Learning to Read the Brain from 5.24 Million 3D Image Sets: NeuroVFM Moves into Clinical Triage Testing

The bottleneck for medical imaging AI is often not that models have seen too little, but that they have not encountered the routine examinations found in real hospitals—with uneven quality, varying scanning protocols, and occasional rare lesions. A study published in *Nature Medicine* introduced NeuroVFM, an effort to build a general-purpose vision model directly from brain and head-and-neck images accumulated by a large healthcare system, supporting multiple diagnostic tasks, report generation, and urgency-based triage.

The research team compiled 566,915 examinations collected over more than 20 years at Michigan Medicine, covering CT and MRI scans of the brain, head and neck, face, and orbits and yielding a total of 5.24 million three-dimensional image sets. NeuroVFM uses a self-supervised method called Vol-JEPA: part of a volumetric scan is masked, and the model is asked to predict the missing region in an abstract feature space, allowing it to learn anatomical locations and pathological changes. The pretraining itself requires neither diagnostic labels nor radiology reports.

Using a temporally independent, held-out clinical dataset, the researchers attached classification modules covering 82 CT and 74 MRI diagnoses. Across the 156 tasks, the overall mean AUROC was 92.68% for CT and 92.49% for MRI, generally outperforming benchmark models trained on the same in-house data using either report supervision or image reconstruction. This comparison suggests that the results stem not only from data volume but also from how the model learns from three-dimensional images.

The team further combined NeuroVFM’s image encoder with the Qwen3-14B language model to generate structured radiology findings, which were then passed to a reasoning model to classify each examination as normal, routine, or urgent. In a test of 300 expert-verified CT and MRI cases, the critical-finding error rate of NeuroVFM-generated content was approximately 10%, compared with 20% for the GPT-5 comparison group; in blinded testing, experts selected the former’s reports more than twice as often as the latter’s. These measurements reflect performance with specific imaging inputs and workflows and should not be interpreted as a comprehensive comparison of the capabilities of all general-purpose models.

A test more closely resembling real-world practice was a one-week prospective “silent” feasibility study conducted in January 2026: the model processed routine imaging from 1,155 patients, but its results were not used for actual medical decisions. The NeuroVFM workflow achieved a balanced triage accuracy of 92.6%, higher than the GPT-5 workflow’s 71.2%. Among 155 patients with at least one critical finding, sensitivity was 86.5%.

That figure also defines the safety boundary. NeuroVFM missed 21 patients with critical findings, primarily because imaging lesions were omitted from the generated reports rather than because the subsequent triage reasoning was inaccurate. For conditions that cannot tolerate delay, such as acute hemorrhage and mass effect, omissions of this kind are sufficient to rule out using the model as an autonomous screening tool at this stage.

In addition, the training data came mainly from a single US academic healthcare system, and the model may inherit biases related to its local patient population, equipment, and clinical workflows. After being connected to a language model, it may also still generate findings that do not exist. The models and code released by the research team are labeled for research use only and are not medical devices. The next steps involve not only validating accuracy at more hospitals but also addressing data governance, continuous monitoring, accountability, and regulatory review. Until these issues are clarified, clinician review remains indispensable.