Medical imaging AI

Medical imaging AI is the most mature slice of clinical AI. It has the deepest bench of FDA-cleared products, the largest published-evidence base, and the longest track record of real-world deployment inside health systems. The FDA’s AI-enabled device list crossed a thousand entries in 2026, and imaging is the plurality of that list by a wide margin.

What actually got cleared

Reading down the FDA list, four patterns dominate imaging:

• Triage and prioritization (CADt). Move the likely-emergency case to the top of the radiologist’s worklist — the classic use case is large-vessel-occlusion detection on head CT, which turned Viz.ai into a category leader. Similar tools now exist for pulmonary embolism, aortic dissection, intracranial hemorrhage, and pneumothorax.
• Computer-assisted detection (CADe). Second-reader-style flags on mammography, chest X-ray nodules, colonoscopy polyps. The evidence base here is the longest and the most rigorously scrutinized.
• Quantification. Volume measurements (hippocampus in Alzheimer’s workup, MS lesion load, ejection fraction) done automatically, more consistently than a human, and integrated into the report.
• Image enhancement and reconstruction. Denoising, low-dose reconstruction, motion correction — often built into the scanner pipeline itself, invisible to the radiologist. GE HealthCare, Siemens Healthineers, Philips, and Canon have all shipped AI-augmented reconstruction on their newer scanners.

Where the evidence actually sits

The imaging-AI evidence base is deeper than other clinical-AI subfields, but it is uneven. Screening mammography and diabetic-retinopathy screening have accumulated enough prospective evidence — some of it from national screening programs in Europe and Asia — to have moved into the “settled science” tier. Chest-X-ray triage and stroke-triage tools have strong controlled-deployment evidence but the population-level outcomes claims are still contested. Newer subfields — pathology whole-slide analysis, cardiology (echo, ECG), and ophthalmology beyond diabetic retinopathy — have promising evidence but are earlier on the diffusion curve.

The deployment questions

When we cover an imaging-AI product, the four questions we return to are:

1. How was it evaluated, and on whom? Most pivotal datasets remain heavily skewed toward U.S. and Western European populations and toward a specific scanner manufacturer. Coverage of the FDA’s Predetermined Change Control Plan (PCCP) guidance explains how manufacturers are increasingly committing to defined post-market changes at submission time.
2. What integrates with what? A high-performing model that requires a parallel viewer usually does not survive contact with a working radiology department. Native PACS integration is the difference between a pilot and a rollout.
3. What is the failure mode? Sensitivity/specificity numbers in the FDA summary are a starting point, not the whole story. Radiologists want to know how the tool behaves on the edge cases their department actually sees.
4. Who pays? Medicare’s evolving reimbursement stance on AI — the New Technology Add-on Payments (NTAP), CPT category-III codes, and MAC-level local coverage — is the plumbing that decides whether a product survives past the pilot budget.

What we cover

Our reporting in this topic focuses on: new FDA clearances of note, PCCP-enabled post-market updates, evidence and post-market surveillance signals from MAUDE and the Joint Commission RUAIH certification program, CMS reimbursement decisions, and the health-system procurement realities of imaging AI at scale. Explore related articles and news below.