The distribution of FDA-cleared AI devices is not uniform across medicine. As of Q2 2026, radiology holds the largest share of authorized AI/ML-enabled devices — by a wide margin — followed by cardiology and pathology. Gastroenterology and primary care have smaller but growing footprints, with a handful of cleared tools and a more active clinical trial pipeline than the device count alone suggests.
This page maps that distribution. For each specialty, it identifies the categories of cleared tools, the evidence quality supporting them, active registered trials, the companies with the most significant presence, and the equity concerns that are specific to each specialty's patient population and data characteristics. It does not duplicate individual FDA device records or evidence appraisals — it links to them and provides synthesis context.
Radiology
Radiology is where the majority of healthcare AI regulatory activity has concentrated. The FDA's AI/ML-enabled device list is dominated by radiology applications — chest X-ray triage, CT pulmonary embolism detection, mammography CAD, brain MRI segmentation, and bone age estimation among the most common cleared categories. Most of these cleared devices follow the 510(k) pathway, with a smaller subset authorized via De Novo for genuinely novel device types.
Cleared Tool Categories
- Chest X-ray prioritization and triage (pneumothorax, consolidation, nodule flagging)
- CT pulmonary embolism detection — workflow triage, not standalone diagnosis
- Mammography computer-aided detection (CAD) and density assessment
- Brain MRI lesion segmentation and volumetric measurement
- Bone age estimation from hand radiographs
- Intracranial hemorrhage detection on non-contrast CT
- Diabetic retinopathy screening from fundus photographs (cleared under ophthalmology but often deployed in radiology workflows)
Evidence Quality
The evidence base for radiology AI is the most developed of any specialty, but that does not mean it is uniformly strong. The majority of published studies are retrospective, using curated single-institution datasets. External validation — testing on data from a different institution or population than the training set — is present in some but not all cleared devices. Prospective randomized trials remain rare; the few that exist tend to focus on workflow efficiency (time-to-read, radiologist miss rate) rather than patient outcome endpoints.
Equity Concerns
Training dataset composition is the central equity problem in radiology AI. Most foundational training sets — including large public datasets like CheXpert and MIMIC-CXR — skew toward academic medical centers in the United States, with documented underrepresentation of certain racial and ethnic groups, older patients, and patients with comorbidities that alter image appearance. Chest X-ray models trained predominantly on one demographic have shown measurable performance gaps when evaluated on others. Bone age models trained on historical datasets derived from specific populations have known limitations when applied broadly.
Cardiology
Cardiology AI spans two distinct application areas with very different evidence maturity: ECG interpretation and echocardiography analysis. ECG-based AI has the longer track record, with cleared devices for atrial fibrillation detection, QT prolongation flagging, and STEMI identification. Echocardiography AI — automated measurement, ejection fraction estimation, and structural abnormality detection — is more recent and has a thinner post-market evidence base.
Cleared Tool Categories
- 12-lead ECG analysis: atrial fibrillation detection, STEMI flagging, QTc prolongation
- Wearable ECG AI (single-lead, consumer-grade devices with FDA De Novo authorization)
- Echocardiography: automated left ventricular ejection fraction, chamber measurement
- CT angiography: coronary artery stenosis quantification, fractional flow reserve (CT-FFR)
- Cardiac MRI segmentation and function quantification
Active Trials
Several registered trials are evaluating whether AI-assisted ECG interpretation changes clinical outcomes — not just diagnostic accuracy. NCT05273021 (AI-ECG for low-ejection-fraction detection) and NCT04601415 (screening for structural heart disease via AI-ECG in primary care settings) are among the trials with enrollment documented as of early 2026. These are worth tracking because they target outcome endpoints rather than surrogate performance metrics.
Equity Concerns
ECG interpretation AI trained on predominantly male datasets has shown documented differences in sensitivity for conditions that present differently by sex — including certain arrhythmias and ST-segment patterns. Wearable ECG AI cleared for consumer devices raises a distinct equity concern: the populations most likely to purchase and use these devices are not the same populations with the highest cardiovascular disease burden. Deployment in health system settings partially addresses this, but access gaps persist.
Pathology
Digital pathology AI — analyzing whole-slide images (WSI) for cancer detection, grading, and biomarker prediction — has moved from research to cleared device status in several applications. Prostate cancer grading (Gleason scoring) and breast cancer lymph node metastasis detection are the most developed cleared categories. Predictive biomarker applications (predicting microsatellite instability or HER2 status from H&E slides without IHC) are in active clinical trials but not yet broadly cleared.
| Application | Regulatory Status | Evidence Maturity | Key Limitation |
|---|---|---|---|
| Prostate Gleason grading AI | FDA-cleared (multiple vendors) | Moderate — prospective studies exist, mostly single-institution | Interobserver variability in training labels |
| Breast lymph node metastasis detection | FDA-cleared | Limited post-market RWE | WSI scanner hardware variability affects model performance |
| MSI prediction from H&E | Investigational — active trials | Retrospective only, no cleared device | Not validated across cancer subtypes |
| HER2 scoring augmentation | Some cleared, some investigational | Mixed — concordance studies available | Heterogeneous tumor expression patterns limit AUC generalizability |
Equity Concerns
Pathology AI training datasets are heavily concentrated in academic medical centers with high-volume cancer programs. Demographic representation in training cohorts is rarely reported at the level of granularity needed to assess bias. One documented concern: prostate cancer grading models trained on predominantly non-Hispanic white patient cohorts have not been systematically validated on patient populations with documented differences in tumor morphology. Staining protocol variation across labs — a known technical confounder — is also inconsistently addressed in validation studies.
Gastroenterology
Gastroenterology has a narrower but well-defined cleared AI footprint. Colonoscopy AI — specifically, computer-aided detection (CADe) of colorectal polyps — is the most established application with the most published prospective RCT data of any GI AI tool. Several CADe systems have FDA clearance. The evidence for polyp detection improvement is reasonably consistent across trials, though adenoma detection rate (ADR) effect sizes vary by baseline endoscopist ADR and study setting.
Cleared Tool Categories
- Colonoscopy CADe: real-time polyp detection during colonoscopy
- Colonoscopy CADx: optical characterization of detected polyps (diminutive lesion characterization)
- Capsule endoscopy AI: automated small bowel lesion flagging to reduce review time
Colonoscopy CADe is notable because it has more RCT evidence than almost any other AI-in-endoscopy application. Multiple prospective trials have shown statistically significant improvements in adenoma detection rate compared to unassisted colonoscopy, with effect sizes generally in the range of 3–6 percentage points. However, several trials also show that the benefit concentrates in lower-baseline-ADR endoscopists, and that high-volume experienced endoscopists may see minimal additive benefit.
Equity Concerns
Colorectal cancer screening rates remain unequal by race, insurance status, and geography. CADe systems improve detection during colonoscopy — but they cannot address the access gap in who receives colonoscopy in the first place. There is also limited published data on CADe performance across different bowel preparation quality levels, which correlates with socioeconomic factors affecting patient compliance with prep instructions.
Primary Care
Primary care AI is more fragmented than the other specialties in this landscape. The cleared device footprint is smaller, but the deployment footprint — particularly for AI scribes, clinical decision support tools, and chronic disease risk models — is arguably larger in terms of patient touchpoints. This creates a gap: many AI tools operating in primary care settings are not classified as medical devices and therefore are not subject to FDA premarket review.
Cleared and Regulated Applications
- Diabetic retinopathy screening (autonomous AI, De Novo authorized — typically deployed in primary care settings without an ophthalmologist present)
- Sepsis prediction algorithms (some cleared as SaMD, others deployed as non-device clinical decision support)
- Hypertension and cardiovascular risk stratification tools (regulatory status varies by intended use)
AI Scribe and Ambient Documentation
Ambient AI documentation — tools that listen to clinical encounters and generate structured notes — has seen the fastest adoption growth of any AI application in primary care. These tools generally do not require FDA clearance because they are positioned as documentation aids rather than diagnostic or treatment decision tools. The evidence base is thin: most published data comes from health system-sponsored implementation reports rather than controlled studies. Hallucination risk — the generation of plausible but inaccurate clinical documentation — is a documented concern that most vendors acknowledge but few have published systematic failure rate data on.
Equity Concerns
Primary care AI equity concerns are among the most complex across all specialties. Chronic disease risk models trained on electronic health record data inherit the biases in that data — including documented differences in care intensity, coding practices, and diagnostic rates by race, insurance type, and geography. Ambient documentation tools trained predominantly on English-language encounters perform worse for patients who communicate in other languages or with non-standard speech patterns. The autonomous diabetic retinopathy screening device (IDx-DR, now Luminare Health) was cleared with a specific intended use population; deployment outside that population's characteristics requires separate evaluation.
Cross-Specialty Patterns
Several themes recur across all five specialties and are worth naming explicitly rather than burying in specialty-specific sections.
| Pattern | Where It Appears | Practical Implication |
|---|---|---|
| Retrospective-to-prospective gap | All specialties | Most cleared devices have stronger retrospective than prospective evidence; post-market studies are inconsistently required |
| External validation deficit | Radiology, pathology, primary care | Models validated at one institution routinely underperform at others; single-site AUC figures should not be treated as generalizable |
| Non-device AI operating without premarket review | Primary care, cardiology (wearables), GI (some CDS) | Large patient exposure without FDA oversight; requires evaluation through other frameworks |
| Training dataset demographic gaps | All specialties | Underrepresentation of specific racial, ethnic, age, and socioeconomic groups is documented across nearly every specialty's foundational datasets |
| Outcome endpoint scarcity | All specialties | Most trials measure surrogate endpoints (detection rate, AUC, time savings); patient outcome data (mortality, morbidity, cost) is rare |
Dominant Companies by Specialty
Company presence across specialties is uneven and shifting. The following reflects the concentration of FDA-cleared products and published peer-reviewed studies as of Q2 2026 — not market share estimates, which carry known methodological problems.
| Specialty | Companies with Multiple Cleared Devices | Notable Investigational Players |
|---|---|---|
| Radiology | Aidoc, Viz.ai, Nanox.AI, Intelerad (acquired MedAI assets), Siemens Healthineers, GE HealthCare | Enlitic, RayzeBio imaging division |
| Cardiology | AliveCor (KardiaMobile), Eko Health, Caption Health (GE), HeartFlow | Cardiologs (Philips), Cleerly |
| Pathology | Paige.AI, PathAI, Proscia, Hologic (digital pathology) | Ibex Medical Analytics |
| Gastroenterology | Medtronic (GI Genius), Fujifilm (CAD EYE), Olympus (EndoBRAIN) | Odin Vision (acquired by Olympus) |
| Primary Care | Luminare Health (IDx-DR), Nuance (DAX ambient AI, Microsoft) | Abridge, Suki, Nabla |
Unresolved Clinical Questions
Across all five specialties, several questions remain genuinely open — not because the research is immature, but because the trials needed to answer them have not been completed or, in some cases, not yet started.
- Does radiology AI triage (e.g., critical finding flagging) reduce time-to-treatment and improve patient outcomes, or does it primarily reduce radiologist workload without downstream clinical benefit?
- For colonoscopy CADe, does improved adenoma detection rate translate to reduced colorectal cancer incidence and mortality at the population level?
- Can ECG AI deployed in primary care settings (via wearables or standard 12-lead) identify high-risk patients early enough to alter outcomes, or does it primarily generate referrals without net clinical benefit?
- Does ambient AI documentation reduce clinician burnout in a measurable, durable way, or do initial time savings erode as documentation complexity increases?
- How should AI-predicted biomarkers (e.g., MSI status from H&E) be validated before replacing laboratory-based testing in treatment decision pathways?
How to Use This Landscape Page
This page is a navigation and synthesis layer. It is designed to orient you to a specialty's AI landscape and then route you to the specific records that support deeper verification.
- To verify whether a specific tool is FDA-cleared, use the FDA AI Device Records section, filtered by specialty and company.
- To assess the evidence quality for a specific tool or application, consult the relevant evidence appraisals, which include study design, dataset characteristics, and limitations.
- To understand how a tool performs in real deployment conditions (not controlled study settings), see the clinical deployment reports section.
- To track regulatory changes affecting a specialty's AI tools, the regulatory tracker records FDA guidances, final rules, and enforcement actions by affected device category.
Comments
Join the discussion with an anonymous comment.