A new AI model learns to compare multiple medical images instead of examining just one at a time.
Researchers built PMC-MI, a training set of 234,956 instruction examples pulled from compound figures in biomedical journal articles, the multi-panel images common in medical papers. Of those, 10,555 multi-image instances were curated for reinforcement learning and checked by medical reviewers. The team trained a model called M3LLM using a three-stage process: standard fine-tuning, reinforcement learning that teaches the model which images actually answer the question asked, and a final broad fine-tuning pass. On the team's own benchmark, PMC-MI-Bench, M3LLM outperformed the strongest existing multimodal models on single-image, multi-image, and relative-position tasks, and it also posted new high scores on two independent public medical benchmarks, OmniMedVQA and MMMU-Med.
Most medical AI tools are built and tested on one image at a time, but real diagnosis often means comparing a chest X-ray from six months ago against today's, or weighing several views of the same lesion. Mining multi-panel figures already published in medical journals is a cheap way around the usual bottleneck of scarce, privacy-restricted hospital imaging data, and it's a technique other labs could copy without needing clinical partnerships.
The approach was also tested on real longitudinal chest X-rays and a dermatology case set, which matters more than any benchmark score, though claiming state of the art on a benchmark you built yourself is worth reading with a raised eyebrow.