Medical AI model EVLF-FM achieves top average diagnostic accuracy across multiple imaging types while explaining its reasoning

A new multimodal vision-language foundation model for medicine, called EVLF-FM, achieved the highest average diagnostic accuracy among tested models across a broad set of imaging tasks while also generating explanations that clinicians rated highly for safety and objectivity, researchers report in Nature Medicine. The work, led by senior author Daniel Shu Wei Ting at the Singapore Eye Research Institute and Singapore National Eye Center, with first author Yang Bai, developed the model using more than 1.3 million samples from 18 datasets across 10 imaging modalities covering six clinical specialties: dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology.

The study is a model development and benchmarking analysis, not a prospective clinical trial. EVLF-FM supports three core tasks: disease diagnosis, medical visual question answering, and pixel-level visual grounding with reasoning. Training combined supervised fine-tuning with visual reinforcement fine-tuning, using a sentence-level clinical reasoning alignment reward to encourage explanations that reflect visual evidence rather than memorized answers.

In internal validation across 11 diagnostic tasks, EVLF-FM reached a mean accuracy of 0.895, outperforming the generalist vision-language models and specialist models it was compared against. It also achieved strong visual grounding, with an average accuracy at 0.5 intersection over union (Acc@0.5) of 0.842 across nine imaging modalities. External testing on 13,443 independent samples from 17 additional datasets showed consistent zero-shot and few-shot performance, with competitive F1 scores despite the model being smaller (8 billion parameters) than some comparators, such as MedGemma 27B and MedDr 32B.

To assess the model's reasoning outputs, 12 physicians—four junior doctors and eight specialists in radiology, pathology and ophthalmology—graded cases from five imaging modalities using the S.C.O.R.E. framework (safety, consensus and context, objectivity, reproducibility, and explainability). The overall mean score was 4.25 out of 5, and inter-grader agreement ranged from substantial to almost perfect. Objectivity scored highest, suggesting the generated reasoning was generally supported by identifiable imaging findings.

The potential impact is broad: a single generalizable model that works across many imaging types and explains its reasoning could support triage, decision support and medical education, particularly in settings with limited specialist access. However, the authors caution that EVLF-FM is not ready for routine clinical deployment. Its current accuracy does not yet reach the level typically required for direct clinical use (for example, consistently exceeding about 90% across datasets), and performance on visual question answering was generally lower than on diagnostic classification. The model also processes only static, two-dimensional images and does not handle three-dimensional volumetric data (such as full 3D CT or OCT volumes). Like other large-scale models, it may inherit biases from training data, so performance in underrepresented populations or settings could be less reliable.

The researchers position EVLF-FM as a generalizable backbone for further fine-tuning, not a finished product. They call for prospective clinical studies to evaluate its safety, effectiveness and real-world utility, and for broader validation across diverse healthcare settings before any clinical deployment.

Reference: Bai, Y., Cheng, H., Zhou, Y. et al. A clinical multimodal vision language foundation model with fine-grain explainability. Nat Med (2026). https://doi.org/10.1038/s41591-026-04655-5