Nature Medicine AI model reads 10 imaging modalities and shows its reasoning

Nature Medicine published a study on 29 October 2026 describing EVLF-FM, a vision language foundation model that reads images from 10 modalities across six clinical specialties and explains its diagnoses by pointing to the parts of an image behind them¹. The authors present it as an answer to two complaints that have slowed clinical uptake of foundation models: most handle a single modality, and few show how they reached an answer¹.

What the study found

The model was built and tested on more than 1.3 million samples drawn from 18 datasets¹. The specialties covered were dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology¹. For external validation the team used 13,443 test samples from 17 further datasets, again spanning 10 imaging modalities¹.

EVLF-FM does three jobs: it gives a diagnosis, answers questions about an image, and marks the pixels that support its reasoning¹. On internal validation for diagnosis it posted an average accuracy of 0.895, which the authors report as the highest among the leading generalist and specialist models it was compared with¹. For visual grounding, the task of locating the relevant region, it scored an average Acc@0.5 of 0.842 across nine modalities, a level the paper calls reasonable rather than leading¹.

On the external datasets the model held up in zero-shot and few-shot settings, and its F1 scores were competitive with rivals even though the model is smaller¹. Its explanations come from a hybrid training method that pairs supervised fine-tuning with visual reinforcement fine-tuning¹. Human reviewers scored those explanations with the S.C.O.R.E. framework and gave an overall mean of 4.25 out of 5¹.

What it means for the market

The paper frames modality-specific design as a brake on clinical adoption¹. For buyers in NHS trusts and European hospitals, which typically assess imaging AI product by product, a single model that outperforms specialist tools across six specialties suggests one platform could in principle cover several departments. The study does not test that in a hospital, so the commercial case remains unproven¹.

The explainability work is the part most relevant to regulators and clinical safety officers. A model that shows the image evidence for each finding gives reviewers something to check, and the 4.25 out of 5 rating from human evaluators is the paper's evidence that clinicians found the reasoning interpretable¹. Whether that would satisfy medical device regulators, or how the model would be classified, is not addressed in the abstract.

For suppliers of single-modality tools, the claim that a smaller generalist model can stay competitive on unseen data is the challenge to watch¹. The abstract does not say who would commercialise EVLF-FM, at what price, or whether it is licensed for any market, so its effect on competition cannot yet be judged.

Caveats

The model has not yet been tested prospectively in clinical care¹. The authors say the model still needs validation across a wider range of healthcare settings¹. Grounding accuracy was lower than diagnostic accuracy, at 0.842 against 0.895¹. The authors state that prospective clinical studies must assess its safety, effectiveness and clinical usefulness before any routine deployment¹.