Learning from routine health system data builds better neuroimaging AI models - Nature

We trained a three-dimensional visual foundation model directly on 5.24 million routine clinical computed tomography (CT) and magnetic resonance imaging (MRI) image series so that it learned a shared representation of neuroanatomy and disease. The model demonstrated state-of-the-art diagnosis, in contrast to foundation models that are trained on public Internet and medical data, and enabled preliminary report generation and triage in real health systems.

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

27,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 12 print issues and online access

269,00 € per year

only 22,42 € per issue

Buy this article

39,95 €

Prices may be subject to local taxes which are calculated during checkout

References

- Radford, A. et al. Learning transferable visual models from natural language supervision. In Proc. 38th International Conference on Machine Learning 8748–8763 (2021). This paper established contrastive image-text pretraining (CLIP) as a foundation for modern vision-language models.

- Jiang, L. Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357–362 (2023). This study demonstrated that large-scale health system data can support broadly useful clinical prediction models.

- Assran, M. et al. Self-supervised learning from images with a joint-embedding predictive architecture. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 15619–15629 (2023). This paper introduced I-JEPA, the self-supervised representation learning method adapted here for volumetric medical imaging.

- Lyu, Y. et al. Learning neuroimaging models from health system-scale data. Nat. Biomed. Eng. https://doi.org/10.1038/s41551-025-01608-0 (2026). This paper showed strong performance with neuroimaging learning from MRI–report pairs at health system scale.

- Moor, M. et al. Foundation models for generalist medical artificial intelligence. Nature 616, 259–265 (2023). This perspective outlines the rationale for generalist medical AI models that learn across heterogeneous clinical data modalities and support multiple downstream tasks.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This is a summary of: Kondepudi, A. et al. Health system learning enables generalist neuroimaging models. Nat. Med. https://doi.org/10.1038/s41591-026-04497-1 (2026).

A.K. and T.H. used ChatGPT to help prepare their contribution to this Research Briefing.

About this article

Cite this article

Learning from routine health system data builds better neuroimaging AI models. Nat Med (2026). https://doi.org/10.1038/s41591-026-04567-4

- Published:

- Version of record:

- DOI: https://doi.org/10.1038/s41591-026-04567-4