Medical Artificial Intelligence · global
Aligning diagnoses with images: Medical AI model EVLF-FM combines interpretation with visual explanations
From retinal images to pathology slides, EVLF-FM attempts to identify diseases while highlighting relevant image regions. Testing across modalities provides preliminary evidence, but whether its explanations are reliable and can improve real-world clinical care still requires prospective research.
When physicians review an AI image interpretation, they need to check not only the disease name but also which clues in the image support that judgment. A study published on October 29 in Nature Medicine introduced EVLF-FM, which integrates disease recognition, visual question answering, and region localization into a single model, aiming to provide diagnostic suggestions accompanied by visual evidence that can be checked.
Development and testing in the formally published paper covered more than 1.3 million samples across 10 imaging modalities and six clinical specialties, including ophthalmology, dermatology, and pathology; an additional 13,443 independent samples were used for external validation. This design was intended to test whether the model could work across different forms of imaging, rather than being familiar with only one type of data.
EVLF-FM combines supervised fine-tuning with visual reinforcement fine-tuning. In addition to answering image-related questions, it can link text descriptions to image regions at the pixel level. For clinical use, the potential value of this capability is that physicians can compare each abnormality identified by the model with the original image.
The journal version reported an average accuracy of 89.5% in internal validation of disease diagnosis, higher than the comparison models in the study. However, this is an average result for specific data and tasks. It cannot be taken directly as the accuracy achievable for every disease or at every hospital, nor does it establish that the model can improve patient care.
The study's figures also changed between versions. The arXiv preprint released in September 2025 reported 11 imaging modalities, 8,884 external validation samples, and an average internal diagnostic accuracy of 85.8%; the formal publication updated the scope of the data and the results. Figures from the two versions cannot be used interchangeably, and the difference in accuracy alone cannot be interpreted as the extent of improvement in the model's capabilities.
The ability to identify image regions still does not mean that the model fully reveals its decision-making process. The paper defines these outputs as post hoc explanations and evaluates their consistency with the evidence; clear presentation and correct judgment therefore remain separate questions that require separate evaluation.
The research team noted that validation in more diverse healthcare settings is still needed before routine clinical deployment, along with prospective studies to assess safety, effectiveness, and clinical utility. The key question ahead is whether these verifiable explanations can help physicians identify errors and make more reliable judgments in real-world clinical workflows.