to be recovered from referral capture, from contrast and dose savings, or from studies a conventional scanner cannot do at all. Against a list price of roughly $3–4M and a premium conventional scanner at $2–2.5M, that is $1M or more to justify on soft revenue, and a hospital capital committee will ask for the referral argument in writing. The cost curve runs through the crystals: yield on large defect-free cadmium telluride wafers is the binding constraint, which is why Canon bought Redlen outright and why GE went after silicon instead, where the material is cheap and the physics is harder. Expect the installed base, currently in the hundreds worldwide, to grow with crystal yield rather than with clinical evidence.

3 minPhoton-counting CT explained - part 1Siemens Healthineers · 50k+ views

7 minWhat's so great about photon-counting CT?Aunt Minnie · 1k+ views

Technical Basics and Clinical Benefits of Photon-Counting CT (Investigative Radiology) · 510(k) Premarket Notification K211591, NAEOTOM Alpha (FDA)

A mammography unit is an X-ray system tuned for one tissue. It runs a tungsten or molybdenum anode at 25–35 kV through a rhodium or silver filter, compresses the breast to 4–6 cm with a paddle to cut thickness, scatter and motion together, and reads the beam on an amorphous selenium direct-conversion detector with 70–100 µm pixels. Two-dimensional full-field digital mammography takes two views of each breast. Digital breast tomosynthesis sweeps the tube through an arc of roughly 15–50 degrees, takes 9–25 low-dose projections along the way, and reconstructs a stack of 1 mm slices, which pulls apart the overlapping fibroglandular tissue that hides a real mass or fakes one on a flat projection. Dose is regulated more tightly here than anywhere else in imaging: the Mammography Quality Standards Act caps mean glandular dose at 3 mGy per view for a standard 4.2 cm compressed breast, and a four-view screening exam works out to about 0.28 mSv of effective dose in 2D and 0.34 mSv with tomosynthesis, against roughly 3 mSv a year of natural background.

Screening mammography is the only breast imaging with randomized mortality evidence behind it, which is why it is the population test and why everything else is supplemental. Tomosynthesis improves on it by a real but modest margin in community practice: recall rates fall by roughly 1–2 percentage points and cancer detection rises by roughly 1 per 1,000 screens, with most of the benefit in the first screening round and in women with heterogeneously dense breasts. It does much less in extremely dense breasts, where cancer and fibroglandular tissue attenuate similarly from every angle, so no amount of geometry separates them. Reading time roughly doubles with tomosynthesis, which is a genuine operational cost when a radiologist reads a hundred screens in a session. The dense-breast gap is closed with ultrasound or MRI rather than with a better mammogram.

Screen with mammography as the default, following the USPSTF recommendation of biennial screening from age 40 to 74, and use tomosynthesis as the acquisition wherever you have it, since it costs the patient a small extra dose and lowers the chance of an unnecessary callback. Add supplemental ultrasound or MRI for women with dense breasts or high lifetime risk instead of expecting more from the mammogram itself. If a woman has a palpable lump or a nipple discharge, order diagnostic mammography plus ultrasound rather than a screening exam, because the codes, the workflow and the evidence standards are all different. If you are building a device and want a screening claim, budget for outcome data in an asymptomatic population, which is the most expensive trial design in medicine; if you can live with a diagnostic claim, a reader study is a far cheaper path.

25–35 kV, amorphous selenium detector at 70–100 µm pixels, breast compressed to 4–6 cm · tomosynthesis sweeps 15–50 degrees in 9–25 projections and reconstructs 1 mm slices · MQSA caps mean glandular dose at 3 mGy per view · 0.28 mSv per screening exam in 2D and 0.34 mSv with tomosynthesis against 3 mSv a year of background · recall down roughly 1–2 percentage points, detection up roughly 1 per 1,000 · 9,118 certified US facilities and 27,692 accredited units, 13,515 of them tomosynthesis · roughly 45 million mammograms a year in the US · 2D unit $80–150k, tomosynthesis-capable unit $250–400k.

Hologic Selenia Dimensions and 3Dimensions, approved by premarket approval in February 2011 (P080003); GE SenoClaire (P130020, August 2014); Siemens Mammomat Revelation and B.brilliant (P140011); Fujifilm Aspire Cristalle (P160031, January 2017); the ECOG-ACRIN TMIST randomized trial; the FDA rule requiring breast density notification, effective September 2024.

The commercial split that matters here is screening versus diagnosis. A diagnostic claim can be supported by a multi-reader study on a few hundred cases; a screening claim in asymptomatic women needs outcome data, which is the most expensive kind of trial in medicine. FDA took a middle path with tomosynthesis, which is why the adoption path is worth studying closely. Hologic's Selenia Dimensions was approved as a Class III device by premarket approval in February 2011 on a reader study rather than a mortality trial, and because tomosynthesis is classified Class III, every competitor had to file its own PMA instead of a 510(k): GE's SenoClaire followed in August 2014 and Fujifilm's Aspire Cristalle in January 2017, giving Hologic roughly a three-year exclusive window that a predicate-based pathway would never have allowed. Then CMS created an add-on code for screening tomosynthesis worth roughly $50 on top of the roughly $140 it pays for a bilateral screening mammogram, and the purchase arithmetic became easy: a unit doing 25 screens a day for 250 days is 6,250 exams a year, so the add-on alone is about $310k a year against a $150–250k premium for the tomosynthesis-capable machine. Adoption followed the code, and 8,638 of 9,118 US certified facilities now have at least one tomosynthesis unit. TMIST, the randomized trial that would actually settle whether tomosynthesis reduces advanced cancers, enrolled roughly 129,000 women and only finished accrual in 2025, more than a decade after the technology became standard practice.

2 minTomosynthesis: New Breast Cancer Screening - Mayo ClinicMayo Clinic · 50k+ views

15 minTomosynthesis (Does it offer 3D Mammography)How Radiology Works · 10k+ views

Recommendation: Breast Cancer: Screening (United States Preventive Services Taskforce) · Screening Performance of Digital Breast Tomosynthesis vs Digital Mammography in Community Practice by Patient Age, Screening Round, and Breast Density (JAMA Network Open)

Cone-beam CT replaces the fan beam and detector arc of a medical CT with a cone-shaped beam and a single flat-panel detector, so one rotation of 180–360 degrees captures the whole volume in 150–600 projections and reconstructs it with the Feldkamp algorithm. The payoff is isotropic voxels, 75–200 µm in dental units and 0.2–0.5 mm in interventional and radiotherapy systems, where a medical CT's in-plane resolution is much finer than its slice thickness. The price is scatter. With a wide cone angle the scatter-to-primary ratio at the detector can exceed 1, which flattens low-contrast detectability and makes the reconstructed values unreliable, so cone-beam numbers are not true Hounsfield units and cannot be used for tissue characterization or for radiotherapy dose calculation without correction. Dose sits between plain radiography and CT: a dental cone-beam scan is roughly 0.18 mSv, about seven times a panoramic radiograph at 0.025 mSv and roughly a tenth of a head CT at 1.6 mSv.

Because the hardware is just a tube and a flat panel on a rotating arm, cone-beam CT can be built into places a CT gantry cannot go: a chairside dental unit, a C-arm in an angiography suite, a linear accelerator gantry, a mobile ring in an operating room, and a standing extremity scanner that images a joint under load. Bone, teeth and metal hardware all image well, and the geometry is accurate enough to plan an implant or register a navigation system. The weakness is soft tissue: low-contrast detectability is roughly an order of magnitude worse than diagnostic CT because of scatter and detector lag, so this is a geometry instrument rather than a tissue instrument. Cone-beam artifacts also worsen away from the central plane, and acquisitions of 5–30 seconds make the image sensitive to patient motion in a way a sub-second CT is not.

Reach for cone-beam CT when you need three-dimensional geometry at the point of care and the question is about bone, hardware or position. Planning a dental implant, treating a tooth with suspected extra canals, or assessing sinus anatomy are all good fits. If you are positioning a patient on a linear accelerator before each fraction, cone-beam CT is already built into the machine and there is no decision to make. Order a diagnostic CT instead whenever the question is soft tissue, an abscess, a tumor margin, or anything that needs reliable Hounsfield units. Do not use dental cone-beam CT as a routine screening exam: ADA and AAOMR guidance is explicit that it should be selected for a specific clinical question, because the dose is several times a panoramic view's and the field of view often covers structures nobody intended to read.

150–600 projections in one 180–360 degree rotation, reconstructed by Feldkamp · isotropic voxels 75–200 µm dental, 0.2–0.5 mm interventional and radiotherapy · scatter-to-primary ratio can exceed 1, so values are not true Hounsfield units · dental scan roughly 0.18 mSv against 0.025 mSv panoramic and 1.6 mSv head CT · acquisition 5–30 seconds · dental unit $50–120k, extremity scanner $300–500k, C-arm option on a $1–2.5M angiography suite · dental scan billed at roughly $150–400.

Planmeca ProMax and Carestream CS 9600 dental units; Medtronic O-arm for spine navigation; syngo DynaCT on Siemens Artis and XperCT on Philips Azurion in interventional suites; kV cone-beam imaging built into Varian and Elekta linear accelerators for daily patient positioning; Planmed Verity and CurveBeam weight-bearing extremity scanners; the ADA and AAOMR patient selection recommendations for dental radiography and cone-beam CT.

The most instructive thing about cone-beam CT is who signs the purchase order. A dental unit costs $50–120k and is bought by a single practice out of operating cash flow, with no capital committee and no construction project, which makes it the only tomographic imaging sold that way. Dental scans are billed to dental plans or paid by the patient at roughly $150–400, so the payback is direct arithmetic: a $90k unit at $250 a scan needs about 360 scans to cover the box, and a practice doing three a week bills $36k a year, which clears the machine in about two and a half years. That self-funded path is why the dental installed base grew faster than any hospital modality. The hospital side works the opposite way. Cone-beam capability on an interventional C-arm is a detector-and-software option on a $1–2.5M angiography suite, and kV cone-beam imaging on a linear accelerator is standard equipment with no separate payment at all, since image guidance is bundled into the radiotherapy episode. Extremity and weight-bearing scanners at $300–500k sit in between and have the best margin story of the three: they bill under the same CT codes as a full scanner while costing a fifth as much, which is a real advantage for an orthopedic practice that owns its own imaging.

4 minDifference Between CT and CBCT |CT and CBCT basic understanding| Fan beam vs Cone beam|Dento Radiology · 50k+ views

3 minHow CBCT Helps Endodontists Expose the Source of Tooth PainAmerican Association of Endodontists · 50k+ views

Dental Cone-beam Computed Tomography (FDA) · Mobile C-arm cone-beam CT for guidance of spine surgery: image quality, radiation dose, and integration with interventional guidance (Medical Physics)

An MRI scanner puts the patient inside a strong static magnetic field, tips the hydrogen nuclei in tissue with a radiofrequency pulse, and listens to the signal they give off as they realign. The field comes from a niobium-titanium solenoid cooled to about 4 K, where the wire carries current with no resistance, so once the magnet is ramped the current circulates in a closed loop with the power supply disconnected. Clinical systems run at 1.5 T or 3 T with a 60 cm or 70 cm patient bore, and the magnet alone weighs 3 to 6 metric tons. Signal rises roughly linearly with field strength, which is why 3 T became standard for neuro and musculoskeletal work and why 7 T systems are cleared in the US for head and knee imaging only. Many exams add an injected gadolinium contrast agent, which shortens T1 locally and makes vascular and enhancing tissue stand out. If any part of the winding goes normal, the stored energy dumps into the coil and boils the helium, and the resulting quench vents several thousand cubic meters of cold gas that has to leave the building through a dedicated pipe.

MRI gives the best soft-tissue contrast of any modality and uses no ionizing radiation, so it can be repeated on the same patient without a dose budget, and changing the pulse sequence changes the contrast mechanism without changing the hardware. The weaknesses are cost, throughput, and infrastructure. A scan takes 20 to 45 minutes against about 10 seconds for CT, so a busy scanner does 10 to 12 patients a day rather than 40. The magnet is always on, which means a permanent 5-gauss safety perimeter, a ferrous-object screening process, and a real injury mechanism when screening fails. Older magnets hold 1,500 to 2,000 liters of liquid helium and lose some to boil-off, and MRI accounts for roughly 17% of US helium consumption, so a supply squeeze shows up as a cryogen bill rather than a shortage of scanners.

Buy 1.5 T as the default: it covers nearly every clinical indication, it is more forgiving of implants and metal artifact, and it costs roughly a third less than 3 T. Go to 3 T if your volume is weighted toward neuro, prostate, or research spectroscopy, where the extra signal buys either resolution or scan time. If your building cannot take a 6-metric-ton magnet, a quench pipe to the roof, and a copper-lined RF room, price the construction before the scanner, because siting often costs as much as the magnet. Choose a sealed low-helium magnet if you are in a region where cryogen delivery is unreliable or expensive, and accept that the magnet cannot then be serviced by topping it up. If the question is bone detail, acute trauma, or anything that has to be answered in under a minute, use CT instead.

1.5 T and 3 T clinical standard, 7 T cleared for head and knee · magnet at about 4 K weighing 3–6 metric tons · bore 60 or 70 cm · conventional magnets hold 1,500–2,000 liters of liquid helium, sealed designs under 10 · MRI is roughly 17% of US helium consumption · 20–45 minute exams, 10–12 patients per scanner per day · scanner $1–3M with siting often the same again.

Siemens Magnetom, GE Signa and Philips Ingenia platforms cover most of the installed base; Philips BlueSeal ships a sealed magnet holding under 10 liters of helium instead of 1,500; Siemens Magnetom Terra was the first 7 T system cleared in the US, for head and knee imaging.

A 1.5 T scanner lists around $1M and a 3 T around $2–3M, but the project cost is what matters: RF shielding, a quench vent to outside air, floor reinforcement for a 6-metric-ton magnet, a chiller loop, and the construction to put all of that in an existing building routinely add as much again. Service contracts run about 8–12% of capital per year, so a $2M scanner carries a $160–240k annual line before staff. Revenue comes from volume against a fixed cost base: Medicare pays a few hundred dollars for a typical outpatient MRI, so a scanner needs roughly 10 to 12 exams a day to work, and utilization rather than image quality decides whether the room pays. That is why hospitals extend hours rather than buy a second magnet, and why an imaging center's business case is mostly a scheduling argument. Helium is a small operating line for a modern zero-boil-off magnet and a large one-time cost if a magnet quenches and has to be refilled and re-ramped, which is the case sealed magnets are sold against. The installed base is enormous and slow to turn over, magnets last 15 to 20 years, and most competition happens on the console and the coils rather than on the magnet.

18 minThe Insane Engineering of MRI MachinesReal Engineering · 1m+ views

4 minWhat Is an MRI Quench?Block Imaging · 10k+ views

A Narrative Review of Advancements in Magnetic Resonance Imaging (MRI) Technology: Evaluating the Shift From Helium-Cooled to Helium-Free Systems (Cureus) · Helium: Mineral Commodity Summaries 2025 (U.S. Geological Survey)

Low-field MRI drops the static field from 1.5 T to somewhere between 0.55 T and 64 mT, which lets the magnet be a permanent neodymium array or a lightly cooled superconductor instead of a helium bath. Signal falls steeply with field, roughly as the square of the field strength for a fixed acquisition, so a 64 mT scanner starts with a small fraction of the signal a 1.5 T scanner has and buys it back with long averaging, thick slices, and trained reconstruction. The payoff is siting: a point-of-care head scanner such as the Hyperfine Swoop weighs about 630 kg, runs off a standard wall outlet, and has a 5-gauss line inside about a meter, so it rolls to the bedside with no shielded room and no quench pipe. At 64 mT a brain exam typically acquires about 1.5 mm in plane with 5 mm slices, against roughly 1 mm isotropic at 3 T. The 0.55 T tier is a different product: a whole-body scanner with an 80 cm bore holding under a liter of helium, aimed at the standard MRI market rather than at the bedside.

The physics that costs signal also buys things back. Susceptibility artifact scales with field, so low field images better next to metal implants and in the lung, T1 contrast is intrinsically better, and specific absorption rate is far lower, which loosens the restrictions on scanning patients with implanted hardware. The weakness is what the images can answer. At 64 mT the scanner reliably shows ventricle size, midline shift, hydrocephalus, large intracranial hemorrhage, and large territorial infarcts, and it does not reliably show small infarcts, microbleeds, small metastases, or posterior fossa detail. Deep learning reconstruction improves how the images look faster than it improves what they detect, which is the honest caution on every published comparison.

Use a 64 mT point-of-care scanner where the alternative is no MRI at all: an ICU patient too unstable to move, a stroke unit that needs a repeat look at a known bleed, a field hospital, or a low-resource setting with no shielded room. Do not use it to rule out a small posterior fossa lesion or an early small infarct, because a negative scan does not carry that weight. Choose 0.55 T instead if you want a general-purpose scanner and the constraint is construction cost or ceiling height, since it covers most body imaging and needs neither a quench pipe nor heavy shielding. If the patient can travel to a conventional scanner and the question is diagnostic rather than triage, send them to 1.5 T. Before buying, confirm with your payers that a low-field exam bills under the existing MRI codes, because there is no separate code for it and no separate payment.

Field 64 mT to 0.55 T against 1.5–3 T conventional · signal falls roughly with the square of field strength for a fixed acquisition · about 1.5 mm in plane and 5 mm slices at 64 mT, against roughly 1 mm isotropic at 3 T · point-of-care unit about 630 kg on a standard wall outlet, 5-gauss line inside about a meter · 0.55 T whole-body with an 80 cm bore and under a liter of helium · roughly $250k for a head-only unit and $0.5–1M for 0.55 T whole-body.

Hyperfine Swoop, a 64 mT bedside brain scanner FDA-cleared since 2020 and deployed in ICUs, stroke units and research programs in Uganda and Malawi; Siemens Magnetom Free.Max at 0.55 T with an 80 cm bore and a sealed sub-liter helium magnet; the Open Source Imaging Initiative's OSI² ONE, a published open-hardware low-field scanner.

The cost argument is siting, not the magnet. A conventional MRI project pays for a copper-lined room, a quench vent, floor reinforcement and a chiller, and a scanner that needs none of those can be installed in an existing room in days. A point-of-care head scanner runs roughly $250k against $1–3M for a conventional system, and a 0.55 T whole-body scanner roughly $0.5–1M, so the capital committee sees a different kind of decision. What is unsettled is revenue. There is no distinct CPT code for a low-field exam, so it bills as an MRI of the same body part or it does not bill at all, and a hospital buying one on a triage argument is buying avoided transport and avoided CT rather than new billing. That makes the strongest early business cases the ones where moving the patient is expensive or dangerous, and the weakest the ones that assume outpatient imaging volume will migrate down in field strength. Expect the 0.55 T tier to grow faster commercially than the 64 mT tier, because it sells into an existing reimbursed workflow.

12 minOSI² ONE: An Open Source Low-Field magnetic resonance imaging (MRI) ScannerOpen Source Imaging Initiative (OSI²) · 1k+ views

13 minMatthew Rosen: Building an MRI Scanner 60 Times Cheaper & Small Enough to Fit in an AmbulanceLDV Capital · 1k+ views

Low-field MRI: Clinical promise and challenges (Journal of Magnetic Resonance Imaging) · Design, Characterisation and Performance of an Improved Portable and Sustainable Low-Field MRI System (Frontiers in Physics)

The magnet supplies the field; everything that turns it into an image sits inside the bore. Three orthogonal gradient coils add a linear field variation along x, y and z so that resonant frequency encodes position, and a clinical system runs about 33–45 mT/m of gradient amplitude with a slew rate of 120–200 T/m/s, driven by amplifiers pushing hundreds of amps at peak powers approaching a megawatt. A body-sized transmit coil delivers the RF pulse, and a close-fitting receive array picks up the signal: 8 to 32 channels for routine work, 64 or more on research head arrays. Channel count matters because independent coil sensitivities let parallel imaging reconstruct an image from undersampled data, so 2–4x acceleration is routine and simultaneous multi-slice pushes diffusion and functional imaging further. The pulse sequence decides what the image means, since the same hardware produces a T1, T2, diffusion or susceptibility image depending only on the timing of the pulses and gradients.

Sequences and reconstruction are where most image-quality gains now come from, because they are software on an installed magnet rather than a new magnet. Compressed sensing and trained reconstruction have cut typical scan times by roughly a third to a half at matched image quality, and vendors ship them as licensed options on scanners already in the field. Gradient performance is limited by the patient, not the amplifier: rapidly switched gradients induce currents in peripheral nerves, and IEC 60601-2-33 requires operation below the peripheral nerve stimulation threshold, which is what caps usable slew rate near 200 T/m/s on a whole-body system. Local head-only gradient inserts sidestep that geometry and reach 200–500 mT/m in research use. Gradient switching also drives Lorentz forces in the coil that make MRI loud, over 100 dB in some sequences, and receive coils are the part patients and staff physically handle, so they get dropped and are a recurring repair line.

Specify gradients by what you actually run: 33–45 mT/m is enough for routine clinical work, and paying for the top gradient tier is worth it only if diffusion imaging, functional imaging, or cardiac work is a real share of your volume. Buy more receive channels if you want acceleration rather than raw sensitivity, since parallel imaging capability scales with independent coils. If throughput is the problem, license the accelerated reconstruction package before you consider a second scanner, because a 40% scan-time cut on an existing magnet is far cheaper than another room. If you are a startup selling into this layer, sell software or coils, since both install on the existing base without touching the magnet. Avoid assuming a research gradient insert transfers to clinical use, because the nerve stimulation limit that bounds a whole-body coil is a regulatory constraint and not an engineering one.

Gradient amplitude 33–45 mT/m clinical, 200–500 mT/m on research head inserts · slew rate 120–200 T/m/s, capped by peripheral nerve stimulation under IEC 60601-2-33 · gradient amplifiers at hundreds of amps and peak powers near a megawatt · receive arrays of 8–32 channels routine, 64+ for research heads · parallel imaging acceleration 2–4x routine · accelerated reconstruction cuts scan time by roughly a third to a half · acoustic noise above 100 dB.

Siemens Deep Resolve, GE AIR Recon DL and Philips SmartSpeed are the shipping deep learning reconstruction options; blipped-CAIPI simultaneous multi-slice, developed for the Human Connectome Project, is now standard on clinical diffusion and functional protocols; head gradient inserts built for connectome imaging reach several hundred mT/m.

This layer is where MRI vendors make margin after the magnet is sold. A 32-channel receive array costs roughly $30–80k, coils get dropped and cables fail, so replacement is a recurring line rather than a one-time purchase. Sequence and reconstruction packages are licensed software options priced from tens of thousands to a few hundred thousand dollars each, and they carry near-zero marginal cost, which is why every vendor's roadmap is weighted toward them. For the hospital the arithmetic is simple: if a package cuts a 30-minute exam to 20, a scanner running 10 hours a day gains several slots a day, and at a few hundred dollars of reimbursement per exam the option pays back inside a year or two at high utilization. Gradient upgrades on an installed magnet are also sold this way, though they need a service visit and sometimes a new amplifier cabinet. The barrier for a third-party entrant is that coils and sequences have to be validated against a specific vendor's scanner and cleared as accessories to it, which is a narrow 510(k) path but a real one, and a few independent coil makers do exactly that.

15 minMRI Machine - Main, Gradient and RF Coils/ Magnets | MRI Physics Course | Radiology Physics Course#2Radiology Tutorials · 100k+ views

7 minPhase encoding helps localize an MRI signal in the body - MRI physics explainedClover Learning · 10k+ views

Gradient specifications (Questions and Answers in MRI) · Recent advances in parallel imaging for MRI (Progress in Nuclear Magnetic Resonance Spectroscopy)

MR-guided therapy puts a treatment device inside an MRI scanner so the target can be seen while it is being treated, rather than only before and after. The largest version is the MR-linac, which combines a whole-body magnet with a megavoltage linear accelerator on a ring gantry: Elekta Unity pairs a 1.5 T magnet with a 7 MV beam, and ViewRay MRIdian pairs 0.35 T with a 6 MV beam. Cine imaging during delivery runs at a few frames per second, and the beam holds automatically when the target drifts outside a drawn boundary, which is what allows planning margins to shrink. The second family is thermal ablation guided by MR thermometry, which reads temperature from the proton resonance frequency shift of about 0.01 ppm per degree C and maps heating to roughly 1 degree C in tissue. That measurement is what makes laser interstitial thermal therapy and transcranial focused ultrasound controllable, because the operator watches the thermal dose accumulate instead of inferring it.

Seeing soft tissue during treatment lets margins shrink, and the MIRAGE randomized trial showed what that is worth: 156 prostate patients treated to 40 Gy in 5 fractions with 2 mm margins under MRI guidance had acute grade 2 or worse genitourinary toxicity of 24.4% against 43.4% with 4 mm margins under CT guidance. The costs are throughput and physics. An adaptive fraction that re-plans on the day's anatomy takes 45 to 60 minutes with a physician and physicist present, against about 15 minutes on a conventional linac, so one MR-linac treats a fraction of the patients one conventional machine does. The magnetic field also bends secondary electrons, and the resulting electron return effect deposits extra dose at tissue-air interfaces, which the planning system has to model rather than ignore. Ablation guidance carries its own limit, since thermometry works well in water-rich tissue and poorly in fat and bone.

Use an MR-linac when the target moves or sits against something you cannot afford to irradiate: pancreas, liver, adrenal, prostate against the urethra and rectum, and oligometastases in the abdomen. Do not buy one to treat routine breast or prostate volumes that a conventional linac with cone-beam CT handles in 15 minutes, because you will pay several times the capital and treat fewer patients for the same reimbursement. If the case for daily online adaptation is real in your practice, run the numbers on physician and physicist time first, since that is the binding cost rather than the machine. For ablation, choose MR guidance when you need a temperature map to know you have treated the target and stopped short of the structure next to it, and choose ultrasound or CT guidance when a simpler endpoint will do. Verify what your payers reimburse for adaptive planning before signing, because the codes were written for a workflow that took far less time.

Elekta Unity at 1.5 T with a 7 MV beam, ViewRay MRIdian at 0.35 T with 6 MV · cine imaging at a few frames per second with automatic beam hold · MR thermometry accurate to roughly 1 degree C from a 0.01 ppm per degree C frequency shift · MIRAGE: 2 mm versus 4 mm margins, acute grade 2+ genitourinary toxicity 24.4% versus 43.4% in 156 patients · adaptive fraction 45–60 minutes against about 15 on a conventional linac · MR-linac roughly $8–10M installed against $3–5M for a conventional linac plus vault.

Elekta Unity and ViewRay MRIdian are the two MR-linac platforms, and ViewRay's Chapter 11 filing in 2023 is the clearest evidence of how hard the business model is; Medtronic Visualase and Monteris NeuroBlate deliver MR-guided laser ablation for epilepsy foci and deep brain tumors; Insightec Exablate Neuro performs transcranial focused ultrasound thalamotomy under MR thermometry.

An MR-linac runs roughly $8–10M installed against $3–5M for a conventional linac plus its vault, and the vault still has to be built with RF shielding on top of the concrete. The revenue side does not move with it. Radiation therapy is reimbursed per fraction and per plan under codes written for conventional delivery, so a center delivering five adaptive fractions is paid roughly what a center delivering five conventional ones is paid, while spending three times the machine time and adding a physicist to every fraction. That gap is why ViewRay filed for Chapter 11 in 2023 despite a working, cleared product with published outcomes, and it is the single most useful case study on this sheet for anyone selling capital equipment into oncology. The systems that do pencil out are at high-volume academic centers treating pancreas and liver SBRT, where the alternative is a case they would not treat at all. Ablation is a better business: a laser or ultrasound system costs $1–3M, the disposable applicator is a per-case consumable, and the procedures have established codes.

2 minMRI-Guided Radiation TherapyUW Health · 10k+ views

8 minThe MIRAGE Trial: MRI-Guided Versus CT-Guided SBRT for Prostate CancerGU Oncology Now · 5k+ views

Clinical Applications of Magnetic Resonance-Guided Radiotherapy: A Narrative Review (Cancers) · A narrative review of MRI acquisition for MR-guided-radiotherapy in prostate cancer (Quantitative Imaging in Medicine and Surgery)

PET images a molecule rather than an anatomy. The patient is injected with a tracer carrying a positron emitter, usually about 370 MBq of fluorine-18 fluorodeoxyglucose, and each positron annihilates with an electron to produce two 511 keV photons flying apart in opposite directions. A ring of scintillator crystals, now usually LYSO read out by silicon photomultipliers, records pairs of hits within a few nanoseconds of each other and reconstructs the line each annihilation happened on. Timing resolution of 200–400 ps lets a time-of-flight system place the event within a few centimeters along that line, which sharpens the image without changing the crystals. The CT half of the scanner does two jobs: it supplies the anatomic map the physician reads the tracer against, and it measures tissue density for attenuation correction, without which the PET numbers are wrong. Clinical spatial resolution is 4–5 mm, so PET says where a molecule concentrates and CT says what structure that is.

PET is the only routine way to see metabolism, receptor expression, or protein deposition in a living patient, which is why it has become the staging test in most solid tumors and the eligibility test for anti-amyloid drugs in Alzheimer's disease. Its central weakness is not the scanner but the tracer, because fluorine-18 has a half-life of 110 minutes and starts decaying the moment it is made. Resolution is coarse compared with CT or MRI, uptake is not specific to cancer so infection and inflammation light up too, and the patient sits for a 60-minute uptake period before a 20-minute scan, which caps throughput near 10 to 12 patients a day. Long axial field-of-view systems change that arithmetic: the uEXPLORER covers 194 cm at once instead of 20–25 cm, with roughly 40 times the sensitivity, so the same study can be done in a fraction of the time or a fraction of the dose.

Order PET/CT when the clinical question is whether disease is present somewhere you have not looked, or whether a known lesion is metabolically active, and use CT or MRI when the question is anatomic. Site the scanner where a tracer can reach it: four hours of driving is 240 minutes, or 2.2 half-lives at 110 minutes each, so a dose arrives at about 22% of what was made and the producer has to start with roughly 4.5 times the activity, which is why the delivered cost per dose rises with distance. If you are more than about four hours' drive from a commercial radiopharmacy, price an on-site cyclotron or plan around generator-produced gallium-68 instead. Buy a long axial field-of-view system only if you have the volume to fill it or a research program that needs dynamic whole-body imaging, since it costs several times a conventional scanner. If your PET volume is mostly cardiac perfusion, look at rubidium-82 from a strontium-82 generator, which removes the cyclotron dependence entirely.

511 keV photon pairs detected in coincidence · fluorine-18 half-life 110 minutes · typical FDG dose about 370 MBq · time-of-flight timing resolution 200–400 ps · clinical resolution 4–5 mm · 60-minute uptake plus a 20-minute scan, 10–12 patients per day · long axial field-of-view 194 cm against 20–25 cm conventional, roughly 40 times the sensitivity · scanner $1.5–3M, total-body systems above $7M.

Siemens Biograph Vision, GE Omni Legend and Philips Vereos are the mainstream digital PET/CT platforms; United Imaging uEXPLORER was the first 194 cm total-body scanner, installed at UC Davis; FDG remains the volume tracer, with PSMA agents for prostate cancer and amyloid and tau agents for dementia growing fastest.

A digital PET/CT costs $1.5–3M and needs a shielded room, a hot lab, and a licensed radiation safety program, so the capital decision is close to the MRI decision. The distinguishing cost is the tracer. FDG is a per-dose consumable that a site buys from a commercial radiopharmacy for a few hundred dollars, and because it decays in transit the supplier prices in the extra activity that has to be made, which is why doses cost more the farther a site sits from a cyclotron. Medicare pays roughly $1,200–1,400 for a hospital outpatient PET/CT study, so at 10 to 12 patients a day the scanner covers its capital and the dose cost with room to spare, and utilization is again what decides. A long axial field-of-view system costs above $7M, which is why the installed count is still in the dozens rather than the thousands. A 16–18 MeV cyclotron plus hot cells and a vault runs $5–10M, which only makes sense for a network serving many scanners or a center running research tracers, and that concentration is why a handful of radiopharmacy chains supply most US PET doses. The growth is coming from new tracers rather than new scanners, since each approved agent brings its own indication and its own payment.

2 minHow Does a PET Scan Work?NIBIB · 1m+ views

4 minHow does a PET scan work?Imperial College London · 500k+ views

PET/CT - Positron Emission Tomography/Computed Tomography (RadiologyInfo.org, ACR and RSNA) · Innovations in Instrumentation for Positron Emission Tomography (Seminars in Nuclear Medicine)

A gamma camera images single photons rather than annihilation pairs, so it needs a physical collimator to know where each one came from. A lead plate drilled with thousands of parallel holes sits in front of a sodium iodide crystal, absorbing every photon that is not traveling nearly along a hole axis, and the flashes that get through are located by an array of photomultipliers behind the crystal. Most of the work is done at 140 keV, the gamma energy of technetium-99m, which labels the majority of nuclear medicine studies. SPECT rotates two camera heads around the patient and reconstructs a tomographic volume, typically in 15 to 30 minutes at 10–15 mm resolution, and adding a CT gives attenuation correction and an anatomic map. Newer cardiac-dedicated cameras replace the crystal and photomultipliers with pixelated cadmium zinc telluride, which converts gamma photons directly to charge and gives roughly 5–6% energy resolution against 9–10% for sodium iodide.

The collimator is both what makes the camera work and what limits it, since it discards well over 99.9% of the emitted photons to gain direction information, which is why SPECT sensitivity is orders of magnitude below PET and its resolution is millimeters coarser. What SPECT has instead is a supply chain that works: technetium-99m comes from a generator sitting in the hospital's hot lab and is eluted on demand, so no cyclotron and no same-day delivery are involved. That makes SPECT by far the most common nuclear medicine study worldwide. CZT cameras with focused multi-pinhole geometry recover a large part of the sensitivity gap for cardiac work, cutting a 15-minute acquisition to a few minutes or the injected dose by a similar factor, but they are heart-shaped and do not serve general imaging.

Use SPECT when a technetium-99m or iodine-123 tracer answers the question and you want a study any hospital hot lab can run: bone scans, renal function, parathyroid localization, dopamine transporter imaging for parkinsonism, and post-therapy dosimetry after lutetium-177. Use PET instead when you need quantitative uptake, better resolution, or a tracer that only exists as a positron emitter, and accept the delivery dependency that comes with it. For a cardiology practice doing mostly myocardial perfusion, a CZT cardiac camera is the better buy, since the throughput and dose reduction are real and the machine costs less than a PET/CT. If you are buying general-purpose, buy SPECT/CT rather than SPECT, because attenuation correction and anatomic localization change the read on a large minority of studies. Before committing capital, look at whether your perfusion volume is migrating to cardiac CT or rubidium-82 PET, because that shift is underway.

140 keV imaging energy for technetium-99m · parallel-hole collimation discards well over 99.9% of emitted photons · SPECT resolution 10–15 mm against 4–5 mm for PET · acquisition 15–30 minutes, a few minutes on a CZT cardiac camera · CZT energy resolution about 5–6% against 9–10% for sodium iodide · general-purpose SPECT/CT roughly $400–900k, CZT cardiac cameras $500–700k.

GE Discovery NM/CT and Siemens Symbia are the general-purpose SPECT/CT workhorses; Spectrum Dynamics D-SPECT and GE Discovery NM 530c are the CZT cardiac cameras; DaTscan with iodine-123 ioflupane and technetium-99m MDP bone scans are the highest-volume non-cardiac studies.

A general-purpose SPECT/CT costs roughly $400–900k and a CZT cardiac camera $500–700k, which puts nuclear medicine within reach of a community hospital in a way PET/CT is not. Operating cost is dominated by the technetium generator, a shielded column of molybdenum-99 delivered weekly that a technologist elutes each morning, priced in the low thousands of dollars and sized to the department's expected volume, so a slow week wastes activity that has already decayed. Medicare pays roughly $1,000–1,300 for a hospital outpatient myocardial perfusion study, and cardiac perfusion has historically been the volume that paid for the department. That volume is now contested by coronary CT angiography and by rubidium-82 and flurpiridaz PET, all of which read better in obese patients, so the growth case for a new general-purpose camera is weaker than the installed base suggests. The clearest new demand is post-therapy imaging: every patient treated with lutetium-177 needs quantitative SPECT for dosimetry, which is turning a mature scanner into a companion device for a growing drug class.

5 minGamma camera | Components & Function l Visual explanationDr. Paulien Moyaert · 50k+ views

3 minSPECT/CT Scan animationANSTO · 100k+ views

Nuclear Medicine (National Institute of Biomedical Imaging and Bioengineering) · Technological Advances in SPECT and SPECT/CT Imaging (Diagnostics)

A radiotracer is a molecule that goes where you want to look, carrying an atom that announces its arrival. These are drugs rather than devices, approved by FDA's drug center under an NDA or ANDA with PET production held to the 21 CFR 212 rules for PET drugs, so there is no 510(k) path and no FDA device class, which is why this entry carries none. Every tracer inherits the logistics of its isotope. Technetium-99m, which labels most nuclear medicine studies, has a 6-hour half-life and is milked on demand from a shielded column of molybdenum-99, whose own 66-hour half-life is short enough that the generator is a weekly delivery and long enough that it can cross an ocean. Fluorine-18 for PET has a half-life of 110 minutes and is made in a 16–18 MeV cyclotron by driving protons into oxygen-18 enriched water, then converted to FDG, tested, and driven to the scanner the same morning. Therapeutic isotopes sit at the other end of the scale: lutetium-177 lasts 6.6 days and ships worldwide, actinium-225 lasts 9.9 days and barely exists.

Half-life is the whole design constraint, and it cuts both ways. A short half-life means low patient dose and a clean scan, and it also means nothing can be stockpiled, so the supply chain runs with no inventory buffer and a production outage shows up in clinics within days. Molybdenum-99 is made by fissioning uranium targets in a handful of aging research reactors, and when Canada's NRU and the Netherlands' HFR were both offline in 2009 and 2010, world supply fell by roughly half for months and hospitals rationed scans. Molybdenum-99 loses about 1% of its activity per hour, which is why it is sold in six-day curies, the activity still present six days after processing, rather than by mass. Fluorine-18 is the opposite problem: it cannot be shipped far at all, so the map of PET scanners follows the map of cyclotrons.

If you are designing a tracer, pick the isotope for the biology first and then check whether its supply exists at clinical scale, because a beautiful agent on an isotope nobody makes is a research program rather than a product. Use technetium-99m when a generator in the hospital hot lab is the advantage you want, and accept its coarser imaging. Use fluorine-18 when you need PET resolution and your sites are within a few hours of a radiopharmacy, and use a gallium-68 generator instead when they are not, since a germanium-68 parent lasts 271 days and makes the site independent of daily delivery. For therapy, treat lutetium-177 as available and actinium-225 as allocated, and build a clinical program around the second only with a written supply agreement. If you are investing in production, the durable position is in isotopes with proven demand and structurally short supply rather than in another tracer for an isotope already made in quantity.

Technetium-99m half-life 6 hours from a molybdenum-99 parent at 66 hours · fluorine-18 half-life 110 minutes from a 16–18 MeV cyclotron · gallium-68 at 68 minutes from a germanium-68 parent lasting 271 days · lutetium-177 at 6.6 days, actinium-225 at 9.9 days · molybdenum-99 decays about 1% per hour and is sold in six-day curies · the 2009–2010 reactor outages cut world molybdenum-99 supply by roughly half · global actinium-225 production about 68 GBq per year.

HFR Petten, BR2 in Belgium, SAFARI-1 in South Africa, OPAL in Australia and LVR-15 in the Czech Republic supply most of the world's molybdenum-99, all converted from highly enriched to low-enriched uranium targets; NorthStar Medical Radioisotopes restarted US domestic molybdenum-99 supply in 2018 and SHINE Technologies is building accelerator-driven production; Oak Ridge National Laboratory, JRC Karlsruhe and IPPE Obninsk are the three thorium-229 sources that between them make roughly 68 GBq of actinium-225 a year.

The isotope is usually a small part of what a scan costs and the entire reason a scan can happen. A technetium generator costs a hospital a few thousand dollars a week and covers a department's studies, while an FDG dose costs a few hundred dollars and rises with distance from the cyclotron because the producer has to make extra activity to cover decay in transit. Molybdenum-99 was sold below its production cost for decades, since the reactors making it were funded as research facilities and treated isotope output as a byproduct, and the OECD Nuclear Energy Agency's push for full cost recovery after the 2009 shortage raised prices several-fold without adding much capacity. A 16–18 MeV cyclotron costs roughly $2–3M, and the vault, hot cells and radiochemistry around it bring a production facility to $5–10M, which is why a handful of radiopharmacy networks supply most US PET doses rather than hospitals making their own. Therapeutic isotopes have the better margins and the worse supply: a lutetium-177 dose is priced as part of a drug selling for tens of thousands of dollars per course, and actinium-225 output is allocated against demand that clinical trial enrollment alone could exhaust. That mismatch is what has drawn DOE, national labs and several startups into accelerator production routes.

5 minTechnetium generator | Everything you need to knowDr. Paulien Moyaert · 50k+ views

3 minProduction of Technetium 99mA Level Physics HQ · 50k+ views

Molybdenum-99 for Medical Imaging (National Academies of Sciences, Engineering, and Medicine) · Challenges and future options for the production of lutetium-177 (European Journal of Nuclear Medicine and Molecular Imaging)

Radioligand therapy attaches a therapeutic radionuclide to a molecule that binds something tumors display and healthy tissue mostly does not, then relies on short-range radiation to do the killing where the ligand lands. These are drugs, approved through FDA's drug center and paid for as drugs, so they carry no device class and no 510(k) predicate, which is why this entry has no FDA class. The workhorse emitter is lutetium-177, a beta emitter with a 6.6-day half-life whose electrons deposit their energy within about 2 mm of tissue, wide enough to hit neighboring cells the ligand missed and narrow enough to spare the organ next door. Alpha emitters such as actinium-225 deliver 5–8 MeV over 50–100 micrometers, which produces double-strand breaks that cells repair poorly, at the cost of a much harder supply problem. Dosing is a course rather than a single treatment: Pluvicto is 7.4 GBq every 6 weeks for four to six cycles, given in a shielded room by a nuclear medicine team.

The clinical evidence is real and randomized. In the VISION trial, 831 men with PSMA-positive metastatic castration-resistant prostate cancer had median overall survival of 15.3 months on lutetium-177 PSMA-617 plus standard care against 11.3 months on standard care alone, with imaging-based progression-free survival of 8.7 against 3.4 months. The weaknesses are logistics and supply. Each dose is made to order for a named patient on a scheduled day, and a 6.6-day half-life means a missed appointment wastes the dose rather than delaying it, so the manufacturing plant, the courier, and the clinic calendar are one system. Novartis launched Pluvicto into exactly that constraint and spent much of 2022 and 2023 supply-limited while it qualified additional production sites. Actinium-225 is far scarcer: global output is roughly 68 GBq a year, and at about 10 MBq per dose that is 6,800 doses, or on the order of 1,700 patient courses of four cycles each, an arithmetic step taken here rather than a published figure.

Use radioligand therapy when a companion PET scan shows the target is actually expressed, since these agents are prescribed off an image rather than off a biopsy and a PSMA-negative patient will not respond. Treat it as a later-line option in the settings where it is approved, and expect that to move earlier as trials read out. If you are building a program, count the fixed costs first: a shielded administration room, a radiation safety officer, waste handling, and a scheduler who can hold a slot to the day, because the drug arrives whether or not the patient does. If you are investing, be skeptical of any alpha-emitter program without a written isotope supply agreement, since actinium-225 output is allocated and clinical trial enrollment alone can consume it. Choose beta emitters for bulky disease where crossfire helps and alpha emitters for micrometastatic or resistant disease where a 50–100 micrometer range is the point.

Lutetium-177 half-life 6.6 days with a beta range near 2 mm in tissue · actinium-225 alphas at 5–8 MeV over 50–100 micrometers · Pluvicto dosed at 7.4 GBq every 6 weeks for four to six cycles · VISION: median overall survival 15.3 versus 11.3 months and imaging-based progression-free survival 8.7 versus 3.4 months in 831 patients · list price roughly $42k per dose, about $250k for a full course · global actinium-225 supply about 68 GBq a year, roughly 1,700 patient courses.

Novartis Lutathera, lutetium-177 dotatate for gastroenteropancreatic neuroendocrine tumors, approved in 2018; Novartis Pluvicto, lutetium-177 vipivotide tetraxetan for PSMA-positive prostate cancer, approved in 2022 and extended to the pre-chemotherapy setting in 2025; Bayer Xofigo, radium-223, the first alpha emitter approved for bone metastases in 2013; actinium-225 PSMA agents remain investigational.

This is a drug business wearing a nuclear medicine department's clothes. Pluvicto lists at roughly $42,000 per dose, so a six-cycle course is about $250,000, and Novartis reported roughly $1.4B of Pluvicto sales in 2024, which makes it the proof point the whole field is financed against. The margin sits with the drug owner, not the hospital: a US hospital buys the dose and bills it under Medicare Part B at average sales price plus a small percentage, then collects a procedure fee for administering it, so the department's return comes from volume and from the imaging that goes with it rather than from the drug. Capital is modest by this sheet's standards, since a shielded administration room and hot lab upgrades run in the low hundreds of thousands of dollars, which is why community centers can enter. The binding constraints are isotope supply and cold-chain-style scheduling rather than equipment, and that is where the acquisitions went: Bristol Myers Squibb paid about $4.1B for RayzeBio and Eli Lilly about $1.4B for Point Biopharma in late 2023, both buying pipelines plus manufacturing. For a startup, owning isotope production or a differentiated ligand is defensible; being a third lutetium-177 PSMA agent is not.

11 minLutetium-177: PSMA Guided Treatment | #ProstateCancer | #MarkScholzMD | #PCRIProstate Cancer Research Institute · 10k+ views

16 minLu-PSMA Radioligand TherapyUniversity of California Television (UCTV) · 10k+ views

DailyMed - PLUVICTO- lutetium lu 177 vipivotide tetraxetan injection, solution (DailyMed, U.S. National Library of Medicine) · Supply and Clinical Application of Actinium-225 and Bismuth-213 (Seminars in Nuclear Medicine)

A diagnostic ultrasound probe holds an array of roughly 64 to 256 piezoelectric elements. Each element is pulsed to send a short burst of sound into the body and then listens for echoes returning from boundaries where acoustic impedance changes. Sound travels through soft tissue at about 1,540 m/s, so the round-trip time of an echo gives its depth directly, and steering the beam electronically across the array builds a two-dimensional image tens of times a second. Frequency sets the central trade-off, because attenuation runs at roughly 0.5 dB per centimeter per MHz: a 2–5 MHz curved array reaches 20–25 cm into an abdomen, while a 7–15 MHz linear array resolves well under a millimeter but only sees the first few centimeters. Doppler processing adds blood velocity from the frequency shift of moving scatterers, and the same probe and console also do elastography and contrast imaging with injected microbubbles. A complete cart system costs $20,000–250,000, plugs into a normal wall outlet, and needs no shielding, no cryogen, and no construction project.

It is real time, carries no ionizing dose, and can be wheeled to the patient, so it is the only cross-sectional modality that works during a resuscitation. The physics also sets hard limits. Bone and gas have acoustic impedances far from soft tissue, so nearly all the sound reflects at those boundaries, which makes lung, bowel gas, and anything behind bone invisible. Attenuation scales with body thickness, so image quality falls off in larger patients exactly where the clinical question is often hardest. The weakness that matters commercially is operator dependence: the diagnostic content is created at the moment of scanning by whoever holds the probe, so a stored study documents what that person chose to record, and a missed finding and a normal exam look identical afterward.

Reach for ultrasound first when the question is about fluid, flow, or a soft-tissue structure you can find an acoustic window to: gallbladder, kidney, pregnancy, cardiac function, veins, thyroid, and any needle you want to watch go in. If the target sits behind bone or air, use CT or MRI instead and do not waste the attempt. If the study has to be re-read months later by someone who was not in the room, weigh the fact that ultrasound documents the operator's choices rather than a fixed volume. If you are buying a screening program rather than a diagnostic tool, insist on a written protocol and credentialed operators, which is how abdominal aortic aneurysm screening works and how informal bedside scanning does not. As a rule of thumb, if a mid-range cart and a trained sonographer can answer the question, nothing else on this sheet is close on cost.

Speed of sound in soft tissue about 1,540 m/s · probes 2–15 MHz, attenuation roughly 0.5 dB/cm/MHz · 20–25 cm depth at 2–5 MHz, a few centimeters at 7–15 MHz · arrays of roughly 64–256 elements · cart systems $20,000–250,000 · Medicare allowed about $173 for a complete echocardiogram and $99 for a complete abdominal study in an office in 2024 · roughly 6.8 million complete echocardiograms billed to Medicare Part B that year.

GE HealthCare Vivid and LOGIQ, Philips EPIQ, Siemens Healthineers ACUSON, Canon Aplio, and Mindray Resona carts in essentially every hospital; FAST exams in trauma bays; obstetric anomaly scanning; ultrasound-guided central line placement, which became standard practice because watching the needle cut complication rates.

The capital is small and the labor is not. A $20,000–250,000 cart sits one to two orders of magnitude below a CT or MRI, needs no siting work, and is bought out of a department budget rather than a capital committee, which is why ultrasound spread everywhere. Service contracts on imaging equipment usually run 5–10% of purchase price a year, and probes at roughly $3,000–10,000 each are the item that actually gets replaced, most often because one was dropped. Revenue per study is thin and volume is enormous: Medicare's 2024 physician utilization file shows about 6.8 million complete echocardiograms at roughly $173 allowed in an office setting and $67 for the professional component in a hospital, and about 720,000 complete abdominal studies at roughly $99 in an office. Because each of those studies is 20–45 minutes of a sonographer's time, the cumulative salary of the person running a cart over its service life dwarfs the purchase price, so the real productivity lever is scan time rather than machine cost. Five vendors hold most of the market, and price pressure has come from below as Mindray and Samsung took share at the mid range. For a new entrant the hard part is that there is no unmet capital need to sell against, so the pitch has to be about who can operate the machine rather than what the machine costs.

6 minUltrasound medical imaging | Mechanical waves and sound | Physics | Khan Academykhanacademymedicine · 100k+ views

14 minUltrasound Transducer (Part 1) Piezoelectric Material and Matching Layer | Ultrasound Physics #9Radiology Tutorials · 50k+ views

Ultrasound Physics and Instrumentation (StatPearls, NCBI Bookshelf) · Ultrasound (National Institute of Biomedical Imaging and Bioengineering)

In a handheld scanner the probe is the whole machine: array, transmit electronics, beamformer and digitizer all sit in the housing, and a phone or tablet supplies the screen, the storage and the network. Butterfly's iQ took the furthest step by replacing the diced piezoelectric stack with a two-dimensional array of roughly 9,000 individually addressable silicon MEMS membranes fabricated directly on the CMOS that drives them, alongside more than 11,000 amplifiers and over 1,100 analog-to-digital converters on the same die, at 3 W or less. A conventional array is cut from a single block of piezoelectric ceramic and wired element by element, which is slow and expensive at fine pitch; a micromachined array is defined lithographically, so the same die spans the 1–10 MHz range that a cart needs three separate probes to cover. Most competitors kept piezoelectric arrays in a small housing, including Philips Lumify, GE Vscan Air, Mindray TE Air and Clarius, while Exo uses piezoelectric micromachined elements as a third route. Butterfly shipped in 2018 at a $1,999 list price and the current probes sit in the low thousands, against $20,000–250,000 for a cart.

The price and the form factor are the whole argument: a department can hand out twenty probes for what one mid-range cart costs, and the scan happens where the decision is made rather than after a transport and a wait. Image quality is the honest weakness. In a 2024 cross-sectional study, 35 point-of-care ultrasound experts scanned the same standardized patients with six handhelds and rated the images, and no single device came out best across right upper quadrant, cardiac apical four-chamber, and superficial neck and lung views. Chip heating and battery life cap continuous scanning, and small apertures cost penetration in larger patients. The probe is also useless without an active software license, so the buyer takes on a recurring cost that a cart does not carry.

Buy handhelds when the clinical questions are close to binary and the value is speed: free fluid, bladder volume, gross left ventricular function, pericardial effusion, and finding a vein or a nerve before you put a needle in. Buy a cart instead when the study has to be a complete diagnostic exam, formally interpreted and archived, because that is what the payment and the medico-legal record are built around. If you are equipping trainees or a low-resource clinic, cost per learner is the deciding number and handhelds win outright. Budget for the subscription, the archive path into the hospital record, and a quality-assurance program before you count any revenue, since a scan with no stored image and no written interpretation cannot be billed under codes like 76705 for a limited abdominal study, which Medicare allowed at about $63 in an office in 2024, or 76942 for needle guidance at about $55. If someone pitches AI guidance as a substitute for training, read the trials first: in the study behind the FDA's first authorization for acquisition guidance, 8 nurses with no prior echocardiography experience produced diagnostic-quality images in 237 of 240 patients, but the endpoints were left ventricular size and function, right ventricular size and pericardial effusion, which is triage rather than a full echocardiogram.

About 9,000 addressable MEMS elements on one die, with 11,000+ amplifiers and 1,100+ ADCs, at 3 W or less · one probe spans 1–10 MHz, replacing three cart probes · $1,999 launch price in 2018, current probes in the low thousands, against $20,000–250,000 for a cart · per-user software subscription in the hundreds of dollars a year · Medicare allowed roughly $63 for a limited abdominal ultrasound in an office and $55 for ultrasound needle guidance in 2024 · 237 of 240 novice-acquired studies rated diagnostic quality for three gross cardiac findings.

Butterfly iQ3, Philips Lumify, GE Vscan Air, Mindray TE Air, Clarius, EchoNous Kosmos and Exo Iris; Caption Guidance, authorized through FDA De Novo DEN190040 in February 2020 as the first AI software to guide image acquisition, whose developer GE HealthCare acquired in 2023.

This is the clearest case on the sheet of a semiconductor process displacing a hand-assembled one, and it changed the revenue model along with the bill of materials. A cart is a $20,000–250,000 capital sale with a service contract usually running 5–10% of price a year and a replacement cycle measured in the better part of a decade. A handheld is a low-thousands hardware sale plus a per-user subscription in the hundreds of dollars a year, so the vendor's revenue scales with the number of clinicians scanning rather than the number of machines, and the software tiers carry most of the margin because hardware at $2,000–3,000 does not. Wafer economics also mean unit cost falls with volume and yield, which a diced ceramic array does not do. The buyer's case is usually not a billing code, since focused exams pay $50–65 under Medicare and are frequently bundled into the visit, so purchases get justified on time saved, formal studies avoided, and fewer failed procedures. Incumbents answered by shipping their own handhelds rather than cutting cart prices, which is the rational move when the installed base is where the service revenue sits.

4 minButterfly Network Puts 3D Ultrasound on a Chip with CadenceEE Journal · 10k+ views

14 minButterfly IQ: Review of the portable handheld ultrasoundAli Haider MD · 100k+ views

Ultrasound-on-chip platform for medical imaging, analysis, and collective intelligence (Proceedings of the National Academy of Sciences) · Point of Care Ultrasound (AIUM)

A therapeutic focused ultrasound helmet holds a hemispherical phased array of 1,024 elements driven at about 650 kHz. Each element's phase is corrected using a CT scan of that patient's own skull, so the beams arrive together at a single spot a few millimeters across, deep in the brain, after passing through intact scalp and bone. Intensity is high only at the focus, so tissue along the path stays cool while the target reaches roughly 55–60 °C in seconds and coagulates. MRI supplies both the targeting and a live temperature map from the proton resonance frequency shift, which lets the operator run low-power test sonications at 40–45 °C, watch the awake patient's tremor and side effects, adjust the target, and only then raise the power enough to make a permanent lesion. The skull is the gating factor: the FDA screening criterion is a skull density ratio of at least 0.45 measured on CT, and patients below it transmit too little energy to heat the target. The same physics at lower power does something entirely different, and that distinction runs through the rest of this entry.

Nothing is implanted and nothing is opened, so there is no burr hole, no lead, no pulse generator to replace, and no hardware infection risk, which is the whole case against deep brain stimulation. The efficacy is real and was measured against a sham: in the pivotal trial, hand tremor scores fell from 18.1 to 9.6 points at three months in the treated group against 16.0 to 15.8 in the sham group, a between-group difference of 8.3 points on a 32-point scale, and the improvement held at 12 months. The cost is that a lesion is permanent and cannot be tuned down later the way a stimulator can be reprogrammed. Side effects in that trial were common: gait disturbance in 36% of patients and paresthesias or numbness in 38%, still present at 12 months in 9% and 14%. Throughput is also poor, because a single treatment occupies an MRI suite for a few hours.

Offer MR-guided focused ultrasound first to a patient with medication-refractory essential tremor who cannot have surgery or does not want an implant, since that is the indication it was approved for and the one with sham-controlled evidence. Choose deep brain stimulation instead when the patient needs bilateral treatment up front, needs the option to adjust or reverse the effect, or screens out on skull density ratio. If you are assessing a company in this space, first establish which physics it is selling: thermal ablation at around 650 kHz destroys tissue and has an approved indication, while blood-brain barrier opening runs near 220 kHz with injected microbubbles, transiently loosens tight junctions instead of heating anything, and has no approved indication at all. Treat the barrier-opening trials in glioblastoma drug delivery and Alzheimer's disease as the frontier they are, with the added detail that skull density ratio stops being a limit at the lower frequency, so the eligible population is much larger if it ever works. Before buying a system, count the diagnostic MRI hours it displaces, because that opportunity cost usually decides whether a program pencils out.

1,024-element hemispherical array at about 650 kHz, focus a few millimeters across · target heated to roughly 55–60 °C, test sonications at 40–45 °C · skull density ratio of at least 0.45 to qualify · hand tremor 18.1 to 9.6 points at 3 months against 16.0 to 15.8 for sham, an 8.3-point difference on a 32-point scale · gait disturbance 36% and paresthesias 38%, persisting in 9% and 14% at 12 months · blood-brain barrier opening runs near 220 kHz with microbubbles and has no approved indication.

Insightec's Exablate Neuro, approved under PMA P150038 in July 2016 for essential tremor and extended by later supplements to Parkinson's disease and to staged bilateral treatment; the same company's Exablate 2000 for uterine fibroids, approved back in October 2004; Profound Medical's Sonalleve; Sonablate and Focal One for prostate tissue ablation; ongoing barrier-opening trials in glioblastoma and Alzheimer's disease.

The capital stack is the first problem. The helmet and console are a seven-figure purchase and do not work alone, since the system needs a dedicated 1.5 T or 3 T MRI at another $1–3 million plus siting, so a program clears $3 million before treating anyone, and each treatment then ties up that magnet for a few hours instead of running eight to ten diagnostic scans. Reimbursement is the second and larger problem. Medicare's 2024 physician utilization file still shows brain focused ultrasound billed under 0398T, a Category III tracking code, with about 1,600 services nationally, and because Category III codes are priced by local contractors rather than a national fee schedule, the allowed amount differed sharply by state, at roughly $880 in California against $1,670 in Florida. Nine years after approval, that is a very small business for a technology that works. The health-economic case is strong where anyone has modeled it: an England analysis put five-year costs at £19,779 for focused ultrasound against £62,348 for deep brain stimulation with slightly more quality-adjusted life years, so it came out cheaper and better. Uterine fibroid treatment is the cautionary version of the same story, approved in 2004 and still small because it competes against hysterectomy and uterine artery embolization, both of which already had codes and referral habits. For a Class III device carrying a new procedure, the coding pathway is usually the binding constraint rather than the approval.

5 minFocused Ultrasound for Parkinson’s Disease and Essential Tremor | Penn MedicinePenn Medicine · 100k+ views

12 minFocused Ultrasound Treatment for Essential Tremor | UCSF NeurosurgeryUCSF Neurosurgery · 10k+ views

The cost-effectiveness of unilateral magnetic resonance-guided focused ultrasound in comparison with unilateral deep brain stimulation for the treatment of medically refractory essential tremor in England (The British Journal of Radiology) · State of the Field (Focused Ultrasound Foundation)

OCT is an interferometer used as a depth gauge. Near-infrared light is split between the tissue and a reference path, and the two only interfere when the path lengths match to within a few microns, so the interference signal encodes how deep each reflection came from. Scanning that across the retina builds a cross-section with roughly 5–10 µm axial resolution over the 1–3 mm the eye is transparent through, which is fine enough to resolve individual retinal layers. In clinic that means a technician captures a macular cube in a few seconds without touching the eye, and the software reports central macular thickness and nerve fiber layer thickness in microns against a normative database. OCT angiography adds capillary flow maps by comparing repeated scans at the same spot, with no injected dye. The catheter version spins a fiber optic inside a coronary catheter under a millimeter across, a few hundred rotations a second, during a pullback with contrast flushed through to clear the blood, and resolves stent struts and plaque at 10–20 µm where intravascular ultrasound manages 100–150 µm.

It is fast, needs no contact and no dye, and produces numbers that are repeatable enough to trend visit over visit, which is what turned it into the decision instrument for retinal disease rather than just a picture. The limit is depth, since scattering confines useful imaging to 1–3 mm, so it works in a transparent eye and inside a vessel you can flush and nowhere else. Blood scatters strongly, which is why intravascular OCT needs a contrast injection on every pullback and why it is a poor fit for patients with kidney impairment. Cataract and vitreous hemorrhage degrade retinal scans in the same way they degrade the clinical exam. The normative databases are also population-specific, so highly myopic and otherwise unusual eyes get flagged as abnormal by the software, and the automated thickness numbers have to be read alongside the image rather than instead of it.

If the question is retinal layer structure, fluid, or nerve fiber thinning, use OCT and repeat it on a schedule, because the value is in the trend. If you need to see leakage and staining rather than flow, fluorescein angiography still answers a question OCT angiography does not. Inside a coronary artery, choose OCT when the decision turns on fine detail such as stent apposition, edge dissection, or calcium thickness, and choose intravascular ultrasound when you need to see out to the vessel's outer wall, are treating an ostial lesion, or cannot give the patient more contrast. Do not buy OCT for anything that needs more than a couple of millimeters of penetration; that is what ultrasound is for. As a purchasing rule of thumb, an ophthalmic system at $30,000–100,000 sits well inside a practice's own budget, so the question is scan volume rather than capital approval.

Axial resolution roughly 5–10 µm, imaging depth 1–3 mm in scattering tissue · intravascular OCT 10–20 µm against 100–150 µm for intravascular ultrasound · ophthalmic systems $30,000–100,000 · Medicare allowed about $40 for a retinal OCT in an office in 2024 · roughly 8.2 million retinal OCT scans and 11 million OCT scans of all types billed to Medicare Part B that year · about 3.6 million intravitreal injections in the same file.

Zeiss Cirrus, Heidelberg Spectralis, Topcon Maestro and Nidek systems in essentially every retina and glaucoma practice; Abbott's intravascular OCT consoles and single-use imaging catheters, and Terumo's competing system; intraoperative OCT built into surgical microscopes for retinal and corneal work.

Ophthalmology is one of the largest imaging markets in medicine by procedure count, and almost nobody outside the specialty knows it. Medicare's 2024 physician utilization file records about 8.2 million retinal OCT scans and another 2.8 million of the optic nerve, roughly 11 million in total, against 6.8 million complete echocardiograms in the same file. Payment per scan is small, about $40 allowed in an office, so a $60,000 machine covers its purchase price after about 1,500 scans of gross revenue (that division is done here, not published), which a busy retina practice reaches in a few months. The scan matters more than its own fee because it determines whether the patient gets injected that month: the same file shows about 3.6 million intravitreal injections at roughly $123 to the physician, and the drug dominates the bill, with aflibercept lines multiplying out to roughly $2.2 billion in allowed charges (2.7 million billed units at about $822 each, arithmetic done here rather than a published total). A cheap instrument sitting upstream of a multi-billion-dollar drug spend is a good position to hold, and it is also why payers audit scan frequency. Intravascular OCT is a different business entirely: the console is usually placed cheaply and the revenue comes from single-use catheters in the high hundreds to low thousands of dollars each, which is a consumable model, and its ceiling is that most operators can still place a stent on angiography alone.

14 minHOW TO READ MACULAR OCT PRINTOUT? made easy!!Insight Ophthalmology · 100k+ views

4 minOptical coherence tomography - Looking into the vessel with lightMEDICA · 5k+ views

Optical Coherence Tomography (EyeWiki) · [Intravascular optical coherence tomography [Invited] (Biomedical Optics Express)](https://pmc.ncbi.nlm.nih.gov/articles/PMC5480504/)

A flexible video endoscope is a steerable tube with a CMOS image sensor at the tip, illumination delivered by LED or fiber, one or two working channels of roughly 2.8–3.7 mm, air, water and suction lines, and four-way tip angulation pulled by control wires from the handle. A colonoscope's insertion tube is about 12–13 mm across and 1.6 m long; a gastroscope is closer to 9–10 mm and 1.05 m. The processor does much of the diagnostic work, since narrow-band imaging filters the illumination to the wavelengths hemoglobin absorbs and makes mucosal vessel and pit patterns stand out without any dye. Fluorescence imaging is a second camera path on the same tower: illuminate near 780–805 nm, filter the camera to the roughly 830 nm emission, and inject indocyanine green, which binds plasma proteins so it stays in the vasculature and is cleared by the liver within minutes. Surgeons use that to judge whether a bowel anastomosis is perfused, to map lymphatic drainage, and to find bile ducts and liver lesions. A separate agent, 5-aminolevulinic acid, makes glioma tissue fluoresce red under blue light and is regulated as a drug rather than as part of the imaging system.

The combination of direct vision and a working channel is what makes endoscopy unusual: the same session can find a polyp, biopsy it, and remove it, so screening and treatment happen in one visit. Reprocessing is the structural weakness. Long narrow lumens and, on duodenoscopes, a moving elevator mechanism cannot survive steam sterilization, so scopes get high-level disinfection instead, and outbreaks of carbapenem-resistant infections were traced to duodenoscopes that had been reprocessed exactly according to the manufacturer's instructions. That drove the FDA toward disposable-component and fully single-use designs. Fluorescence has a different weakness: indocyanine green signal strength depends on dose, timing, camera gain and working distance, so a bright image is a qualitative impression rather than a perfusion measurement, and quantification is still unsolved. Ontario Health's assessment found that adding it to colorectal surgery reduced anastomotic leaks and reoperations, but graded the certainty of that evidence as low.

Use flexible endoscopy whenever the target is mucosa you need to see, sample, or treat, and accept that nothing else gives you all three in one sitting. If the question is depth of invasion or what sits behind the wall, add endoscopic ultrasound or cross-sectional imaging instead of pushing the optics harder. Choose single-use scopes where contamination risk and turnaround time dominate the decision, such as ERCP in patients already colonized with resistant organisms or bronchoscopy at the bedside in an ICU; at ordinary elective volumes reusables still cost less per procedure. Add indocyanine green when you will actually change the operation based on perfusion or lymphatic drainage, and do not add it expecting a number you can put in the note. If you are equipping a unit, size the scope fleet and the reprocessing room first, because those decide throughput more than the number of procedure rooms does.

Working channel roughly 2.8–3.7 mm; colonoscope insertion tube about 12–13 mm across and 1.6 m long · indocyanine green excited near 780–805 nm, emitting near 830 nm, hepatically cleared in minutes · scopes roughly $25,000–45,000 each and a tower $60,000–150,000 · reusable duodenoscope reprocessing measured at €80 per ERCP observed and €183–254 modeled, at about 223 ERCPs per scope per year · Medicare allowed about $241 to the physician for a colonoscopy with biopsy in 2024 and about $277 for an average-risk screening colonoscopy, with the facility payment on top.

Olympus, Fujifilm and Pentax Medical GI and bronchoscopy platforms; Boston Scientific's EXALT Model D single-use duodenoscope, cleared in December 2019, and its single-use bronchoscope; Ambu's single-use scope line; Stryker's SPY fluorescence systems, whose first intraoperative clearance under Novadaq dates to 2005, and Intuitive's Firefly on the da Vinci; GI Genius, authorized through FDA De Novo DEN200055 in April 2021 as the first AI aid for polyp detection during colonoscopy.

The capital here is a fleet rather than a machine. A tower runs $60,000–150,000 and each scope $25,000–45,000, so a multi-room endoscopy unit with 20 to 40 scopes and an automated reprocessing room lands in the $1–3 million range, and the scope inventory is sized by turnaround time rather than by procedure count. The recurring costs are reprocessing and repair: a four-hospital French study measured €80 per ERCP for purchase, maintenance, microbiological control and reprocessing combined, with modeled figures of €183–254 depending on drying and automation, against a workload of about 223 ERCPs per duodenoscope per year. Revenue comes from volume at modest unit prices, with Medicare allowing the physician about $241 for a colonoscopy with biopsy across 1.2 million services in 2024 and about $277 for an average-risk screening colonoscopy, plus a separate and larger facility or ambulatory surgery center payment. Single-use scopes are priced in the low thousands of dollars per procedure, which is more than reprocessing costs at any realistic volume, so hospitals adopt them where infection risk justifies it rather than to save money. Fluorescence sells as a camera upgrade into a tower the hospital already owns, and the dye is cheap at tens of dollars a vial, so the margin sits in the imaging head; Stryker paid roughly $700 million for Novadaq in 2017 to own that position. Olympus holds the large majority of the global GI endoscope market, commonly put near 70%, which is why new entrants sell alongside the scope through AI, single-use designs and capsules rather than trying to displace it.

13 minEndoscope AnatomyOfstead Insights · 50k+ views

10 minIndocyanine Green in Surgery | Basics of Fluorescence-Guided SurgeryEdusurg Clinics · 1k+ views

Infections Associated with Reprocessed Duodenoscopes (FDA) · Indocyanine Green Fluorescence Imaging for Colorectal Surgery: A Health Technology Assessment (Ontario Health Technology Assessment Series)

Photoacoustic imaging fires a nanosecond laser pulse, typically 5–10 ns somewhere in the 680–970 nm window, into tissue. Hemoglobin, melanin or lipid absorbs it, heats by a few millikelvins, expands, and launches a broadband ultrasound pulse that an ordinary clinical ultrasound array picks up. The result has optical contrast, meaning it shows what absorbs light, at ultrasonic resolution and depth, and acquiring at several wavelengths separates oxygenated from deoxygenated hemoglobin, so the image is a map of blood oxygen saturation with no injected agent. Useful depth is about 3–5 cm, because skin exposure limits cap the pulse energy at roughly 20 mJ/cm² in the near infrared and light attenuates steeply below that. Diffuse optical methods give up on forming an image from unscattered light entirely: shine near-infrared light in at one point, collect it centimeters away, and fit a photon-diffusion model to recover bulk absorption and scattering, which yields spectroscopy with roughly 1 cm of localization rather than a picture. Pulse oximetry and cerebral tissue oximetry are the two members of that family that became routine clinical equipment, and both report a number rather than an image.

The appeal of photoacoustics is that it gets functional contrast, hemoglobin concentration and oxygenation, at depths where pure optical imaging has long since diffused into noise, and it does so on detection hardware that already sits in every hospital. The unsolved problem is quantification: signal amplitude scales with the local light fluence, and the fluence at depth depends on the tissue's own optical properties, which is exactly what you were trying to measure. That circularity is why oxygen saturation numbers from deep photoacoustic images are still not trustworthy, and no amount of better detection fixes it. Practical limits stack on top: tunable nanosecond lasers are large and expensive, pulse repetition rates constrain frame rate, and surface melanin absorbs a large share of the light before it gets anywhere. Diffuse optical methods have the opposite profile, cheap and wearable and safe, with resolution so coarse that they can report a trend at a fixed site and cannot localize a lesion.

If you need oxygenation trends at one anatomical site and a single number is enough, use near-infrared tissue oximetry, which is commercial, inexpensive, and already used routinely in cardiac surgery and neonatal intensive care. If you are characterizing a breast mass at one of the few sites with the equipment, opto-acoustic imaging added to ultrasound is the single photoacoustic application with a US premarket approval, and everything else you read about is investigational. If you are building in this area, pick an indication where the current comparator is genuinely bad, because adding a modality on top of ultrasound only pays when it changes a decision ultrasound gets wrong, and in breast imaging the comparator is a cheap core needle biopsy that answers the question outright. If you are evaluating a company, ask which code it bills under and which trial supports the claim; without both, it is an instrument business selling to grant-funded labs, which is a real business but a much smaller one.

Pulses of 5–10 ns in the 680–970 nm window · optical contrast at ultrasonic resolution, useful depth about 3–5 cm · skin exposure limited to roughly 20 mJ/cm² in the near infrared · diffuse optical localization about 1 cm · one US premarket approval for a photoacoustic imaging system, granted in January 2021 · research and clinical-research systems typically $150,000–500,000.

Seno Medical's Imagio Breast Imaging System, approved under PMA P200003 in January 2021, which overlays opto-acoustic maps on B-mode ultrasound to help classify breast masses; iThera Medical's MSOT systems and FUJIFILM VisualSonics Vevo LAZR, both sold as research instruments; Medtronic INVOS and Edwards ForeSight cerebral oximeters, which are the diffuse-optical technology that actually reached the bedside; functional near-infrared spectroscopy and diffuse optical tomography of the breast, both still research.

As of today this is an instrument business rather than a device business. Systems sell at $150,000–500,000 to research groups and academic medical centers, the buyer is a grant rather than a capital committee, the sales cycle follows funding cycles, and the worldwide installed base is in the hundreds. The bill of materials is dominated by the light source, since a tunable nanosecond optical parametric oscillator is a large share of system cost and most of its bulk, which is why nearly every roadmap points at pulsed laser diodes and LEDs that cost one to two orders of magnitude less and deliver far less pulse energy, trading depth for price. Detection is close to free by comparison, because a standard clinical ultrasound array works, so a photoacoustic front end can be added to a cart the hospital already owns. The one part of this family with durable clinical revenue is tissue oximetry, sold as a monitor with disposable sensor pads at tens of dollars each, which is a consumable model with predictable margins. For photoacoustics to become a device business it needs an indication where the alternative is invasive or unavailable, a pivotal trial, and eventually a Category I code; until then the honest description is that one product has an approval and the rest of the field is funded by research budgets.

7 minThe Incredible Cancer-Detecting Potential of Photoacoustic Imaging | Lei Li | TEDTED · 50k+ views

4 minPhotoacoustic Imaging BIMA2016Turku Bioimaging · 10k+ views

4 minFunctional Near Infrared Spectroscopy (fNIRS) Explained! | Neuroscience Methods 101Psyched! · 10k+ views

Optical Imaging (National Institute of Biomedical Imaging and Bioengineering) · Clinical photoacoustic imaging (Photoacoustics)

A soft-tissue surgical robot is a teleoperated instrument holder. The surgeon sits at a console a few meters from the patient, looks into a 3D stereo display, and moves two hand controllers, while a patient-side cart drives three or four arms holding 8 mm instruments through ports in the insufflated abdomen. Each instrument has a cable-driven wrist near the tip, so it bends and rotates inside the body in a way a rigid laparoscopic stick cannot. Each arm pivots about a fixed remote center at the port, which keeps the machine from levering against the abdominal wall as the tip moves. Software scales hand motion down, typically by 2:1 or 3:1, and filters tremor. Until Intuitive's da Vinci 5 added force sensing in 2024, the console gave the surgeon no sense of touch at all, and tissue tension had to be judged by watching the tissue deform.

The wrist and the stereo view make fine suturing deep in the pelvis practical, which is why radical prostatectomy converted first and low rectal resection followed. Sitting at a console with the image and the hands aligned is also easier on the surgeon than standing over a laparoscopic stack for four hours. The weaknesses are cost and evidence. Docking the arms adds operating room time on top of a case that already costs more in consumables, and for common operations the advantage over laparoscopy is small: randomized comparisons in benign hysterectomy, cholecystectomy and colon resection generally show similar complication rates and length of stay at higher cost. The FDA has not granted marketing authorization for any robotically-assisted surgical device specifically for the prevention or treatment of cancer, and it issued a safety warning in 2019 about using them in cancer surgery, after a randomized trial found worse survival with minimally invasive radical hysterectomy for cervical cancer.

Pick the robot when the operation happens in a confined space and involves reconstruction: radical prostatectomy, low rectal resection, complex ventral hernia repair, revisional foregut surgery. If the case is a straightforward cholecystectomy or a benign hysterectomy, laparoscopy is cheaper and the published outcomes are about the same, so use the robot only when that particular surgeon is genuinely faster on it. If you are evaluating a purchase, model utilization before anything else, since a system doing 100 cases a year carries a fixed cost per case roughly three times that of one doing 300. If the operation is open by necessity or the patient cannot tolerate insufflation, none of this applies. And if the reason for buying is that a competing hospital advertises one, that is a marketing decision and should be judged as one.

System price $0.7–3.1M · instruments and accessories $900–3,700 per procedure · service contract $95,000–225,000 per system per year · 11,106 da Vinci systems installed at the end of 2025 · about 3,153,000 da Vinci procedures in 2025, roughly 280 per system · instruments, accessories and service were $7.59B of Intuitive's $10.06B 2025 revenue

Intuitive's da Vinci Xi, SP and 5 systems plus the Ion bronchoscopy platform; Medtronic Hugo, cleared by 510(k) in December 2025; CMR Surgical Versius, granted a De Novo in October 2024; Johnson & Johnson Ottava, granted a De Novo in July 2026; Asensus Senhance, bought by Karl Storz in 2024.

The system is the smaller half of the business. A da Vinci lists between $0.7M and $3.1M depending on model, configuration and country, and Intuitive sold $2.47B of systems in 2025 against $6.02B of instruments and accessories and $1.57B of service, so $7.59B of the $10.06B total came from hardware that was already installed. Instruments carry a use counter and stop working after a set number of procedures, which is what converts a surgical program into a subscription at $900–3,700 per procedure, and service runs $95,000–225,000 per system per year once the first year included in the purchase price expires. Utilization decides whether that works for the buyer: 3,153,000 procedures across 11,106 systems is about 280 per system per year (derived from those two figures), and a $150,000 service contract spread over 280 cases adds roughly $540 per case before any instrument is opened. The hospital usually cannot bill for the difference, because a robotic hysterectomy and a laparoscopic one fall in the same DRG, so the incremental cost comes out of the hospital's margin and is bought back in referrals. Patent expiry changed less than expected: the original da Vinci claims ran out years ago, and Intuitive still placed 1,721 systems in 2025 and grew procedures 18%, because the barrier is trained surgeons, a stocked instrument room and a signed service contract rather than a claim chart. New entrants have moved slowly through the FDA, with Versius granted a De Novo in October 2024, Hugo cleared in December 2025 and Ottava granted a De Novo in July 2026. The thing to watch is price competition on consumables, since that is where the $6B sits.

2 minda Vinci® Surgery - How It WorksSamaritan Health Services · 500k+ views

6 minInside the world of a robotic surgeonFreethink · 500k+ views

Computer-Assisted Surgical Systems (FDA) · Intuitive Surgical Annual Report on Form 10-K, fiscal year 2025 (U.S. Securities and Exchange Commission)

These robots hold a trajectory rather than a pair of hands. The surgeon plans cuts or screw paths on a CT or X-ray model of the patient's own bone, the system registers that model to the physical anatomy by touching known landmarks with a tracked probe, and an optical camera then follows reflective arrays screwed into the bone so the plan stays attached to the patient as they move. In a knee replacement the arm enforces a haptic boundary: the surgeon drives the burr, and the arm stiffens and stops it at the edge of the planned resection. In spine surgery the arm holds a rigid tube on the planned pedicle trajectory and the surgeon drills through it. Bone is the reason this works at all, because bone is rigid, so a registration made at the start of the case is still valid an hour later, which is not true of soft tissue. Reported pedicle screw placement accuracy averages about 97% on older spine robots and 99% on the newer ones with an optical camera and a multijointed arm.

Accuracy and repeatability are real. Robotic knee replacement puts components within a degree or two of the plan where jigs and a saw give a few degrees of spread, and a spine robot places screws into pedicles that a freehand surgeon would decline to instrument, with far less fluoroscopy for the whole team. The weakness is that better geometry has not yet shown up as better long-term outcomes. Trials find modestly faster early recovery and shorter stays after robotic total knee replacement, but joint registries have not so far demonstrated a lower revision rate at five to ten years, which is the number that would settle the argument. Setup also costs time and dose: array pins are extra incisions that can fracture bone, registration failures force a fallback to the conventional technique mid-case, and CT-based planning means an extra scan before surgery.

Use a spine robot when the screws are hard: deformity, revision, obese patients where landmarks are buried, or percutaneous fixation where you would otherwise be shooting a lot of fluoroscopy. Use a knee robot when you are doing partial knee replacement, where component position has to be exact and the volume is low enough that a jig set never gets familiar. If you are doing 300 routine primary total knees a year with good results, the robot will buy you alignment precision and not much else, so treat it as a workflow and marketing purchase rather than a clinical one. If your surgeons will not accept the added setup time, do not buy it, because a robot used on a third of cases carries the same capital and service cost as one used on all of them. For simple stereotactic tasks like a single deep brain stimulation lead, a conventional frame is still accurate to roughly a millimeter and costs almost nothing.

Pedicle screw accuracy about 97% on older spine robots and 99% on newer ones · robotic knee components within 1–2° of plan against a few degrees for jigs · capital cost roughly $0.8–1.5M for an arm-based system · disposable arrays, pins and cutting accessories a few hundred to about $1,000 per case · over 2 million Mako procedures performed worldwide to date, of which more than 1 million were total knees · Stryker knee and hip implant sales of $2.66B and $1.87B in 2025

Stryker Mako, launched as Mako 4 in 2025 and available in more than 45 countries; Zimmer Biomet ROSA Knee and ROSA Spine; Smith+Nephew CORI, a handheld image-free system; J&J DePuy Synthes VELYS; Medtronic Mazor X, which traces back to the SpineAssist robot the FDA authorized in 2004; Globus Medical ExcelsiusGPS; Renishaw Neuromate for stereoelectroencephalography electrode placement.

The money here works the opposite way from soft-tissue robotics. The per-case consumable is small, a few hundred to about $1,000 of arrays, pins and cutting accessories, so the robot is not the annuity. The implant is. Stryker does not report Mako revenue as a line item at all; what shows up is $2.66B of knee and $1.87B of hip implants in 2025, and the robot exists to make sure those implants are the ones in the room. More than 2 million Mako procedures have been done worldwide to date, over 1 million of them total knees, and every one of those cases came with a Stryker implant. That is why systems get placed on consignment or bundled into multi-year implant commitments rather than sold outright at the $0.8–1.5M capital price, and why a hospital that installs one vendor's arm has effectively chosen that vendor's implants for the life of the machine. The hospital gets no extra payment for using it, since a total knee falls in the same DRG either way and computer-assisted navigation carries a Category III code that generally pays nothing extra, so the case for buying is throughput, surgeon recruitment, and local advertising. Utilization is the whole argument on the buyer's side, because capital and service cost the same whether the arm runs 60 cases a year or 400. Spine is the larger long-term market, with roughly 4.83 million spinal operations performed globally each year, which is why Medtronic bought Mazor rather than building a spine robot from scratch.

4 minTotal Knee Replacement: Mako Robotic Surgery Explained | Tufts MedicineTufts Medicine · 50k+ views

3 minRobotic spine surgery with Mazor XCorewell Health in Southeast Michigan · 100k+ views

A Comparison of Spinal Robotic Systems and Pedicle Screw Accuracy Rates: Review of Literature and Meta-Analysis (Asian Journal of Neurosurgery) · Stryker Corporation Annual Report on Form 10-K, fiscal year 2025 (U.S. Securities and Exchange Commission)

Everything in this class goes in through a needle puncture in an artery or vein. A modern catheter is a composite tube built over a mandrel: a PTFE liner for lubricity, a braid or coil of stainless steel or nitinol for torque and kink resistance, and a Pebax outer jacket whose stiffness steps down along the length so the shaft pushes at the hub and stays floppy at the tip, all reflowed together under heat shrink. Coronary work runs through a 6 French guide catheter, about 2 mm across, usually from the wrist. A drug-eluting stent is a laser-cut cobalt-chromium or platinum-chromium tube with struts 60–80 µm thick, crimped onto a balloon and expanded at 12–16 atmospheres against the vessel wall, carrying a polymer coating that releases sirolimus or everolimus over several weeks to suppress the scar tissue that would otherwise close the vessel again. A transcatheter aortic valve is bovine pericardium sewn into a balloon-expandable cobalt-chromium frame or a self-expanding nitinol frame, crimped to about 6 mm, pushed up the aorta through a 14 French sheath, and deployed inside the diseased native valve, which gets crushed against the aortic wall rather than removed.

Avoiding a sternotomy changes who can be treated: TAVR patients are typically discharged the next day, and many go home the same day, where surgical valve replacement means a week in hospital and months of recovery. The trade-offs are specific and well documented. Pushing a frame through the aortic annulus presses on the conduction system, so 6–17% of TAVR patients need a permanent pacemaker depending on valve design; paravalvular leak is more common than after surgery; and nobody has 20-year durability data on a crimped bioprosthetic valve, which matters enormously for a 60-year-old and not at all for an 85-year-old. Against that, unadjusted 30-day mortality in the US registry sits at 2.2–2.4% in a population whose median age is 79. Coronary stents have their own version of this: the drug that stops restenosis also delays healing over the struts, so patients take dual antiplatelet therapy for 6–12 months and accept a bleeding risk in exchange for a stent thrombosis risk of well under 1% a year.

If the patient is over about 75 with severe symptomatic aortic stenosis and femoral access, TAVR is the default and surgery needs a specific reason. If the patient is under 65 with a long life expectancy, favor surgery, because a surgical valve has decades of durability data and leaves cleaner options for the second operation. If the anatomy is bicuspid, the annulus is very large or small, or the patient needs bypass grafts at the same time, take them to surgery. For coronary disease, stent focal lesions and send left main or three-vessel disease with diabetes to bypass surgery, since the trials favor surgery there. If you are building a program, check the coverage rules before the catheter order, because Medicare's national coverage determination sets volume floors that a small hospital cannot meet.

FDA approvals walked from inoperable patients in November 2011 to low surgical risk in 2019 · trial populations went from 11.6% STS predicted mortality in PARTNER 1B to 1.9% in PARTNER 3 · 210,495 TAVRs at 786 US hospitals between January 2019 and March 2022, roughly 65,000 a year · unadjusted 30-day mortality 2.2–2.4% · pacemaker required in 6–17% of TAVR patients · drug-eluting stent struts 60–80 µm, valve delivered through a 14 French sheath

Edwards SAPIEN, approved as PMA P100041 in November 2011, and Medtronic CoreValve and Evolut; Abbott MitraClip (2013) and TriClip G4 (2024) for mitral and tricuspid leaflet repair; Boston Scientific Watchman (2015) for left atrial appendage closure; Abbott Xience and Boston Scientific Synergy drug-eluting stents.

TAVR is the cleanest example in medical devices of a company buying its own market with randomized trials. PARTNER 1B in 2010 randomized 358 inoperable patients against medical therapy and won on one-year survival, which got the SAPIEN valve approved for inoperable patients in November 2011, a population of a few thousand a year in the US. PARTNER 1A and the CoreValve pivotal trials moved it to high risk by 2014, PARTNER 2A and SURTAVI to intermediate risk in 2016 and 2017, and PARTNER 3 and Evolut Low Risk to low surgical risk in 2019, at which point the reimbursed population was essentially everyone with severe symptomatic aortic stenosis. Each trial cost tens of millions of dollars and returned a multiple of the previous addressable market: PARTNER 1B enrolled patients with an 11.6% predicted surgical mortality, PARTNER 3 enrolled patients at 1.9%, and the median real-world risk score in the registry now sits at 3.3% in between them. The unit economics hold up because the valve is expensive, roughly $30,000, inside a DRG paying the hospital somewhere in the $40,000–50,000 range, so hospital margin depends almost entirely on getting the patient out the next day. Coverage rules concentrate the volume: CMS requires a program to have at least 50 open heart operations and 300 percutaneous coronary interventions a year before it can start, which is a large part of why 786 hospitals accounted for all 210,495 US TAVRs performed between January 2019 and March 2022, roughly 65,000 a year. Coronary stents show what happens when the opposite is true. Four vendors sell an equivalent device at roughly $1,000–1,500 with no coverage barrier, prices grind down every year, and that is why Edwards, Abbott and Boston Scientific all moved their development money into structural heart implants where the premarket approval requirement keeps the field small.

3 minTAVR - Transcatheter Aortic Valve Replacement AnimationRWJBarnabas Health · 1m+ views

4 minEdwards SAPIEN 3 TAVR (Transfemoral Implant Procedure Animation)HeartValveSurgery.com · 100k+ views

NCA - Transcatheter Aortic Valve Replacement (TAVR) (CAG-00430R) - Decision Memo (Centers for Medicare & Medicaid Services) · Trends in Transcatheter Aortic Valve Replacement Outcomes: Insights From the STS/ACC TVT Registry (JAMA Cardiology)

Catheter ablation for atrial fibrillation destroys a ring of tissue around each pulmonary vein so the electrical triggers inside the vein can no longer reach the atrium. Three energy sources do the job. Radiofrequency drives 25–30 W through an irrigated tip and heats tissue resistively past about 50 °C, one lesion at a time, with contact force sensing to tell the operator the tip is really pressed against the wall. Cryoablation expands nitrous oxide inside a 28 mm balloon wedged into the vein ostium and freezes the whole ring in one shot. Pulsed field ablation is not thermal at all: it delivers microsecond biphasic pulses at around 1,500 V that open permanent pores in cell membranes, and because heart muscle electroporates at a lower field strength than esophagus, phrenic nerve or vessel wall, a dose that kills myocardium leaves the structures behind it intact. That threshold difference is what PFA is sold on.

Thermal ablation works and has twenty years of practice behind it, but heat and cold cannot tell which tissue they are in. Reported rates with radiofrequency and cryoablation run about 0.5% pulmonary vein stenosis, 2.7% phrenic nerve palsy, 0.15–0.5% stroke or transient ischemic attack, and 0.02–0.15% atrioesophageal fistula, and that last one kills 83–100% of the patients who get it without surgical repair. Selectivity removes most of that list: PFA registries report acute isolation of the veins above 99%, no pulmonary vein narrowing at all against mild to severe narrowing in roughly 12% of a radiofrequency cohort, and no phrenic nerve injury. PFA brings its own failure modes. Intravascular hemolysis rises with the number of applications and with poor catheter contact and occasionally causes acute kidney injury, reversible coronary spasm occurs when energy is delivered near a coronary artery, and pericardial tamponade remains the leading major complication inside an overall major complication rate of 1.7–1.9%.

For straightforward pulmonary vein isolation, use PFA if your lab has it, because efficacy is comparable and the complications it removes are the ones that kill people. Use point-by-point radiofrequency when the lesions are away from the veins, in atypical flutter, mitral isthmus lines and ventricular tachycardia, where titratable power and a mapping-integrated tip still matter more than speed. Keep PFA away from a coronary artery, or accept that you may induce spasm and need to treat it. Cryoablation is still a reasonable single-shot choice if you already own the console and your paroxysmal results are good, since head-to-head data put it level with radiofrequency on efficacy and ahead on procedure time. Set expectations by substrate rather than by energy source: one-year freedom from atrial fibrillation is 80–82% in paroxysmal patients and 66–72% in persistent ones in the best real-world registries, and no energy source changes that much.

Radiofrequency point-by-point isolation averages about 95 minutes · PFA registry procedure time 58–61 minutes with acute isolation above 99% · thermal complication rates 0.5% pulmonary vein stenosis, 2.7% phrenic nerve palsy, 0.02–0.15% atrioesophageal fistula · PFA major complications 1.7–1.9% · one-year freedom from AF 80–82% paroxysmal, 66–72% persistent · PFA pulses at roughly 1,500 V, radiofrequency at 25–30 W

Boston Scientific Farapulse, PMA-approved January 2024 and extended to persistent atrial fibrillation in 2025; Medtronic PulseSelect, approved December 2023 under PMA P230017 on the strength of the PULSED AF study; Medtronic Affera Sphere-9, which combines focal PFA and radiofrequency on one tip; Johnson & Johnson Varipulse; Medtronic Arctic Front cryoballoon; contact-force radiofrequency catheters used with the CARTO and EnSite mapping systems.

PFA converted an existing procedure market unusually fast. Boston Scientific's electrophysiology unit reported $800M of sales in 2023, $1.9B in 2024 and $3.3B in 2025, and the company states that PFA is now the predominant component of that business, all of it from converting procedures that were already being done rather than from new patients. Nothing about the payment changed to cause it: a hospital gets the same DRG or outpatient rate for an ablation whatever the energy source, so the value to the buyer is throughput and avoided complications. A 58–61 minute case instead of a 95 minute one is an extra case or two per lab per day, and never having an atrioesophageal fistula removes a low-probability, very high-cost event from the ledger. The vendors capture that in the disposable, which is single-use and priced above the thermal catheters it replaced, on top of a mapping system placed as capital in the same razor-and-blade pattern as surgical robotics. The barrier to entry is the Class III premarket approval: PULSED AF enrolled 150 paroxysmal and 150 persistent patients and reported 66.2% twelve-month efficacy in the paroxysmal arm, which is a multi-year, tens-of-millions-of-dollars pivotal program, and it is why the competitors here are all large public companies rather than startups.

2 minPulsed Field Ablation (PFA): A New, Safer Option for AFib | Duke HealthDuke Health · 50k+ views

4 minDeep Dive into FARAPULSE™: PFA System Workflow AnimationBoston Scientific Cardiology · 100k+ views

Summary of Safety and Effectiveness Data: PulseSelect Pulsed Field Ablation System, PMA P230017 (FDA) · Pulsed Field Ablation of Atrial Fibrillation: A Novel Technology for Safer and Faster Ablation (Biomedicines)

Navigation puts a preoperative CT or MRI into the same coordinate frame as the patient on the table, then shows where the instrument tip sits inside that volume. Two tracking technologies do the measuring. Optical tracking uses a stereo infrared camera watching retroreflective spheres on rigid bodies clamped to the instrument and to the patient, and is good to a few tenths of a millimeter as long as nothing blocks the line of sight. Electromagnetic tracking puts a field generator near the patient and a tiny coil in the instrument tip, works around corners and inside the body, and is good to roughly half a millimeter to a millimeter unless ferromagnetic metal distorts the field. Neither tracker is the limiting error. The limiting error is registration, which means matching the image volume to the physical patient by touching fiducials or anatomic landmarks or by surface-matching a scanned contour, and everything after that inherits whatever error the registration started with.

Navigation lets a surgeon operate confidently where the anatomy is hidden or destroyed, which is why revision sinus surgery, skull base work and percutaneous spinal fixation adopted it early, and it cuts fluoroscopy dose for the whole room. The failure mode is that navigation error is silent. The screen always draws a crosshair somewhere, whether or not the registration is still valid, and in brain surgery it stops being valid almost immediately: target registration error measured against preoperative MRI runs about 5.9 mm before the dura is opened, 6.2 mm after dural reflection and 7.5 mm after resection has started. Reregistering to an intraoperative ultrasound sweep pulls that back to about 2.7 mm early in the case and 4.2 mm after dural opening, and one series reported mean extent of resection rising from 75.9% to 86.8% with intraoperative ultrasound in use. Bumping the reference array, or moving the patient relative to it, invalidates everything without any warning on the display.

Use navigation when the target is not visible and the consequences of missing it are severe: pedicle screws in deformed spines, skull base tumors, revision sinus surgery, biopsy of a deep lesion. Verify the registration against a known landmark before you trust it, and re-verify after any bump to the array, because the display gives you no other way to know. If the anatomy moves during the operation, stop trusting the preoperative scan and update it, with intraoperative ultrasound if you want cheap and fast or intraoperative MRI if you need the same contrast you planned on. If all you need is confirmation that the hardware ended up where you intended, a single intraoperative 3D scan at the end of the case is much cheaper than navigating the whole procedure. Do not buy an intraoperative MRI suite for volume you do not have, since the room is unavailable for ordinary surgery whenever the magnet is in use.

Optical trackers accurate to a few tenths of a millimeter, electromagnetic trackers to roughly 0.5–1 mm · registration error dominates, not tracker error · target registration error against preoperative MRI about 5.9 mm pre-dura, 6.2 mm post-dura, 7.5 mm after resection begins · intraoperative ultrasound reregistration returns it to about 2.7 mm · extent of resection 86.8% with intraoperative ultrasound against 75.9% without · navigation station a few hundred thousand dollars, an intraoperative MRI suite several million plus the shielded room

Medtronic StealthStation with the O-arm mobile CT; Brainlab Curve and Cirq; Stryker Q Guidance; 7D Surgical's machine-vision registration using visible light instead of a scan; NDI Polaris optical and Aurora electromagnetic trackers, which sit inside most of the above; Augmedics xvision, a head-mounted display that draws the trajectory over the surgeon's view of the spine.

Navigation is rarely bought as a standalone product. Implant vendors place stations and hand out the disposable arrays and clamps because the navigation only works with their instruments, so the station is a channel lock on the screws and cages, in the same pattern as orthopedic robotics. A station that would list at a few hundred thousand dollars often arrives at no separate charge, which is why a hospital switching implant vendors usually finds it is switching navigation systems too. The buyer gets no revenue for using it: the US add-on codes for stereotactic computer-assisted navigation pay the surgeon a small amount and leave the hospital's inpatient DRG payment unchanged, so the business case is fewer revisions, less radiation, and shorter cases. Intraoperative imaging is a different purchase, because it is real capital with an opportunity cost. A mobile intraoperative CT is roughly the price of a fixed imaging room and gets used across many services, while an intraoperative MRI suite costs several million dollars plus a shielded operating room and needs a steady flow of glioma and epilepsy cases to be worth the floor space. Intraoperative ultrasound is the outlier: a probe and a tracking bracket cost a fraction of either, and they recover most of the accuracy the preoperative scan loses.

5 minSurgical Navigation System | Biomedical Engineers TV |Biomedical Engineers TV · 10k+ views

2 minUsing Intraoperative MRI During Brain SurgeryUTMBHealth · 5k+ views

Brain Shift in Neuronavigation of Brain Tumors: An Updated Review of Intra-Operative Ultrasound Applications (Frontiers in Oncology) · Performance of image guided navigation in laparoscopic liver surgery – A systematic review (Surgical Oncology)

A clinical linear accelerator pushes electrons down a copper waveguide driven by a magnetron or klystron at about 3 GHz, reaching 6–18 MeV in roughly a meter. For photon treatment those electrons hit a tungsten target and make bremsstrahlung X-rays; a multileaf collimator of 60 to 160 tungsten leaves, each around 5 mm wide at the isocenter, shapes the beam, and the gantry rotates around the patient so dose piles up at the target while the entrance dose is spread over many directions. Intensity-modulated and volumetric arc techniques move the leaves and vary the dose rate continuously during that rotation, which is how a concave dose distribution gets wrapped around a spinal cord. A conventional course is 2 Gy a day to 60–80 Gy over 30–40 weekday fractions, though hypofractionated schedules now deliver 26 Gy in five fractions for breast cancer and about 36 Gy in five for prostate. Proton therapy replaces the linac with a cyclotron or synchrotron running to 70–250 MeV, and the Bragg peak puts most of the energy at a depth set by the beam energy with nothing beyond it, so there is no exit dose at all. The IAEA's DIRAC directory lists more than 8,500 radiotherapy centers and over 20,000 treatment units worldwide, and the overwhelming majority of those units are photon linacs.

The photon linac is the workhorse, it treats roughly half of all cancer patients at some point, and modern planning shapes dose tightly enough that the technical argument is mostly settled. Its weaknesses are the beam going in and the beam coming out, plus the fact that the patient and the tumor are in slightly different places every day. MR-linacs address that directly by imaging soft tissue during treatment and replanning on the spot, at two to four times the capital cost and with treatment slots that run 45–60 minutes instead of 10–15, which halves the number of patients a machine can treat. Protons remove exit dose entirely, which is better dosimetry by inspection, but the clinical evidence is strong mainly in pediatric, skull base and ocular tumors, and the field still plans with a fixed relative biological effectiveness of 1.1 that turned out to underestimate central nervous system damage in some pediatric medulloblastoma cases.

Treat almost everything on a photon linac, and hypofractionate wherever the trials support it, because five visits are better for the patient than twenty even though the department is paid less for them. Use protons where the tissue behind the target cannot take any dose and the patient has decades of life left: pediatric tumors, skull base, uveal melanoma, reirradiation. Do not build a proton case around prostate cancer, since private payers deny it and the randomized evidence does not carry the argument. Consider an MR-linac when your case mix is genuinely adaptive, meaning pancreas, liver and other targets that move next to bowel, and only if you can fill the schedule at an hour a patient. If the question is one to five brain metastases, a Gamma Knife or a linac-based radiosurgery platform is usually the cheaper way to the same plan.

Electrons to 6–18 MeV, RF at roughly 3 GHz, multileaf collimator leaves about 5 mm wide at isocenter · conventional course 2 Gy × 30–40 fractions to 60–80 Gy, hypofractionated 26 Gy in 5 for breast · protons at 70–250 MeV with no exit dose · a linac runs a few million dollars plus a shielded vault and typically treats 25–35 patients a day, an MR-linac costs two to four times as much, a multi-room proton center $100–200M · more than 8,500 radiotherapy centers and 20,000 units worldwide · 30 US proton centers operating in 2019 with 10 more under construction

Varian TrueBeam and Halcyon, from the business Siemens Healthineers bought in 2021 for about $16B; Elekta Versa HD and the Unity 1.5 T MR-linac; ViewRay's MRIdian, whose maker filed for Chapter 11 in August 2023; Accuray CyberKnife and TomoTherapy; Elekta Leksell Gamma Knife; proton systems from IBA, Varian, Mevion and Hitachi.

US radiation oncology is paid per fraction, which means the clinical improvement of the last decade cuts provider revenue. A breast course that used to be 25 visits is now five, and the department's payment falls roughly in proportion while its fixed costs do not, so the machine has to be filled with more patients to stand still. Medicare's answer was the Radiation Oncology Model, a prospective episode payment that would have made fraction count irrelevant, and Congress blocked its start twice before CMS delayed it to a date still to be determined, so it has never run. The base machine is not the problem: a linac runs a few million dollars plus a shielded concrete vault, and it typically treats 25 to 35 patients a day for a decade or more. Proton therapy is where the arithmetic has failed most visibly. A multi-room center costs $100–200M, 30 were operating in the US in 2019 with 10 more under construction, and Scripps in San Diego opened in 2014 against a $220M investment and a plan to treat 2,000 patients a year, treated about 1,400, and filed for bankruptcy protection in 2017; Indiana University closed its center for the same reason. Single-room systems cut the entry price by most of an order of magnitude, but they do not fix the underlying problem, which is that the high-volume simple indication that would fill a center is prostate cancer and payers will not cover protons for it. ViewRay shows what happens when the physics works and the billing does not: the MRIdian was a technically strong machine with real adaptive capability, it had no separate payment for the MR guidance, its hour-long slots cut throughput against a conventional linac, and the company filed Chapter 11 in 2023.

8 minHow a Linear Accelerator works – ElektaElekta · 500k+ views

7 minHow does proton radiation therapy work?Fermilab · 100k+ views

Radiation Oncology Model (Centers for Medicare & Medicaid Services) · Three discipline collaborative radiation therapy (3DCRT) special debate: The United States should build additional proton therapy facilities (Journal of Applied Clinical Medical Physics)

A pacemaker is a titanium can holding a lithium battery, a pulse generator and a small computer, wired to one or two insulated leads that are threaded down the subclavian vein and fixed into the right atrium and right ventricle. It watches the heart's own electrical activity and fires a pulse of roughly 2–3 V lasting about 0.4 ms whenever the intrinsic rate drops below a programmed floor. An implantable cardioverter-defibrillator does all of that and adds a high-voltage capacitor that charges off the same small battery over about 10 seconds and then dumps 30–40 J across the heart to stop ventricular fibrillation. Cardiac resynchronization therapy adds a third lead through the coronary sinus onto the outside of the left ventricle so both ventricles are paced together, which is used in heart failure with a wide QRS and a low ejection fraction. All three go in under local anesthetic in a cath lab in 45 to 90 minutes, with the can in a pocket under the collarbone. Two newer shapes avoid the vein entirely: a leadless pacemaker of about 0.8 cm³ is delivered up the femoral vein and tined directly into the right ventricular wall, and a subcutaneous defibrillator runs its electrode under the skin alongside the sternum instead of inside the heart.

Pacing for symptomatic bradycardia works about as well as any treatment in medicine: the rhythm is disabling or fatal untreated, and a device that costs a few thousand dollars corrects it for a decade at a time. The weak part is almost never the electronics, it is the lead. A transvenous lead flexes with every heartbeat, on the order of 35 million cycles a year, and insulation abrasion and conductor fracture accumulate at roughly 0.2–2% per year depending on the design, so a patient implanted at 55 should expect lead work later. Batteries are the other clock, at 8–12 years for a pacemaker and 5–8 for a defibrillator, and every replacement means reopening the pocket with another 1–2% infection risk. Once a lead has been in place more than a year it is bound into fibrous tissue, so removing it takes laser or mechanical cutting sheaths and carries roughly a 1–2% major complication rate and a few tenths of a percent procedural mortality, which is why abandoning a dead lead in place is often the safer choice and why the abandoned lead then becomes the reason the patient cannot be scanned.

If the problem is a slow rhythm, a pacemaker is the answer and the only real questions are how many leads and whether to go leadless. Choose a leadless pacemaker when the patient needs ventricular pacing only, has poor venous access, or has already had a device infection, and accept that retrieval after several years is not a routine procedure. Choose a subcutaneous defibrillator when the patient needs shock therapy but no pacing and you want the veins left clean, which is the standard argument in a young patient with hypertrophic cardiomyopathy, and choose a transvenous defibrillator when antitachycardia pacing or resynchronization may be needed later. If the patient is likely to need an MRI, confirm MR conditionality for the specific generator-and-lead combination rather than for the can alone, and check for abandoned leads, which usually rule the scan out. If the patient is 85 and pacing-dependent, weigh battery life against how many more generator changes it implies, since each one is another procedure with its own infection risk.

Pacing pulses of about 2–3 V for 0.4 ms, defibrillation shocks of 30–40 J · pacemaker batteries last 8–12 years and defibrillators 5–8 · lead failure roughly 0.2–2% per year · pocket infection 1–2% per implant, extraction major complications 1–2% · leadless pacemaker about 0.8 cm³ · pacemaker generator roughly $4,000–8,000 and a CRT-D roughly $20,000–30,000 · more than a million pacemakers implanted worldwide each year.

Medtronic Micra AV2 and Abbott Aveir leadless pacemakers, the latter approved in a dual-chamber version in 2023; Boston Scientific EMBLEM subcutaneous defibrillator; Medtronic's Sprint Fidelis defibrillator lead, withdrawn in 2007 after roughly 268,000 were implanted, and Abbott's Riata, recalled in 2011 after insulation wear let the conductors work their way outside the lead body.

The device is a small part of the bill. A pacemaker generator sells to hospitals for roughly $4,000–8,000 and a CRT-D for roughly $20,000–30,000, while the total cost of the implant admission runs several times the device price once the lab, the imaging, the implanting physician and the overnight stay are counted. Medicare pays the hospital one fixed DRG amount for that admission, so the hospital's margin is the gap between that payment and the price it negotiated for the can, which is why large systems standardize on one or two vendors and run implant contracts by volume. More than a million pacemakers go in worldwide each year, and four vendors — Medtronic, Abbott, Boston Scientific and Biotronik — hold most of a cardiac rhythm management market of roughly $12–15 billion, which makes this the largest active-implant category by installed base. Unit growth is slow because the indication tracks the population over 70, so most of the vendors' revenue growth comes from mix: a leadless pacemaker sells for several times a conventional one, and a defibrillator for several times a pacemaker again. Replacements are a predictable revenue stream, since every implant schedules another procedure 5 to 12 years out, and that is also why lead reliability shows up in the accounts. A recalled lead converts a routine generator change into an extraction, and it moves the cost from a 45-minute pocket procedure to a case with a surgical backup team on standby.

3 minImplantable Cardioverter Defibrillator - How it worksGebrüder Betz Medical Animation · 100k+ views

3 minMicra TPS Implant Procedure AnimationMedtronic Cardiac and Vascular · 100k+ views

Pacemakers - What Are Pacemakers? (NHLBI, NIH) · Leadless pacing: a comprehensive review (European Heart Journal)

Neuromodulation puts electrodes on or near a nerve and runs a pulse train through them to change how that circuit behaves. Deep brain stimulation drops two leads through burr holes into the subthalamic nucleus or globus pallidus, each carrying four to eight contacts, and drives them at around 130 Hz with 60 microsecond pulses from a generator in the chest; it suppresses Parkinsonian tremor and rigidity well enough to cut medication dose, though what the stimulation does to the circuit is still argued over. Spinal cord stimulation puts an 8- to 16-contact lead in the dorsal epidural space around T8–T10 for chronic back and leg pain, either at 40–60 Hz where the patient feels a tingle over the painful area, or at 10 kHz where they feel nothing. Vagus nerve stimulation wraps a helical electrode around the left cervical vagus for drug-resistant epilepsy and treatment-resistant depression, and sacral neuromodulation runs a lead into the S3 foramen for overactive bladder, urinary retention and fecal incontinence. Sacral is the one implant on this sheet with a screening step built into the pathway: a percutaneous test lead runs for one to two weeks, and the permanent device only goes in if symptoms improve by more than half. Newer systems sense as well as stimulate, with Medtronic's Percept reading beta-band local field potentials off the same deep brain lead and adjusting amplitude, and Saluda's Evoke measuring the evoked compound action potential in the cord and holding it at a set point.

Deep brain stimulation for advanced Parkinson's and essential tremor has randomized evidence behind it and a track record going back to the late 1990s, and unlike an ablation it is adjustable and reversible, which is why it displaced lesioning almost entirely. Its failure modes are the ordinary implant ones: lead migration or fracture, infection at roughly 2–5% at the chest pocket or the burr-hole site, and a battery replacement every three to five years for a primary cell or every nine to fifteen for a rechargeable one. Spinal cord stimulation is where the honest reading is uncomfortable. The 2023 Cochrane review concluded that it probably does not improve chronic low back pain compared with placebo stimulation, sham-controlled trials since have mostly agreed, and 15–30% of implanted systems are removed within five years, usually because the effect faded. Pain societies published detailed objections to that review's scope and patient selection and the argument is not settled, but a therapy sold on a tingle the patient can feel, tested mostly in unblinded trials, is exactly the setup where placebo response is largest. Reporting on FDA's adverse-event database in 2018 counted more than 80,000 injury reports for spinal cord stimulators over the previous decade, among the highest totals of any device class.

If a Parkinson's patient still responds to levodopa but spends hours a day off or dyskinetic, deep brain stimulation is the standard next step, and it is better considered early than after the disease has progressed far enough to make the surgery risky. If the diagnosis is essential tremor and the patient is elderly or on anticoagulation, focused ultrasound thalamotomy avoids an implant and its replacements, at the cost of being irreversible and usually one-sided. For chronic pain, take the trial seriously as a gate rather than a formality: if the one-week percutaneous trial does not produce a clear response, do not implant, because that trial is the only screen the pathway has. Be skeptical of any spinal cord stimulation business case built on unblinded outcome data, and ask for the explant rate at three and five years rather than the responder rate at three months. If the indication is overactive bladder that has failed drugs, sacral neuromodulation and botulinum toxin injections are the two real options, and the choice usually turns on whether the patient would rather accept an implant or repeat injections every six months.

Deep brain stimulation at roughly 130 Hz with 60 microsecond pulses · primary-cell generators last 3–5 years, rechargeable 9–15 · deep brain stimulation infection roughly 2–5% · spinal leads of 8–16 contacts at T8–T10, run at 40–60 Hz or 10 kHz · 15–30% of spinal systems explanted within five years · more than 80,000 spinal cord stimulator injury reports to FDA over a decade · neuromodulation market roughly $7–8 billion a year.

Medtronic Percept with BrainSense adaptive stimulation; Abbott Infinity and Boston Scientific Vercise deep brain systems; Nevro Senza HF10 at 10 kHz and Saluda Evoke, the first closed-loop spinal cord stimulator cleared in the US; LivaNova VNS Therapy for epilepsy and treatment-resistant depression; Medtronic InterStim and Axonics sacral neuromodulation; NeuroPace RNS, which detects seizure onset from intracranial electrodes and stimulates in response.

A spinal cord stimulation system sells for roughly $20,000–30,000 and the total US commercial cost of implanting one runs $40,000–60,000, which for about a decade made it one of the better-paying procedures in interventional pain. That is also what turned it into a target: payers now require documented conservative care, a psychological evaluation and a successful trial before approving the permanent implant, and several narrowed coverage after the sham-controlled trials read out. Deep brain stimulation runs roughly $20,000–30,000 in hardware and $50,000–100,000 for the whole bilateral episode in the US, and it has established CPT codes and settled Medicare coverage, which is the main reason it is a stable business while the pain segment is not. Vagus nerve stimulation for depression shows the other side of that: FDA approved it in 2005, CMS declined to cover it, and it took until 2019 for coverage to arrive at all, and then only for patients enrolled in a study. The neuromodulation market is roughly $7–8 billion a year, spinal cord stimulation is the largest single segment, and that segment stopped growing around 2019 and has been flat to down since, which is unusual for a device category and traces directly to the evidence. Rechargeable generators cut both ways for the vendors: they removed a replacement procedure every three to five years, which had been predictable revenue, and they made the device easier to sell. Sacral neuromodulation is the clearest recent share shift, with Axonics taking a large piece of Medtronic's position in about five years mostly by shipping a rechargeable device first.

2 minDeep Brain Stimulation: How It WorksThe Wall Street Journal · 100k+ views

2 minSpine Stimulator for PainMayo Clinic · 100k+ views

4 minMore Than 80,000 Spinal Cord Stimulator Injury Reports Filed With FDA | NBC Nightly NewsNBC News · 100k+ views

Deep Brain Stimulation (DBS) (National Institute of Neurological Disorders and Stroke) · Spinal cord stimulation for low back pain (Cochrane Database of Systematic Reviews)

Both devices replace a dead sensory transducer with an electrode array that stimulates the surviving neurons behind it. A cochlear implant threads a 12- to 22-contact array into the scala tympani, roughly 16 to 31 mm around the cochlear spiral, and stimulates spiral ganglion neurons directly, bypassing the hair cells that no longer respond. The external processor splits the microphone signal into frequency bands and sends each band's energy to the electrode sitting where a healthy cochlea would have coded that frequency; power and data cross the skin over an RF link between two magnetically aligned coils, so nothing inside the body carries a battery. Retinal implants do the same job for photoreceptors, with an epiretinal array on the inner surface of the retina stimulating ganglion cells, or a subretinal chip stimulating the bipolar cells that photoreceptors normally drive. The channel count is the difference that decides how well each one works. Speech survives being coded into roughly eight effectively independent channels, and 22 contacts is comfortably more than that; form vision needs thousands of pixels, and the Argus II, the only retinal implant ever approved in the US, shipped 60 electrodes.

Cochlear implants work well by any standard: most adults who lost hearing after learning to speak reach 70–90% open-set sentence recognition in quiet, and children implanted before about age two often reach age-appropriate spoken language. Music and speech in a noisy room stay poor, because current spreads through the conductive fluid in the cochlea and neighboring electrodes excite overlapping neuron populations, so the number of genuinely independent channels is closer to eight than to 22. Hardware reliability is high, with cumulative device failure under roughly 0.5% a year, and the recurring cost is the external processor rather than the implant, replaced every five to seven years. Retinal implants have not come close to that bar. The Argus II gave most recipients light localization and large-shape discrimination rather than reading, with best measured grating acuity around 20/1260, and its manufacturer stopped production in 2019 and cut most of its staff in 2020, leaving roughly 350 implanted patients with a device that nobody supports, repairs or updates.

If a postlingually deafened adult scores below roughly 60% on best-aided sentence testing, refer for a cochlear implant workup rather than fitting another hearing aid, since the usual mistake here is a decade of delay rather than a wrong device. If a child is born profoundly deaf, implant early, because the language outcome depends heavily on whether stimulation arrives before about age two. If the patient will need MRI, ask which magnet the implant carries: older designs require the magnet to be surgically removed before a scan, and current self-aligning designs are conditional at 3 T. For blindness from photoreceptor loss, treat every retinal implant as investigational, enroll in the trial rather than buying the product, and ask what happens to the implant if the sponsor stops funding it. Ask that question of any single-product implant company, because explanting a dead device is real surgery and the manufacturer's balance sheet is not something a patient can inspect.

Cochlear arrays of 12–22 contacts inserted 16–31 mm, roughly 8 effectively independent channels · about 736,900 cochlear implants registered worldwide as of December 2019, of which 118,100 US adults and 65,000 US children · postlingual adult sentence recognition typically 70–90% in quiet · cochlear device $25,000–40,000 and a total US episode of $50,000–100,000, at roughly $10,000–25,000 per quality-adjusted life year · Argus II carried 60 electrodes, reached grating acuity near 20/1260, and went into roughly 350 patients · in the PRIMA pivotal trial, 26 of the 32 patients assessed at 12 months improved by at least 0.2 logMAR.

Cochlear's Nucleus, Sonova's Advanced Bionics HiRes Ultra 3D and MED-EL's SYNCHRONY, which between them cover nearly the whole cochlear market; Second Sight's Argus II, approved in the US in 2013 under a humanitarian device exemption and discontinued in 2019; Retina Implant AG's Alpha AMS subretinal chip, whose maker closed in 2019; Science Corporation's PRIMA, a 2 mm subretinal photovoltaic chip powered by near-infrared projected from glasses, whose 12-month pivotal results in the New England Journal of Medicine in 2025 showed 26 of the 32 patients assessed improving by at least 0.2 logMAR.

A cochlear implant is one of the clearest cost-effectiveness cases in medicine. The device costs roughly $25,000–40,000 and the whole US episode $50,000–100,000, and published analyses put it at roughly $10,000–25,000 per quality-adjusted life year in adults, well inside what payers routinely fund for drugs; in children implanted early, the reduction in special-education cost makes several analyses come out cost-saving outright. About 736,900 devices were registered worldwide as of December 2019, of which 118,100 went into US adults and 65,000 into US children, and even so well under 10% of eligible adults in high-income countries receive one. The binding constraint is referral pathways and audiology capacity rather than evidence or price, and that gap is what any new entrant is actually selling into. Three vendors hold nearly all of a market of roughly $2–2.5 billion a year, and a large share of their revenue is the processor upgrade every five to seven years, sold to an installed base that already carries the implanted half. Retinal implants are the inverse case: roughly 350 Argus II devices at about $150,000 each never came near covering development cost, and the humanitarian device exemption that allowed approval also caps the eligible population at 8,000 US patients a year, which is a ceiling on revenue written into the approval. Second Sight's shutdown is the risk worth naming plainly. An implant lasts only as long as someone funds its service, no regulation requires anyone to maintain an orphaned device, and the patient carrying it cannot switch vendors without another operation.

3 minCochlear Implant AnimationFauquier ENT · 500k+ views

4 minWhat a Cochlear Implant Actually Sounds LikeSciShow · 1m+ views

5 minThe Workings of the Retinal ProsthesisSupport Center for Microsystems Education · 10k+ views

What Are Cochlear Implants for Hearing? (NIDCD) · An update on visual prosthesis (International Journal of Retina and Vitreous)

A total hip replacement is four parts: a titanium or cobalt-chrome stem down the femoral canal, a 28–36 mm cobalt-chrome or ceramic head on its taper, a titanium shell pressed into the reamed acetabulum, and a liner inside that shell which is the surface the head actually slides on. A total knee is a cobalt-chrome femoral component, a titanium tibial tray, a polyethylene insert between them, and sometimes a patellar button. Fixation is either bone cement, which is polymethyl methacrylate mixed in the operating room and sets in about ten minutes, or a porous or hydroxyapatite-coated surface that bone grows into over six to twelve weeks. The bearing surface is the part that decides how long the construct lasts, and the material that fixed it was highly crosslinked polyethylene, irradiated to 5–10 Mrad and remelted, which cut wear rates by roughly 80–90% against conventional polyethylene and largely removed particle-driven bone loss as the dominant late failure. Registry data now puts the 10-year revision rate for a primary hip at under 3%, so the leading reasons for revision have shifted to infection, instability and fracture around the implant rather than wear.

Joint replacement is among the most reliably effective operations there is: pain relief is large and durable, and the UK National Joint Registry puts the 10-year revision estimate for a hip implanted in 2014 at 2.7%, comfortably below the 5% at ten years that NICE sets as the bar for a construct to be used at all. The weakness is that the whole class clears through 510(k) against a predicate, so a new bearing couple, taper geometry or coating can reach the market on bench data alone, and problems only appear in a registry three to seven years later. Metal-on-metal bearings are the case study. They were sold on the argument that a metal head on a metal cup would wear less than polyethylene and allow larger heads with better stability, they cleared on predicates, and the resulting cobalt and chromium debris caused soft-tissue destruction and pseudotumors in a fraction of patients; DePuy's ASR was withdrawn in 2010 after roughly 93,000 had been implanted, litigation cost Johnson & Johnson billions, and the registries had flagged the revision signal before regulators moved. FDA published a proposed order in January 2013 and a final order in February 2016 requiring premarket approval for metal-on-metal total hips from May 2016, which ended their sale in the US rather than approving any of them.

For end-stage hip or knee arthritis with failed conservative care, replacement is the standard answer and the interesting decisions are all about which construct. Pick highly crosslinked polyethylene against a ceramic or cobalt-chrome head as the default hip bearing, since it has the registry evidence and there is no wear-driven argument left for anything exotic. Pick cemented fixation in older patients with poor bone quality, where uncemented stems carry a real periprosthetic fracture rate, and uncemented in younger patients with good bone. If a vendor is selling a new bearing, coating or modular junction, ask for registry survivorship rather than wear-simulator data, because that is precisely the gap metal-on-metal went through. If the patient is under 65, healthy and motivated, an ambulatory surgery center is now a reasonable site of care, and it changes the economics more than the surgery.

Femoral heads of 28–36 mm on cobalt-chrome or ceramic · crosslinked polyethylene irradiated to 5–10 Mrad, cutting wear by roughly 80–90% · 10-year hip revision estimate 2.7% against a NICE threshold of 5% · roughly 800,000 knee and 500,000 hip replacements a year in the US, and about 110,000 hips and 119,000 knees a year in the UK registry · implant construct $3,000–6,000, down from $5,000–8,000 in the 2000s · Medicare pays roughly $12,000 for the hospital admission and $10,000–12,000 for the same procedure in an ambulatory surgery center · DePuy ASR withdrawn in 2010 after roughly 93,000 implants.

Stryker Triathlon and Zimmer Biomet Persona knees; DePuy Synthes Corail and Pinnacle hips; Smith+Nephew Oxinium oxidized zirconium bearings; DePuy's ASR metal-on-metal hip, withdrawn in 2010; the UK National Joint Registry and the American Joint Replacement Registry, each holding well over a million procedures and both used as the primary post-market surveillance for the class.

The implant is cheap relative to the episode and expensive relative to the site of care, and that gap is what drives everything in this market. A hip or knee construct sells for roughly $3,000–6,000, down from $5,000–8,000 in the 2000s, and prices have eroded a few percent a year since Medicare started paying hospitals a fixed amount for the whole 90-day episode under bundled payment models. Moving cases to ambulatory surgery centers accelerated that: Medicare took knee replacement off the inpatient-only list in 2018 and hip in 2020, then added both to the ASC list, and at an ASC paid roughly $10,000–12,000 for the case the implant is the single largest cost line rather than a rounding error against a hospital stay. A surgery center that pays $5,000 for a construct has spent close to half its payment before anyone is in the room, so ASC owners negotiate implant prices the way a manufacturer negotiates a bill of materials, and several buy directly or use lower-cost private-label suppliers. Volumes are large and growing with the over-65 population, at roughly 800,000 knees and 500,000 hips a year in the US and about 119,000 knees and 110,000 hips a year in the UK registry, so vendors defend price by tying the implant to surgeon-facing capital: a Stryker Mako robot placed in a hospital pulls Stryker implants for years, which is the same razor-and-blade logic as surgical robotics and the main reason robotic systems are given aggressive placement terms. Litigation is a real line item in this business, not a tail risk, and the metal-on-metal settlements ran into billions of dollars.

2 minTotal Hip ReplacementTriStarHealth · 1m+ views

7 minKnee Replacement Implants - What Do They Look Like And What Are My Options?Talking With Docs · 50k+ views

Metal-on-Metal Hip Implants (FDA) · National Joint Registry 22nd Annual Report 2025 (National Joint Registry)

A continuous glucose monitor is a filament a few millimeters long, pushed into subcutaneous fat by a spring-loaded applicator and wired to a coin-sized transmitter stuck to the skin. The filament carries glucose oxidase immobilized over a working electrode: glucose and oxygen react to make hydrogen peroxide, the peroxide is oxidized at the electrode, and the current that flows is proportional to glucose in the interstitial fluid. Interstitial glucose lags blood glucose by about 5 to 15 minutes, which matters most when it is falling fast, and is why alarm thresholds are set with margin. Current sensors report every one to five minutes, are factory-calibrated so the user never pricks a finger to calibrate them, and hold a mean absolute relative difference against a lab reference of about 8%. Wear time is 10 days for Dexcom's G7 and 14 for Abbott's Libre 3; the exception is Senseonics' Eversense, a fluorescence-based sensor placed under the skin of the upper arm in a short office procedure and read through the skin by an external transmitter, which runs 180 days or a full year. Automated insulin delivery closes the loop by feeding the sensor stream to an algorithm on a pump or phone, which adjusts basal insulin every five minutes and gives correction boluses, with the user still announcing meals.

The clinical case is settled. Automated insulin delivery raises the share of the day spent between 70 and 180 mg/dL from roughly 50–60% on injections to 70–75%, and lowers HbA1c by about 0.3–0.6 points, without the extra hypoglycemia that used to be the price of tighter control. The weaknesses are mechanical and behavioral. Sensors fail early or read badly often enough that vendors ship replacements as a routine cost of business, adhesive lets go in heat and water, lying on a sensor produces false lows that wake people up at night, and some drugs interfere with the electrochemistry directly, which is why every sensor carries its own interference list. Insulin delivery is the harder half, because an infusion set can kink or occlude and a stopped basal produces ketoacidosis within hours, whereas a bad reading only produces a bad decision. The algorithm is also bounded by the drug: a rapid-acting analog peaks 60 to 90 minutes after dosing, so no controller can respond to a meal faster than the insulin it is dosing.

If a patient has type 1 diabetes, they should be on a sensor, and the only real questions are which one and whether to add a pump. If they are on multiple daily injections and still struggling, moving to automated insulin delivery usually helps more than any further adjustment to the injection regimen. If they have type 2 diabetes and take insulin, Medicare and most commercial plans now cover a sensor, so cost is rarely the obstacle; if they have type 2 on oral agents alone, an over-the-counter sensor is a behavior-change tool rather than a treatment and should be judged against that expectation. Choose the implanted 180-day or one-year sensor when adhesive failure or repeated insertions are what keeps breaking, and accept two office procedures a year in exchange. If the patient will not carry a phone or will not act on alarms, the sensor mostly produces data nobody uses, so fix that before buying hardware.

Sensor wear 10 days for Dexcom G7 and 14 for Abbott Libre 3, against 180 days or a year for the implanted Eversense · mean absolute relative difference around 8% · interstitial lag of 5–15 minutes · automated delivery lifts time in range from roughly 50–60% to 70–75% and cuts HbA1c 0.3–0.6 points · sensors run roughly $35–75 each at cash retail, or $1,300–3,000 a year · gross margins around 60–65% · the first over-the-counter CGM cleared in March 2024 at about $99 a month.

Dexcom G7 and Stelo, the first over-the-counter CGM, cleared in March 2024; Abbott FreeStyle Libre 3 and its consumer versions Lingo and Libre Rio; Senseonics Eversense 365, an implanted sensor rated for a year; Tandem t:slim X2 with Control-IQ, Insulet Omnipod 5 and Medtronic MiniMed 780G automated delivery systems; FDA's 2018 De Novo for integrated CGM, which moved the category out of premarket approval into Class II with special controls.

This started as a device sold to endocrinologists and became a consumer product, and that shows up in every line of the business. A sensor sells for roughly $35–75 at cash retail against gross margins of about 60–65%, so most of the price is not manufacturing, and the model is a subscription: the reader or transmitter is nearly free, the recurring sensor is where the revenue sits, and continuous wear runs $1,300–3,000 a year at list. Most of the recent volume came from coverage rather than from technology, since Medicare widened CGM coverage in 2023 to essentially every insulin user and to patients with a history of problematic hypoglycemia, which moved a large population from cash-pay to covered overnight. The regulatory structure mattered as much as the coverage: FDA's 2018 De Novo created an integrated CGM class with special controls, which took sensors out of premarket approval and let one company's sensor legally drive another company's pump, and Dexcom, Tandem and Insulet all built businesses on that interoperability. Over-the-counter clearance in 2024 opened the far larger population of type 2 patients not on insulin, where nothing is reimbursed and the product competes with fitness wearables at around $99 a month, which is a different buyer and a much lower price per day. Abbott and Dexcom together sell more than $10 billion of sensors a year and both keep cutting price per day of wear, which is the normal shape of a consumer hardware market and unusual for anything else on this sheet. The strategic question for a new entrant is not accuracy, which is now commoditized around 8%, but distribution, since two companies already sit in the pharmacy channel with a decade of payer contracts behind them.

5 minGlucose Sensors ExplainedDhvani Patel · 10k+ views

3 minClosed-loop insulin delivery in adults with type 1 diabetesThe Lancet · 10k+ views

Continuous Glucose Monitoring (NIDDK) · Recent advances in closed-loop insulin delivery (Metabolism: Clinical and Experimental)

A brain-computer interface records electrical activity from motor or speech cortex and decodes the intended movement or word from it, so a paralyzed person can drive a cursor, a robot arm or a speech synthesizer. The workhorse recording device is the Utah array, a 4 by 4 mm silicon block with 96 electrodes on shanks about 1 to 1.5 mm long that penetrate the cortical surface, wired out to a connector bolted to the skull. Decoding is a supervised learning problem: the participant attempts a movement, the system fits a model from firing rates to intended velocity or phoneme, and the model is retrained regularly as the recorded population drifts. Three device architectures are being tried at once. Penetrating arrays give the highest information rate and the most surgical risk, Neuralink's N1 puts 1,024 electrodes on 64 flexible polymer threads inserted by a robot and transmits wirelessly with no percutaneous connector, and two lower-risk approaches avoid penetrating cortex at all: Synchron threads a 16-electrode stent up the jugular into the vein running over motor cortex, and Precision Neuroscience lays a 1,024-contact thin film on the cortical surface. The recent results are real: speech has been decoded at 60 to 80 words per minute from participants who cannot speak. They also come from a very small number of people.

For someone with locked-in syndrome or high tetraplegia there is no competing technology, so the comparison is against no communication at all, which is why participants and reviewers both tolerate a high risk profile. The unsolved problem is that the electrodes do not last. Penetrating shanks provoke a foreign-body response that encapsulates them in glial scar, and in the largest pooled study of Utah arrays, covering 55 arrays and nearly nine years of recordings, the average usable recording lifespan was 622 days, with some arrays running past 1,000 days and one to 9 years. A device that stops working after two years and needs a craniotomy to replace is a research tool, not a treatment. The other honest limits are scale and setting: total human experience across every group is on the order of a hundred people, most sessions still involve engineers in the room, and decoders need periodic recalibration that a patient at home cannot perform.

Treat this as a research program, not a purchase. If you are a clinician, the only route today is enrolling a patient in a trial under an investigational device exemption, and the right candidates are people with severe paralysis and intact cognition who can commit to years of sessions. If you are an investor, discount channel count and decoding accuracy, because both improve reliably and neither is the constraint; ask instead how long the arrays hold signal in the specific animals or people implanted, and how many were revised. If a company's plan depends on Medicare, ask which benefit category it expects to fall into, because there isn't one yet and the answer determines whether payment takes three years or ten. If the application is restoring communication rather than dexterous control, weigh an endovascular or surface device with fewer channels against a penetrating array, since a device that yields fewer bits but avoids opening the dura may be the better trade for that use.

Utah array of 96 electrodes on a 4 by 4 mm base, shanks about 1–1.5 mm · Neuralink N1 with 1,024 electrodes on 64 flexible threads, Synchron's endovascular device with 16 · speech decoded at 60–80 words per minute · average usable recording lifespan of 622 days across 55 pooled Utah arrays, with some past 1,000 days and one to 9 years · total human experience on the order of a hundred people worldwide · no CPT code, no Medicare benefit category, and no list price, because nothing is for sale.

The BrainGate consortium, which has run intracortical trials since 2004 and produced most of the published human results; Neuralink's N1, first implanted in a person in January 2024, whose initial participant lost usable channels when threads retracted; Synchron's Stentrode and its COMMAND trial; Precision Neuroscience's Layer 7 surface array, cleared by FDA in 2025 for use up to 30 days; Blackrock Neurotech, which supplies the Utah arrays nearly every academic group uses.

There is no revenue in this field and no price to quote, so the economics are entirely about who funds the gap between a trial and a product. The field has raised well over a billion dollars in private capital, mostly against a story about restoring function to a few hundred thousand people with severe paralysis in the US, which is a real but small population and not one that supports a consumer price. The harder problem is that no payment pathway exists. A brain-computer interface does not fit cleanly into any Medicare benefit category — it is not durable medical equipment in the usual sense, not a prosthetic limb, and not an inpatient procedure with an existing code — and until it fits one, an approved device would have no rate attached. CMS's Transitional Coverage for Emerging Technologies pathway gave Breakthrough-designated devices a faster route to a national coverage determination, and was paused for new candidates in August 2026 in favor of the RAPID coverage pathway. Either one takes only a handful of devices a year and still ends in a coverage decision. A credible milestone would not be a higher channel count or a faster typing rate, since both improve steadily. It would be arrays holding usable signal for five years without surgical revision in a double-digit number of participants, systems that work unattended at home for weeks, and a named Medicare benefit category with a payment rate. Until those three exist, the sensible assumption is that this is a decade from a product, and that the electrode-materials work matters more to that timeline than the decoding work does.

5 minHow a Brain Implant and AI Gave a Woman with Paralysis Her Voice BackUC San Francisco (UCSF) · 5m+ views

13 minNeuralink brain chip's first human patient. How does it work? | About ThatCBC News · 100k+ views

Implanted Brain-Computer Interface (BCI) Devices for Patients with Paralysis or Amputation - Non-clinical Testing and Clinical Consideration (FDA) · Longevity and reliability of chronic unit recordings using the Utah, intracortical multi-electrode arrays (Journal of Neural Engineering)

The FDA sorts devices into three risk classes and gives each a route to market. Most devices go through a 510(k) premarket notification, which asks the sponsor to show the device is substantially equivalent to a legally marketed predicate rather than to show it works: same intended use, and either the same technological characteristics or different ones that raise no new questions of safety and effectiveness. A De Novo request covers a low-to-moderate-risk device with no predicate, and the FDA grants it along with special controls, creating a new classification that later devices can then use as a predicate. A premarket approval application (PMA) is the Class III route, and it asks for independent evidence of a reasonable assurance of safety and effectiveness, which in practice means a clinical trial. The volumes are lopsided. In calendar 2025 the FDA cleared 3,186 devices by 510(k), granted 27 De Novo requests, and approved 38 original PMAs, so roughly 98% of premarket authorizations that year were 510(k)s (the three counts come from FDA's own databases; the percentage is arithmetic on them). FY2026 user fees are $26,067 for a 510(k), $173,782 for a De Novo and $579,272 for a PMA, with qualified small businesses paying a quarter of each.

The 510(k) is fast and cheap relative to what it authorizes. Across the 3,183 clearances issued in 2025 with usable dates, the median time from submission to decision was 126 days and the mean 146, both computed here from FDA's clearance records, so a company with a clean predicate can plan on four to six months. The cost of that speed is that the claim is bounded by the predicate: if the device does something the predicate did not, the FDA can find it not substantially equivalent, and the sponsor lands in De Novo or PMA anyway. The chain of comparisons is the structural weakness, because each clearance is a small step from the last and a device several predicates down the line can differ substantially from whatever the chain started with. That is predicate creep, and it is why the 2011 National Academies review of the program and a run of high-profile recalls left the FDA under pressure to tighten predicate selection. PMA fixes the evidence problem and creates a schedule problem: original PMAs approved between 2023 and 2025 took a median of 307 days at the agency, and that clock only starts after a pivotal trial that a 2015 analysis of priority-review devices found ran a median of three years and enrolled a median of 297 patients.

Start in FDA's classification database and find the product code that describes what your device does. If a product code exists and a marketed predicate shares your intended use, write the 510(k) and budget four to six months of review, because that is the cheapest marketing authorization available and it is worth contorting an indication statement to stay inside it. If the technology is genuinely new but the risk is moderate and nothing fits as a predicate, file a De Novo: expect to wait roughly twice as long, and note that the special controls written into the new classification become the bar every follower has to clear. If the device is implanted, life-sustaining, or otherwise Class III, plan the PMA from the beginning, since bolting a pivotal trial onto a program designed around a 510(k) costs more than designing for the trial. Request Breakthrough designation if the device treats or diagnoses a life-threatening or irreversibly debilitating condition, but treat it as schedule help rather than a lower bar: it buys interactive review and queue priority, and it does not change the evidence standard or get you paid.

3,186 510(k) clearances, 27 De Novo grants and 38 original PMAs in calendar 2025, so 510(k) is about 98% of authorizations · FY2026 user fees $26,067 / $173,782 / $579,272, a quarter of that for small businesses · 510(k) median 126 days to decision, original PMA median 307 · PMA pivotal trials median 297 patients over 3 years · 1,284 Breakthrough designations against 198 marketing authorizations · PMA supplements $86,891 for a 180-day, $463,418 for a panel-track.

Intuitive's da Vinci systems, whose predicate chain a 2023 PLOS ONE case study used to demonstrate a method for detecting predicate creep; FDA's Safety and Performance Based Pathway, which lets some well-understood device types clear against published performance criteria instead of a named predicate; the September 2023 guidance on best practices for selecting a predicate device; the Breakthrough Devices Program, which had granted 1,284 designations and seen 198 of those devices reach marketing authorization as of March 31, 2026.

The user fee is the smallest line in either budget. A traditional 510(k) program is mostly bench testing, biocompatibility, sterilization validation and the people who write the submission, and a straightforward one runs a few hundred thousand dollars all in. A PMA program is the pivotal trial, and a median 297-patient trial over three years is a tens-of-millions-of-dollars line before anyone has manufactured at scale, which is a planning estimate rather than a published figure. That gap sets how the company gets financed: a 510(k) product can reach revenue on a seed and a Series A, while a PMA product needs the trial funded as its own tranche, which is why Class III startups usually sell to a strategic acquirer at approval rather than before it. The Class III cost is also permanent, since every meaningful change needs a supplement at $86,891 for a 180-day review or $463,418 for a panel-track one, and the FDA approved roughly 800 PMA supplements a year against 33 to 38 originals over 2023 to 2025. De Novo buys a head start rather than exclusivity, because the classification it creates is public and the next company files a 510(k) against it, so the protection is the two or three years of lead plus whatever clinical evidence and billing codes you accumulate meanwhile. Breakthrough designation moves the review queue and not the payer, so it belongs in the schedule model and not the revenue model.

2 minBasics of 510(k) Clearance ProcessStanford Biodesign · 10k+ views

3 minKey differences between 510(k) and PMA pathways at the FDAStanford Biodesign · 10k+ views

Premarket Notification 510(k) (FDA) · Identification of predicate creep under the 510(k) process: A case study of a robotic surgical device (PLOS ONE)

Software as a medical device is software intended for a medical purpose that achieves that purpose on its own, rather than by driving a piece of hardware, and the FDA regulates it through the same three pathways as any other device. Almost all of it is Class II and clears by 510(k). The FDA publishes an AI-Enabled Medical Device List, which held 1,524 authorizations at its March 2026 update: 1,466 cleared through 510(k), 39 granted through De Novo and 19 approved as PMAs. Radiology accounts for 76% of the list, cardiovascular 10% and neurology 5%, and the rate is climbing fast, with 333 authorizations in calendar 2025 against 18 in 2016. What these products actually do is narrower than the label suggests: they segment an organ, measure an ejection fraction, or flag a suspected large-vessel occlusion on a CT angiogram and page the stroke team. The FDA's De Novo order for the first stroke-triage product said in as many words that identifying the suspected finding is not for diagnostic use beyond the notification, which is a fair description of most of the category.

The cost structure is the attraction. There is no bill of materials, no sterilization, no service contract, and a cleared model ships as a container that talks to the hospital's existing PACS. Regulatory load is lighter than hardware too: across the machine-learning devices authorized in 2024, median review time was 162 days, 151 days for the 510(k)s and 372 for the De Novos. The weakness is that the evidence behind a clearance is often thin and hard to check. Of those 2024 authorizations, only 29.2% reported both sensitivity and specificity in the public summary and 15.5% reported any demographic breakdown of the test set, while the median predicate was 2.2 years old and 64.5% of predicates were themselves machine-learning devices, so the substantial-equivalence chain increasingly compares one algorithm with another. Postmarket performance shows the consequence: a 2026 JAMA Network Open cohort of 903 AI-enabled devices authorized through August 2024 found 43 of them recalled, a median of 458 days after authorization, with recall rates highest among devices that had no clinical performance study behind them.

If the product is a fixed model doing one measurable task on data a hospital already collects, treat it as a Class II 510(k) and spend your effort finding a predicate with the indication you want. If you intend to retrain the model after launch, write a predetermined change control plan into the original submission: section 515C of the FD&C Act and FDA's December 2024 guidance let you specify the modifications, the data and methods used to make them, and the impact assessment up front, and then make those changes without a new marketing submission. Uptake is still low, at 16.7% of 2024 machine-learning device summaries, so this is currently a differentiator rather than table stakes. If the software makes an autonomous decision with no clinician in the loop, expect De Novo and a prospective trial in the intended-use population, not a 510(k). And settle how the product gets paid before you build it, because a cleared algorithm with no billing code and no labor it visibly replaces is a very hard sale.

1,524 AI-enabled devices on FDA's list at the March 2026 update, 96% of them 510(k) clearances · 76% radiology, 10% cardiovascular · 333 authorizations in 2025 against 18 in 2016 · median review 162 days, 151 for 510(k) and 372 for De Novo · 29.2% of 2024 authorizations reported both sensitivity and specificity, 15.5% reported demographics · PCCPs in 16.7% of 2024 summaries · 43 of 903 AI devices recalled, median 458 days after authorization · ContaCT's add-on payment capped at $1,040 per case.

Viz.ai's ContaCT, authorized by De Novo in February 2018 and the first AI product to get a Medicare new technology add-on payment, capped at $1,040 per case for FY2021 and requiring its own ICD-10-PCS code (4A03X5D) before the payment could be billed; IDx-DR, authorized by De Novo in April 2018 for autonomous diabetic retinopathy screening and now billed under CPT 92229; Paige Prostate, the first De Novo for AI in pathology, in September 2021; FDA's December 2024 guidance on predetermined change control plans for AI-enabled device software functions.

Development cost is mostly people and clinical data, and the marginal cost of the thousandth deployment is close to zero, so the whole question is revenue. There are only three ways an algorithm gets paid in the US, and each has a defect. It can be folded into an existing procedure code, in which case there is no incremental revenue and the buyer pays out of an operating budget. It can get its own CPT code, which is usually a Category III tracking code with no national value, so payment is at each contractor's discretion. Or it can get a temporary add-on payment, which is what Viz.ai obtained and which expires after two or three years. CPT 92229, one of the few Category I codes written for an autonomous algorithm, has no professional component at all, because no physician interprets the image, and CMS has said publicly that it is still evaluating how services involving assistive technology should be valued. That is the unresolved part: the fee schedule pays for physician work, and an algorithm that removes physician work removes the basis of the payment. In the meantime most vendors sell a per-study subscription to a radiology group or a hospital, where the return to the buyer is throughput and turnaround time rather than new revenue, which caps the price at roughly what the labor saved is worth.

2 minUsing artificial intelligence in radiology clinical practiceMayo Clinic · 10k+ views

15 minArtificial Intelligence in medical imaging: From research to clinical practice – Koen Van LeemputAalto University · 5k+ views

Artificial Intelligence-Enabled Medical Devices (FDA) · Machine Learning-Enabled Medical Devices Authorized by the US Food and Drug Administration in 2024: Regulatory Characteristics, Predicate Lineage, and Transparency Reporting (Biomedicines)

Getting paid in the US is three separate problems, and none of them is solved by an FDA clearance. Coding asks whether a code exists that describes what was done, coverage asks whether the payer will pay for it in this patient, and payment asks how much. Procedure codes come from the American Medical Association rather than the government: a Category I CPT code requires that every device and drug needed has FDA clearance or approval, that many physicians across the country perform the service, that the volume matches the intended clinical use, and that clinical efficacy is documented in the literature. A Category III code is a temporary tracking code for emerging technology and requires only that the service is performed in humans plus one of an advisor's sponsorship, peer-reviewed literature, or an IRB-approved protocol. The difference in the money is total: a Category III code carries no relative value units, so payment is at each Medicare contractor's discretion or is simply denied, and the code is archived five years after publication unless someone requests conversion or extension. The timelines are slow and fixed. Category III codes accepted at the AMA panel's September 2025 meeting became effective on July 1, 2026 after a six-month implementation period, and a Category I code has to clear the panel, then be valued by the RUC, then survive CMS rulemaking before it is worth anything.

Once a technology has a valued Category I code and a settled coverage position, the revenue is predictable and the sales conversation turns into a purchasing decision. Everything before that point is the weakness. MR-guided focused ultrasound of the brain is still billed under 0398T, a Category III code, nearly a decade after the FDA approved the device in July 2016: Medicare's 2024 Physician and Other Practitioners data shows roughly 1,600 services nationally, and because the code is carrier-priced rather than nationally valued, the allowed amount was about $880 in California against about $1,670 in Florida for the same procedure. Inpatient devices face a different failure. Medicare pays the hospital a fixed amount per admission set by the MS-DRG, so a better device that costs $10,000 more comes straight out of hospital margin unless the DRG weight moves, which takes years of claims data. A new technology add-on payment softens that and does not fix it: the payment is the lesser of 65% of the cost of the technology or 65% of the amount by which the case exceeds the DRG payment, so the hospital absorbs at least 35% of the difference, and the technology stops being "new" after two to three years.

Decide the coding strategy before the pivotal trial, because the endpoints that satisfy the FDA are usually not the endpoints a payer will pay against, and adding a health-economic arm afterwards means running a second study. If an existing Category I code plausibly describes your procedure, use it and spend the effort on coverage instead. If nothing fits, apply for a Category III code as early as the criteria allow, and treat the utilization it generates as the evidence you will need for conversion. If the device is used on inpatients and costs more than the DRG covers, get the ICD-10-PCS procedure code first and then apply for NTAP, remembering that an application due in October pays out starting the following October and that FDA authorization must be in hand by May 1. In the hospital outpatient setting, apply for transitional pass-through within three years of FDA authorization, which pays the device cost on top of the procedure rate for two to three years and then folds into it. If you hold a Breakthrough designation on a Class III device heading for PMA, look at CMS's RAPID coverage pathway, announced in April 2026 and out for comment since August 11, 2026, which proposes to post a proposed national coverage determination on the day of FDA authorization and finalize it as soon as 60 days later; the older TCET pathway is paused for new candidates.

Category III codes carry no RVUs and are archived 5 years after publication · 0398T still Category III nearly a decade after the July 2016 PMA approval, roughly 1,600 Medicare services in 2024, allowed amounts about $880 in California against about $1,670 in Florida · NTAP pays the lesser of 65% of technology cost or 65% of the excess over the DRG, so the hospital absorbs at least 35% · NTAP newness window 2 to 3 years, applications due in October for payment the following October, FDA authorization required by May 1 · outpatient pass-through runs 2 to 3 years and must be applied for within 3 years of FDA authorization · RAPID proposes a proposed NCD on the day of authorization with a final NCD as early as 60 days later.

CPT 0398T for MR-guided focused ultrasound, a Category III code with no national value nearly a decade after approval; Viz.ai's ContaCT, the first AI product to receive an NTAP, capped at $1,040 per case for FY2021 and dependent on a new ICD-10-PCS code (4A03X5D) issued first; CMS's Transitional Coverage for Emerging Technologies pathway, finalized August 2024 and paused for new candidates in August 2026 in favor of the RAPID coverage pathway.

A reimbursement program is a real line item and it is usually underfunded: a specialty society sponsor, a health economics dossier, claims analyses, and a reimbursement lead working the AMA and CMS cycles for two to four years. A useful rule of thumb is to budget the same order of magnitude for coding and coverage work as for the clinical evidence itself, and to assume no meaningful US revenue until a valued code exists. Who pays decides who buys. Under a DRG the hospital captures the payment and absorbs the device cost, so you sell on shortening length of stay, avoiding a readmission, or displacing a more expensive disposable, and a purely clinical improvement with no cost offset gets stalled at the value analysis committee. In the physician office the practice buys the device out of its own margin against a professional fee, so a code without RVUs means the practice funds the equipment and gets paid nothing extra for using it. The consequence is blunt: hospitals buy the mediocre device that has an established Category I code and a favorable predicate ahead of the better device that has neither, and the delay can run five to ten years, which is longer than most startups can fund.

6 minPayment Methodologies for Hospital Inpatient ServicesGreg Palmer · 10k+ views

19 min4.3 Reimbursement Basics for Medical DevicesLuis R Soenksen · 1k+ views

New Medical Services and New Technologies (CMS) · Criteria for CPT® codes (American Medical Association)

Anything labeled sterile has to reach a sterility assurance level of 10^-6, meaning no more than one chance in a million that a given unit carries a viable organism, and industry gets there by four main routes. Ethylene oxide gas alkylates microbial DNA, runs at 37 to 63 °C with roughly 1 to 6 hours of exposure at 450 to 1200 mg/L and 40 to 80% humidity, penetrates sealed packaging and long lumens, and then needs days of aeration to drive residuals down to the limits in ISO 10993-7. Gamma irradiation from cobalt-60 and electron or X-ray beams work by free-radical damage, need no residual removal, and typically deliver 15 to 45 kGy with 25 kGy as the industry default. Steam at 121 to 134 °C is the cheapest of all and destroys anything thermoplastic, electronic, or protein-based. The FDA treats moist heat, dry heat, EtO and radiation as established methods that can be validated against recognized consensus standards, and puts a heavier evidence burden on anything novel. EtO does about half of all sterile devices sold in the US, which EPA puts at 20 billion units a year across 88 commercial sterilizers.

EtO is often the only method that sterilizes a finished device without ruining it, which is why catheters, multi-layer packaging, and heat-sensitive polymer assemblies still go through it. It is also a carcinogen, and the plants are where people get exposed. EPA's April 2024 air toxics rule would have cut emissions from commercial sterilizers by over 90%, facilities said they could not meet the deadlines, a presidential exemption process was created in January 2025, and in March 2026 EPA proposed repealing the 2024 rule with comments closing that May. For a manufacturer, the concentration matters more than the rule: half the country's sterile output moves through 88 sites, and losing one is a supply interruption rather than an inconvenience. Radiation has no residuals and no plant siting fight, and it damages materials instead: irradiation of ultra-high-molecular-weight polyethylene generates free radicals that oxidize the chains, which caused increased wear and premature failure in joint implants until manufacturers moved to irradiating under vacuum or inert gas with radical scavengers. Cobalt-60 also decays with a 5.27-year half-life, so the operator of a gamma facility replaces source every year whether the plant runs or not.

Choose the sterilization method while you are still choosing materials, not after the design freeze, because the method constrains the polymer and the packaging more than the other way round. If the device is a heat-sensitive assembly with lumens or dead spaces, plan on EtO and design the packaging for gas exchange and aeration. If it is a simple, radiation-stable polymer or metal in a sealed pouch, use gamma, e-beam or X-ray and skip the residuals work entirely, but qualify the material at the maximum dose rather than the nominal 25 kGy. If it is reusable metal instrumentation, steam is cheaper than anything else and you should not be arguing about it. Second-source the sterilization site from the start if you can, because a single-site dependency on a Class III device is the most expensive kind: FDA's master file pilot, opened in November 2019, exists specifically so that PMA holders can change EtO site through a post-approval report instead of a PMA supplement.

Sterility assurance level 10^-6 · EtO at 37–63 °C, 1–6 hours exposure, 450–1200 mg/L, 40–80% humidity, plus days of aeration · radiation 15–45 kGy with 25 kGy the default · steam at 121–134 °C · EtO handles about half of US sterile devices, roughly 20 billion units a year, across 88 commercial sterilizers · EPA's 2024 rule targeted a 90%-plus emissions cut and was proposed for repeal in March 2026 · cobalt-60 half-life 5.27 years · ISO 10993-1 sorts devices by surface, external-communicating or implant contact and by limited, prolonged or long-term duration.

ISO 11135 for EtO process validation, ISO 11137 for radiation and ISO 10993-7 for EtO residuals, all recognized by the FDA; the July 2019 FDA innovation challenges for new sterilization methods and lower EtO emissions; the November 2019 EtO Sterilization Master File pilot, whose participants include Boston Scientific, Becton Dickinson, Steris, Oscor and Medtronic; the FDA's September 2023 guidance recognizing ISO 10993-1 for biological evaluation.

Contract sterilization itself is cheap, usually cents to a few dollars a unit, and the cost that matters is qualification and dependency. Biocompatibility is the schedule item. ISO 10993-1, which the FDA recognized in its September 2023 guidance, is a risk-based framework rather than a fixed list: you classify the device by contact type (surface, external-communicating or implant) and by contact duration (limited, prolonged or long-term), then justify which endpoints need data, starting with chemical characterization and a toxicological risk assessment and adding cytotoxicity, sensitization, irritation, systemic toxicity, hemocompatibility, genotoxicity and implantation as the contact gets longer. The trap is that the testing has to run on the final finished device, in its final material, made by the final process, and sterilized by the final method, so nothing can start until the design is frozen, and for a long-term implant the panel typically runs six to eighteen months and costs into the hundreds of thousands of dollars (a planning rule of thumb, not a published rate). That puts a strong economic argument behind reusing materials with an existing history of use, because choosing a novel polymer means paying for the full battery plus the risk of a finding that sends the design back. Changing sterilization method or site on a marketed device costs the same way: revalidation, possibly new biocompatibility data if residuals or material properties shift, and on a PMA device either a supplement at $86,891 or the master file route. Plan a year for that change and treat single-source sterilization as a real supply risk rather than a procurement detail.

3 minMedical Device Sterilization: ETO, Radiation, and SAL 6 ExplainedStarFish Medical · 5k+ views

2 minEO Sterilization ProcessSOLSTEO · 10k+ views

Sterilization for Medical Devices (FDA) · Sterilization of Polymeric Implants: Challenges and Opportunities (ACS Applied Bio Materials)

Section 524B of the Federal Food, Drug, and Cosmetic Act, added by the Consolidated Appropriations Act of 2023 and effective March 29 that year, made cybersecurity a statutory content requirement for premarket submissions. It applies to a cyber device, defined as one that includes software the sponsor validated, installed or authorized, can connect to the internet, and has characteristics that could be vulnerable to cybersecurity threats. Three things have to be in the submission: a plan to monitor, identify and address postmarket vulnerabilities and exploits in a reasonable time, including coordinated vulnerability disclosure; processes that give reasonable assurance the device and related systems are cybersecure, along with a way to make postmarket updates and patches available; and a software bill of materials covering commercial, open-source and off-the-shelf components. The requirement covers 510(k), De Novo, PMA, PDP and HDE submissions and their supplements, and since October 1, 2023 an eSTAR submission is put on technical screening hold if the cybersecurity section is incomplete, which makes this an admissibility gate rather than a review comment. The FDA's current final guidance on the subject is dated June 27, 2025. Interoperability is the other half of the same problem: a connected device has to speak DICOM for images, HL7 v2 or FHIR for clinical data, and the relevant IHE profiles for the workflow, or every hospital becomes its own integration project.

The strength of 524B is that it forces an inventory. A software bill of materials lets a hospital answer whether it owns anything containing a newly disclosed library without calling twenty vendors, and a published disclosure policy gives researchers somewhere to send a finding other than a conference talk. The weakness is that the law is prospective only, since it does not apply to submissions filed before March 29, 2023, so it does nothing for the installed base. That base is the actual problem. An imaging system or an infusion pump has a 10 to 20 year service life against operating system support windows of roughly half that, and the software image is validated as part of the device, so a hospital cannot patch it the way it patches a laptop; the change has to be validated by the manufacturer against the device's intended use. For devices that are past support or whose vendor is gone, there is no patch at any price, and the only mitigations left are network segmentation, monitoring, and taking the device off the network. On the interoperability side, a proprietary interface is a hidden cost that shows up in the sales cycle rather than the bill of materials, because each site needs its own integration work.

Assume 524B applies if the device connects to anything at all, and generate the software bill of materials out of the build system rather than assembling one by hand at submission time, because it has to stay current for the life of the product. Design for patchability from the start: signed updates, a rollback path, and a written revalidation scope that says which classes of change can ship in days rather than months. Publish a coordinated vulnerability disclosure policy and a security contact before you need one, since the first serious report is a bad time to invent the process. Choose DICOM, HL7 FHIR and the applicable IHE profiles over a proprietary API even when the proprietary one is easier, because the cost of a custom integration per hospital usually exceeds the engineering cost of conformance within a handful of sites. If you sell into an existing fleet, expect a security questionnaire and a segmentation requirement before purchasing, and treat the answers as a sales asset rather than a compliance chore.

Section 524B effective March 29, 2023, with three required elements: a postmarket vulnerability plan, patch and update processes, and a software bill of materials · eSTAR submissions on technical screening hold for incomplete cybersecurity sections since October 1, 2023 · current FDA final guidance dated June 27, 2025 · requirement applies to 510(k), De Novo, PMA, PDP and HDE plus supplements, and not to anything submitted before March 29, 2023 · device service life 10 to 20 years against operating system support windows of roughly half that · interoperability standards in practice are DICOM, HL7 v2 and FHIR, plus IHE profiles.

FDA's June 2025 final guidance on cybersecurity quality system considerations and premarket content; the MITRE and FDA legacy medical device report and the Medical Device Cybersecurity Regional Incident Preparedness and Response Playbook; MITRE's 2024 and 2026 white papers on normalizing software bill of materials data, which exist because SBOMs from different vendors are not comparable as shipped; DICOM and HL7 FHIR as the two interfaces almost every clinical buyer will ask about.

Security is a permanent engineering line rather than a submission cost. Threat modeling, penetration testing, SBOM tooling, vulnerability monitoring and a release pipeline all have to stay staffed for the device's whole service life, which for hardware means 10 to 20 years of patch obligations against a single sale. That mismatch is one of the reasons vendors keep moving from one-time capital sales toward service contracts and subscriptions, since a recurring payment is the only structure that funds a recurring obligation. Nobody pays a premium for a secure device, so the return shows up as deals that do not stall: a hospital's security review is a gate in front of the purchase order, and a vendor who can hand over an SBOM, a disclosure policy and a patch history clears it in weeks instead of quarters. On the hospital side the cost of a legacy fleet is real and mostly invisible, showing up as segmentation projects, monitoring, cyber insurance premiums, and eventually capital replacement of equipment that still works clinically but can no longer be defended. Interoperability pays back the same way: a device that ships standards conformance drops into an existing workflow, while a proprietary interface adds an integration project per site and usually turns a three-month sales cycle into a year.

10 minProtecting Medical Devices from Cyberharm | Stephanie Domas | TEDxColumbusTEDx Talks · 10k+ views

6 minHow medical devices like pacemakers, insulin pumps can be hackedCBS Mornings · 5k+ views

Cybersecurity (FDA) · Cybersecurity requirements for medical devices in the EU and US - A comparison and gap analysis of the MDCG 2019-16 and FDA premarket cybersecurity guidance (Computational and Structural Biotechnology Journal)