Internal Medicine · 2 h ago
Multimodal medical AI model shows strong performance across imaging tasks
Researchers developed EVLF-FM, a vision-language model trained and evaluated on more than 1.3 million samples spanning 10 imaging modalities and six specialties. It achieved an average internal diagnostic accuracy of 0.895 and was externally tested on 13,443 samples, but prospective studies are still needed to assess clinical safety and utility.
- Internal diagnostic accuracy averaged 0.895.
- External validation included 13,443 samples.
- Grounding accuracy averaged 0.842 at Acc@0.5.
- Prospective clinical evaluation remains necessary.
Researchers developed EVLF-FM, a multimodal medical vision-language foundation model for image-based diagnosis, visual question answering and image-region localization. Development and testing covered more than 1.3 million samples from 18 datasets, across 10 imaging modalities and six specialties. External validation used 13,443 independent samples from 17 additional datasets. The model was compared with generalist vision-language models and specialist foundation and deep-learning models; investigators also assessed its explanatory outputs through human evaluation.
In internal diagnostic benchmarks, EVLF-FM achieved a mean accuracy of 0.895, the highest among the evaluated models. It also achieved an average Acc@0.5 of 0.842 for visual grounding across nine modalities. External testing found competitive zero-shot and few-shot performance, including F1 scores, while human ratings of explanations averaged 4.25 out of 5 using the S.C.O.R.E. framework.
The findings suggest broad technical capability across imaging tasks, but do not establish clinical benefit or safety. The authors note that the explanations are post hoc and that further validation in diverse healthcare settings is needed. Prospective clinical studies are required before routine deployment; benchmark performance alone does not show that the model improves decisions or outcomes.
In practice
EVLF-FM performed well across diagnostic, visual-question-answering and image-grounding benchmarks spanning multiple modalities, including external testing on 13,443 samples. These findings do not support adopting the model for clinical decisions: the work does not establish safety, effectiveness or improved patient outcomes, and the authors call for prospective clinical studies and validation across diverse care settings. Its textual explanations were human-rated, but are described as post hoc, so their scores do not establish that the explanations faithfully represent the model’s decision process.
AI-generated clinical interpretation · Not editor-reviewed
Is this summary clinically accurate?
Help fellow clinicians: your rating sends inaccurate summaries straight to our editors.
Sign in to rate this summary →Source
Nature Medicine: A clinical multimodal vision language foundation model with fine-grain explainability ↗This is an automated AI-condensed summary that has not yet been reviewed by an editor. Always consult the full item at the original source.