← Back to Home

Medical AI spans 10 imaging modalities: EVLF-FM links diagnostic answers to lesion locations

From identifying diseases to highlighting imaging evidence, EVLF-FM aims to make medical AI judgments easier to verify. The study shows potential across imaging tasks, but performance on test datasets remains some distance from a reliable assistant in clinical practice.

By SURL BioNews

When medical imaging AI provides only a disease name, physicians still need to ask: What exactly did it see? A study published in Nature Medicine on October 29 incorporated this question into model design. EVLF-FM, developed by the research team, combines disease recognition, visual question answering and lesion localization, aiming to let users compare textual judgments with specific regions in images.

According to the formally published study, model development and testing used more than 1.3 million samples in total across 10 imaging modalities, with an additional 13,443 samples used for external validation. Average diagnostic accuracy in internal validation was 89.5%. This figure describes average performance under the study conditions; it cannot be directly interpreted as indicating that the model has the same recognition capabilities for all diseases and images from all hospitals.

Another core capability of EVLF-FM is linking answers to image locations. In addition to answering image-related questions, the model can mark relevant regions at the pixel level, allowing users to check whether its judgments correspond to visible lesions. This presentation provides clues for verification; however, seemingly reasonable markings do not establish that a diagnosis is correct, and a fluent explanation alone cannot confirm that the reasoning is reliable.

The research team's official code repository further explains that training begins with supervised fine-tuning using medical images and text data, followed by a reinforcement learning method called GRPO, incorporating rewards for both task performance and clinical reasoning consistency. The design aims to have the model address both answers and explanations, enabling diagnosis, question answering and localization to operate within a single system.

Interpreting the findings also requires distinguishing between versions of the research. The arXiv preprint released by the authors in September 2025 reported 23 development datasets, 11 imaging modalities and 8,884 external test samples, with an average internal diagnostic accuracy of 85.8%. These early figures differ from those in the formally published paper and cannot be cited interchangeably. Nor can the difference between the two accuracy figures alone establish how much the model improved under identical conditions.

The team has made training and evaluation code, along with downloadable model weights, publicly available, providing a starting point for other researchers to examine the method. However, the official code repository specifically notes that the released model version was trained on public datasets. Whether the downloaded model can reproduce the various results in the formally published paper therefore still requires checking the data, model version and evaluation conditions used.

For clinical practice, the next step is to determine whether these capabilities can improve actual care. The research still requires validation in more diverse healthcare settings, along with prospective clinical studies to assess safety, effectiveness and practical utility. Its coverage of multiple imaging modalities and provision of verifiable visual evidence make EVLF-FM worthy of further research; whether it can become a tool for routine clinical care will depend on evidence from real-world workflows.

References

  1. Nature Medicine
  2. EVLF-FM research team / GitHub
  3. arXiv