University of Michigan · 2026-07-12 · major
NeuroVFM — brain-scan AI outperforms GPT-5 on clinical triage
NeuroVFM is a neuroimaging foundation model trained on 5.24M clinical MRI/CT volumes from Michigan Medicine. Using Vol-JEPA, it achieves 92.68 AUROC across 156 tasks and outperforms GPT-5 by 21.4pp on triage — at 24× lower cost.
Michigan Medicine's 5M-scan neuroimaging AI trained on uncurated hospital data outperforms GPT-5 on triage at 24× lower cost.
Quick facts
| Maker | MLiNS Lab, University of Michigan |
|---|---|
| Training data | 5.24M MRI & CT volumes from 566,915 studies, 20+ years |
| Training method | Vol-JEPA (self-supervised, no radiology labels) |
| CT performance | 92.68 AUROC across 82 diagnostic tasks |
| MRI performance | 92.49 AUROC across 74 diagnostic tasks |
| Triage vs GPT-5 | 92.6% vs 71.2% balanced accuracy (+21.4pp) |
| Cost vs GPT-5 | 24× cheaper for report generation |
| License | MIT (code); CC-BY-NC-SA 4.0 (weights, non-commercial) |
| Published | Nature Medicine, July 2026 |
Benchmarks
What is it?
NeuroVFM is a generalist visual foundation model for clinical neuroimaging, trained on 5.24 million MRI and CT volumes from the University of Michigan health system. Using a self-supervised method called Vol-JEPA, it learned brain anatomy and pathology directly from raw clinical scans — no radiology report labels, no curation required.
How does it work?
NeuroVFM uses Volumetric Joint-Embedding Predictive Architecture (Vol-JEPA), a 3D ViT-Base encoder that predicts masked volume patches from context, trained in under 1,000 GPU hours on 8 NVIDIA L40S GPUs. Diagnostic heads covering 156 tasks (74 MRI + 82 CT) and a lightweight LLaVA-based findings LLM are then fine-tuned on top for report generation and triage.
Why does it matter?
In a real-world prospective trial at Michigan Medicine (1,155 studies, January 2026), NeuroVFM achieved 92.6% triage accuracy versus GPT-5's 71.2% — a 21.4-point gap — while costing 24× less and producing 23× less carbon per report. It proves that private clinical imaging data, trained via self-supervision, can outperform frontier commercial models on specialized clinical tasks.
Who is it for?
ML researchers in medical imaging, hospital AI teams, neuroradiology labs
Frequently asked questions
- What is NeuroVFM?
- NeuroVFM is a visual foundation model for neuroimaging developed by the MLiNS Lab at University of Michigan. It is trained on 5.24 million clinical MRI and CT volumes using self-supervised Vol-JEPA and covers 156 diagnostic tasks across both modalities — without radiology report labels.
- How does NeuroVFM compare to GPT-5 on clinical tasks?
- In a prospective one-week trial at Michigan Medicine (1,155 studies, January 2026), NeuroVFM achieved 92.6% balanced triage accuracy versus 71.2% for GPT-5 — a 21.4 percentage point advantage. NeuroVFM also generates radiology reports 24× cheaper and with 23× lower carbon footprint.
- Is NeuroVFM available as open source?
- Yes. NeuroVFM code is on GitHub (MLNeurosurg/neurovfm) under an MIT license. Pretrained weights are on Hugging Face under CC-BY-NC-SA 4.0 for non-commercial research use and require institutional email approval. The model is research-only — not a medical device.
- What data was NeuroVFM trained on?
- NeuroVFM was trained on 5.24 million uncurated MRI and CT volumes from 566,915 studies at University of Michigan's Michigan Medicine, spanning over 20 years of routine clinical care — using no radiology report labels, only the raw imaging data itself.
Try it
https://github.com/MLNeurosurg/neurovfm