If your strategic view of medical AI market structure treats the FDA-cleared device count as an indicator of broad clinical AI adoption across specialties (implying that AI is becoming pervasive across cardiology, neurology, pathology, oncology, and other clinical domains), the December 2025 FDA authorisation distribution shows a structure that is heavily concentrated. The FDA had authorised 1,357 AI/ML-enabled medical devices from 693 different companies across 17 clinical specialties. Radiology accounts for 76.6% of authorisations (1,039 devices). Cardiovascular: 9.6% (130 devices). Neurology: 4.5% (61 devices). Every other specialty combined: under 10%. The market is heavily concentrated by specialty (radiology) and broadly distributed by company (most companies hold only one or two authorisations). The two concentrations interact in ways that shape clinical AI deployment strategy.

FDA-authorised AI/ML medical devices by specialty:

The radiology concentration has deep structural drivers. Three explain the bulk of the pattern.

The first driver: imaging data has been the workhorse training corpus for medical AI. Radiology produces structured imaging datasets (CT, MRI, X-ray, mammography) at scale, with established clinical workflows for image acquisition, interpretation, and reporting. The data infrastructure has been more conducive to AI development than data infrastructure in other clinical domains. AI model development concentrated in imaging accordingly.

The second driver: radiology has clearly defined decision points (presence/absence of finding, classification of finding type, measurement of finding characteristics) that map cleanly onto supervised machine learning tasks. Other clinical specialties have more diffuse decision-making: pathology has decision-point structure (similar to radiology, hence pathology's growth) but cardiology, neurology, and internal medicine often have decision processes that integrate many data sources, clinical judgement, and patient-specific factors that resist single-task ML formulation.

The third driver: regulatory pathway alignment. The FDA 510(k) pathway works well for incremental imaging AI improvements, where substantial equivalence to predicate imaging devices is straightforward to demonstrate. The same pathway is less well-suited to genuinely novel clinical AI applications in non-imaging domains, where predicate devices may not exist. The regulatory channel itself contributes to radiology concentration.

The company distribution is the inverse pattern. Of 693 companies with at least one FDA-authorised device, the large majority hold only one or two. GE Healthcare leads with 93 devices, Siemens Healthineers 82, Shanghai United Imaging 38, Philips Healthcare 36, Canon Medical Systems 35, Aidoc Medical 30. Qure.ai Technologies 8, RaySearch Laboratories 8, Zebra Medical Vision 9. The top three companies hold 213 of the 1,357 devices (16%); the remaining 84% is spread across hundreds of smaller firms.

FDA-authorised AI/ML medical devices by top company:

The two concentrations, by specialty (heavy in radiology) and by company (broad across many firms), produce a specific market structure with implications for clinical deployment.

Three data observations follow.

The first observation: radiology AI procurement is the highest-volume but most commoditised segment. With 1,039 authorised devices, radiology departments have many options, but the options are increasingly similar in capability (substantial equivalence reflects shared technical approaches). Differentiation increasingly comes from clinical workflow integration, deployment infrastructure, and outcomes measurement rather than core algorithm performance.

The second observation: non-radiology AI procurement faces structurally different market dynamics. Cardiovascular AI has 130 authorised devices, far fewer options than radiology, but each option may serve more specific clinical use cases. Neurology has 61. Other specialties have substantially fewer. Procurement in these areas operates with fewer choices and weaker substitute pressure on pricing or performance.

The third observation: the company distribution favours established medical imaging vendors. GE Healthcare and Siemens Healthineers lead the device count because they have decades of established relationships with hospital systems, established regulatory expertise, established product distribution channels, and (for many devices) substantial-equivalence claims relative to their own predicate devices. The structural advantage for established vendors makes market entry harder for AI-focused startups even when their technology is competitive.

Non-radiology authorisations have grown from 7 in 2016 to 60 in 2025 (see chart: FDA AI/ML medical devices by specialty over time, 2016–25). Cardiology, neurology, anaesthesiology, and gastroenterology-urology have all seen acceleration since 2020. The radiology dominance is being modestly diluted as other specialties build authorisation volume, but the pace of dilution is slow relative to the radiology base.

Three implications follow for health systems setting AI device procurement strategy in 2026.

The first implication: radiology AI procurement should be treated as a workflow integration challenge more than an algorithm selection challenge. With 1,039 radiology AI devices authorised, the variation in core algorithm capability is smaller than the variation in workflow integration, clinical impact measurement, and operational outcomes. Procurement that optimises algorithm benchmarks may miss the actual deployment differentiators. Procurement that optimises clinical workflow integration and outcomes measurement may identify the better deployment partners regardless of algorithm performance.

The second implication: non-radiology AI procurement requires more careful evaluation because fewer options exist. The 130 cardiovascular devices, 61 neurology devices, and smaller specialty counts mean procurement teams need deeper individual device evaluation rather than market-comparison evaluation. The structural data is thinner; the per-device evaluation needs to be more rigorous.

The third implication: the cross-specialty AI deployment portfolio needs to be planned with awareness of the market structure. A health system aiming to deploy AI broadly across specialties will face very different procurement dynamics in radiology (many options, commoditised choices) than in cardiology (fewer options, less substitute pressure) than in neurology (still fewer options). The deployment plan, vendor relationship strategy, and clinical evidence approach should be specialty-specific rather than generic.

The data observation: the FDA authorisation concentration reflects historical AI development patterns and regulatory pathway design. It does not reflect underlying clinical opportunity or where AI can produce clinical benefit. As non-imaging clinical AI matures over 2026-2028, particularly in areas like clinical decision support, drug-dosing AI, and predictive analytics where the FDA authorisation pathway is less central, the deployment opportunity will broaden faster than the FDA authorisation count suggests.

For health systems setting medical AI strategy in 2026, the data observation calibrates expectations: the FDA authorisation distribution shows where the established commercial market is, not where the clinical opportunity is. Strategic planning that engages with the broader deployment landscape (FDA-authorised devices, FDA-exempt clinical decision support tools, internally developed AI, and AI integrated into clinical workflow tools that don't require FDA authorisation) produces a more accurate picture of where AI can be deployed than the FDA authorisation count alone suggests.

The trajectory: through 2026-2028, the FDA authorisation count will continue to grow with continued radiology dominance, gradual non-radiology growth, and continued broad company distribution. The strategic implication: medical AI deployment opportunity will increasingly be shaped by what happens outside the FDA-authorised device count, in clinical decision support, predictive analytics, and workflow AI that operate under different regulatory frameworks. Plans that engage with this broader landscape will be aligned; plans that read the FDA count as the comprehensive AI market measure will be operating from a partial view.