A clinical multimodal vision language foundation model with fine-grain explainability

5 Key Takeaways

1. EVLF-FM achieved the highest average accuracy of 0.895 in internal validation for disease diagnostics, outperforming leading generalist and specialist models.
2. In medical visual grounding, EVLF-FM achieved an average accuracy at 0.5 intersection over union (Acc@0.5) of 0.842 across nine modalities.
3. External testing confirmed consistent zero-shot and few-shot performance with competitive F1 scores despite a smaller model size.
4. A hybrid training strategy combining supervised and visual reinforcement fine-tuning enables clinically interpretable explanations aligned with visual evidence.
5. Human evaluation using the S.C.O.R.E. framework yielded a high overall mean score of 4.25/5, supporting clinical interpretability of reasoning outputs.

OpenRounds transcript synthesis · Verify against the source