Singapore Medical AI Shows Its Reasoning on Scans

Researchers in Singapore have built an AI model that reads ten kinds of medical image, from eye scans to chest X-rays, and shows which part of the image its answer is based on. The model, called EVLF-FM, is described in a paper published in Nature Medicine on 29 October. The authors are clear that it is a research system. It has been tested only on existing datasets and is not ready for use on patients.

The work was led by ASTAR’s Institute of Advanced Intelligence and Computing and the Singapore Eye Research Institute and Duke-NUS Medical School. The senior authors are Yong Liu of ASTAR and Daniel Shu Wei Ting of SERI and Duke-NUS. Co-authors include researchers at Stanford, Harvard Medical School, Oxford, Tsinghua University and Mount Sinai in New York.

What EVLF-FM does

EVLF-FM stands for Explainable Vision Language Foundation Model for Medicine. It is a vision language model, meaning it takes in an image and a question and answers in text. It can suggest a diagnosis, answer questions about an image and draw a box around the relevant area, a task called visual grounding. It also writes out its reasoning step by step, tied to what it has highlighted.

The team developed and tested the model on more than 1.3 million samples from 18 datasets. These cover ten imaging types across six specialties: dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology. The model has 8 billion parameters, which is small next to some rival medical models.

Training ran in two stages. The model was first fine-tuned on labelled examples. It was then refined with a form of reinforcement learning on visual tasks, a method known as GRPO. The step-by-step reasoning examples used in training were generated with other AI models, MedGemma-1.5 and GPT-5-mini. About 30% of those generated explanations were thrown out because they contained hallucinations, the paper says.

How it performed

On held-back data from the training sources, EVLF-FM had the highest average diagnostic accuracy of the models tested, at 0.895. In visual grounding, its boxes matched the correct region (at least 50% overlap) 84.2% of the time on average across nine imaging types.

The team also tested it on 13,443 samples from 17 further datasets it had never seen. Without extra training, it scored best on 8 of 12 of these external datasets. It was competitive with much larger models, including Google’s MedGemma 27B and MedDr 32B.

Twelve physicians rated its written reasoning using a framework called S.C.O.R.E. They gave it an average of 4.25 out of 5. Objectivity scored 4.36, and reasoning on eye OCT scans scored 4.73. Specialists also noted things the model missed, such as vitreomacular traction, a condition in which the gel inside the eye pulls on the retina.

These figures differ from those in the team’s 2025 preprint, which used a different mix of datasets. This article uses the peer-reviewed figures. The authors have also deposited their code in a public Zenodo archive.

Why explainable scan reading still needs a trial

Explanations are the selling point here, and the reason this work could be useful. A doctor who can see which part of a retinal scan led to a suggested diagnosis can check it in seconds. A bare “diabetic retinopathy: yes” offers nothing to check. That is why the authors picture EVLF-FM as “the backbone of human-in-the-loop decision-support tools”, with a clinician always making the call.

But the paper is unusually frank about the distance from here to the clinic. The authors write that the model’s “current accuracy does not yet reach the level typically required for direct clinical deployment”, which they put at consistently above about 90% across datasets. They describe it as a step forward, “rather than representing a deployment-ready system”. Every result comes from past, retrospective data. “Prospective clinical studies are required to evaluate the safety, effectiveness and clinical utility of EVLF-FM before routine clinical deployment,” they write.

There are other limits too. The model handles only still, two-dimensional images, not 3D scans or video. It may carry biases from its training data. And its reasoning was learned partly from AI-written explanations, almost a third of which had to be thrown out. Readers should also note the authors’ declared interests: Ting holds patents and shares in EyRIS and aISIGHT, two medical AI companies.

That pattern is familiar. Microsoft’s diagnostic AI scored 83% on past cases in a Science paper, and a dermatologist’s own test found an older image model beat two LLMs on skin cancer. Tools that do reach patients, such as the MRI-based breast cancer risk tool the FDA cleared or the AI reading X-rays in TB screening, are narrower tools that have gone through regulatory review or real-world use. EVLF-FM’s next step is a prospective trial.