A medical foundation model can now be judged on a more demanding question than whether it correctly names a condition in an image: can it indicate the part of the image supporting that decision, answer medical questions and explain its findings in terms clinicians can inspect? A paper published in Nature Medicine on October 29 presents a model built to handle those tasks across several medical specialties. Its strongest results are on benchmark data, not in a trial of patient outcomes.[1]
The Explainable Vision Language Foundation Model for Medicine, or EVLF-FM, was developed and evaluated with more than 1.3 million samples drawn from 18 global datasets, spanning ten imaging modalities and six specialties. That scope matters because a model optimised only for one sort of scan can look highly effective until it is faced with a different camera, anatomy or clinical setting. The researchers also report external testing on 13,443 independent samples from 17 additional datasets.[1]
Beyond the one-number diagnostic score
The six specialties represented are dermatology, hepatology, ophthalmology, pathology, pulmonology and radiology. The imaging inputs include X-rays, CT, MRI, ultrasound, retinal images, pathology slides and endoscopy. The system combines a disease-level image representation with more detailed spatial features, allowing it to classify a condition while also attempting to mark the region relevant to the answer.[1]
Across the internal diagnostic evaluations, the researchers report a mean accuracy of 0.895, the strongest average among their selected generalist and specialist comparisons. The number should not be read as an 89.5% clinical diagnostic success rate for all patients or diseases. It summarises a collection of controlled benchmark tasks with different prevalence, ground truth, scanners and decision thresholds.[1]
For visual grounding—the task of locating evidence in an image—the model reached 0.842 accuracy at an intersection-over-union threshold of 0.5 across nine imaging modalities, against 0.755 for the InternVL comparator. That threshold tests whether a predicted region overlaps its reference annotation sufficiently. It does not test whether every sentence of a generated rationale faithfully represents the model's decision process.[1]
The external tests are the more useful reality check
What the model was actually tested on
| Assessment | Reported figure | Qualification |
|---|---|---|
| Development and internal evaluation samples | Over 1.3 million | 18 global datasets; ten imaging modalitiesSource 1 |
| Independent external test samples | 13,443 | 17 additional datasetsSource 1 |
| Internal diagnostic average accuracy | 0.895 | Average across selected benchmark tasksSource 1 |
| Visual grounding [email protected] | 0.842 versus 0.755 | Nine modalities; one generalist comparatorSource 1 |
| Physician assessment of explanations | 4.25 out of 5 | Twelve doctors rated sampled explanationsSource 1 |
The investigators also tested datasets outside the development collections, both without task examples and with a small number supplied in context. Retinal OCT diagnostic accuracy rose from 0.503 without examples to 0.797 with 64 examples. Improvements were not universal: the paper reports the highest DRCR-CFP accuracy in the zero-shot setting. More examples can help a particular task without guaranteeing a gain across every dataset.[1]
The study also asked twelve physicians—four junior doctors and eight specialists—to evaluate generated explanatory outputs using a framework covering safety, context, objectivity, reproducibility and explainability. The overall mean score was 4.25 out of five. That is a human appraisal of sampled outputs, not prospective validation that explanations prevent medical mistakes or improve patient care. The paper itself distinguishes readable rationales from reasoning faithfulness.[1]
Why the difference matters
Diagnostic classification, visual localisation, answer generation and clinical decision support are separate capabilities. In this study, they were evaluated in complementary ways, which is more informative than a single leaderboard win. Yet the work does not report random assignment of hospitals or patients to the model, clinician workflow times, treatment decisions or patient morbidity. Those remain important steps before any deployment case can be made.[1]
The research offers a plausible route towards models that help clinicians inspect image evidence rather than merely receive a prediction. Its real contribution is a broad and explicitly testable architecture; the next evidence standard is reliable prospective performance with human oversight in the environments where care is delivered.[1]
Sources & further reading
- A clinical multimodal vision language foundation model with fine-grain explainability Nature Medicine · 29 Oct 2026 · Accessed 30 Oct 2026