FDA-approved AI algorithms: 20%+ CAGR, radiology still dominates, but hospitals are building their own

The number of FDA-approved AI algorithms has been growing at a 20%+ CAGR since 2020. Radiology accounts for 76% of the total approvals. That share has started to shrink over the past year, with cardiovascular and neurology gaining traction. However, many FDA 'radiology' listings are simply image-based, even when the real clinical domain is oncology, cardiovascular, or another specialty. So the count of radiology algorithms may actually be inflated. A more precise classification from the FDA would be welcome to better understand where AI is truly making clinical impact.

The biggest medical imaging incumbents - GE, Siemens Healthineers, and Philips - clearly lead the list in number of FDA-approved algorithms. In recent months, GE acquired icometrix, RadNet acquired Gleamer, and Philips acquired DiA Imaging, further strengthening the incumbents' positions. M&A is becoming a key strategy for incumbents to consolidate their AI portfolios rather than building everything in-house.

The bigger trend is that hospitals are increasingly building their own algorithms rather than outsourcing to MedTech vendors. There are a few reasons for this. Capturing the real value of an AI algorithm takes extra work from providers: these tools are not 'off the shelf.' FDA approval does not mean clinical value. A lot of effort goes into making algorithms more accurate by training them on the hospital's own data and into integrating them into the clinical workflow. Clinical decision support sits at the core of what a hospital does - it is too strategic to outsource. Vendor lock-in is exactly what many leading providers want to avoid in the age of AI.

The FDA list is full of duplicates: many algorithms have to be recertified whenever changes are made. There is a real disconnect here. AI is developed to be improved continuously, while FDA certification only evaluates a model at a single point in time. Every update, retrain, or improvement triggers a new submission process. A number of companies are now tackling this pain point, pushing regulators to adapt their logic to the reality of AI - where models are living systems that evolve with new data, not static software releases.