Most medical imaging models are specialists. A dermatology network reads skin, a pathology network reads slides, and neither can explain itself in terms a clinician can audit. A team led by researchers at A*STAR’s Institute of Advanced Intelligence and Computing and the Singapore Eye Research Institute set out to collapse that fragmentation into one model that also shows its work.
The result, published in Nature Medicine, is EVLF-FM: an explainable vision language foundation model for medicine. Its training and evaluation set drew above 1.3 million samples across 10 modalities and 18 datasets, covering six clinical fields from dermatology and hepatology to ophthalmology, pathology, pulmonology and radiology.
One model, six specialties
On internal disease diagnostics, EVLF-FM posted the highest average accuracy at 0.895, beating both generalist and specialist models it was compared against. It also handles visual question answering and pixel-level grounding, pointing to the exact region behind its answer rather than returning a bare label.
That grounding held up outside the training data. External validation used 13,443 independent samples from 17 additional datasets, and the model kept competitive F1 scores in zero-shot and few-shot settings despite a smaller footprint than many rivals.
Making the reasoning legible
The team used a hybrid strategy combining supervised training with visual reinforcement fine-tuning, which teaches the model to generate explanations aligned to the visual evidence rather than plausible-sounding narration. Clinician evaluation using the S.C.O.R.E. framework gave those explanations a mean of 4.25 out of 5.
The authors are careful about what this does not prove. Performance varies across healthcare settings, and they call for prospective clinical studies to establish safety, effectiveness and real utility before anyone deploys it routinely. Accuracy on curated datasets and accuracy on a Tuesday night in a busy emergency department remain very different tests.
