A new AI system called CineMR can pull actual numbers out of heart MRI videos instead of just describing what it sees.
Researchers built CineMR, a vision-language model that calls external image-analysis tools - for segmentation, phase selection, volumetry, and wall-motion tracking - and folds their outputs back into its own reasoning before answering. They trained it in two stages: supervised fine-tuning on examples of tool use, then a reinforcement-learning step called GRPO that rewards correct tool calls. On a new multi-cohort benchmark testing quantitative metrics, diagnosis, and differential diagnosis from cine cardiac MRI, CineMR hit 35.9% pass@1 and 58.9% pass@4. Its own Qwen3-VL-8B backbone managed just 1.5% pass@1 without the tool scaffolding, and other medical vision-language models like LLaVA-Med and MedGemma-4B scored even lower.
The gap between CineMR and its unassisted backbone is the real story: tool use, not raw model size, closes most of the distance to reliable quantitative reading of cardiac scans. Correct tool invocation jumped from 78.9% to 99.8% after the reinforcement-learning stage, and feeding the model live tool outputs improved measurement accuracy by 20.4-23.7% over letting it guess numbers directly. That suggests the fastest path to trustworthy medical AI may be wiring existing, already-validated analysis software into language models, rather than waiting for ever-bigger models to learn anatomy from pixels alone.
Worth noting that 35.9% pass@1 still means the model gets the full answer wrong roughly two-thirds of the time, so this reads as a promising architecture paper, not a tool any cardiologist should trust unsupervised yet.