AI in medical diagnosis
Algorithms read scans and slides and are now routine tools, mostly in radiology.
Open in the interactive tree →In April 2018 the FDA allowed the first autonomous AI diagnostic system (IDx-DR, which detects diabetic eye disease without a doctor reading the image). Deep-learning models now flag strokes, fractures, lung nodules and breast cancer on images, and large language models are tested for clinical reasoning and documentation. The promise is faster, more consistent and cheaper diagnosis; the risks are bias, over-reliance and unclear liability.
As of October 2026
By the end of 2025 the FDA list held about 1,450 AI-enabled devices (about 1,600 by June 2026 per third-party counts), roughly three quarters in radiology; the FDA notes the list is not comprehensive. Only a few AI decision-support systems have so far been tested prospectively for clinical benefit when physicians use them in practice.
Open steps
- Trials proving patient benefit Medium AI leverageFew tools have randomised evidence that patients do better; trials must show effects on outcomes, not only reading accuracy.
- Staying accurate on new scanners High AI leveragePerformance drops on new machines, hospitals and populations; deployed models need monitoring and updates that regulators accept.
- Language models inside real workflows Medium AI leverageLLMs score well on case vignettes, yet physicians using them did no better in a trial; designs that fit real consultations and stay safe must be tested.
- One model for many findings High AI leverageA foundation-model-based CT triage system for 14 findings is FDA-cleared; a cleared general-purpose model across organ systems with multisite rare-condition proof is still missing.
Where AI could help
Medium AI leverage. The tools exist; what remains is prospective proof, workflow integration and regulation, which AI cannot shortcut.
- Triage that cuts radiologist workload
- Detecting subtle cancers on scans and slides
- Monitoring deployed tools to catch performance drift
- Combining images with health records for risk prediction
Shown so far
- In January 2025 Nature Medicine reported the PRAIM study of 463,094 German screening women: AI-supported reading found 6.7 versus 5.7 cancers per 1,000 (17.6% more) without a higher recall rate. source
- In February 2026 The Lancet published the MASAI trial of about 106,000 women: AI-supported mammography screening had fewer interval cancers (1.55 versus 1.76 per 1,000) than standard double reading. source
Prerequisites
- Electrocardiogram1903AI reading of ECG traces is one of the widest-used medical AI tools
- Medical ultrasound1958AI now guides and reads ultrasound scans at the bedside
- CT and MRI scanning1971
- Deep Learning2012
- Large Language Models2022