Researchers boost AI accuracy in medical image diagnosis by combining two training techniques — accuracy and reasoning both improve simultaneously

A research team published a new method that trains AI systems (called LLMs, which understand and generate text) on radiology reports (written descriptions of medical scans) using two sequential steps: first standard supervised training on disease labels, then a reinforcement learning technique called GRPO that refines the AI's answers without explicit reasoning supervision. Tests across three radiologist-verified datasets showed both accuracy gains and improved reasoning quality.

The breakthrough matters because previous methods faced a tradeoff: teaching AI to be more accurate at disease classification often made its reasoning worse (less explainable to doctors). This approach eliminates that tradeoff — the AI now outputs more correct diagnoses *and* provides clearer explanations of how it reached those conclusions, which hospitals need for clinical adoption.

For radiologists and hospital administrators, this makes AI-assisted diagnosis more trustworthy: doctors can now rely on both the AI's final answer and understand *why* it made that call, reducing the friction for integrating AI into routine scan reviews. For medical AI companies building diagnostic tools, this is an immediate technique to adopt — no new hardware required, just better training methods.