Learning from routine health system data builds better neuroimaging AI models
Nature Medicine, Published online: 31 July 2026; doi:10.1038/s41591-026-04567-4 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…
Researchers have developed a new approach to train artificial intelligence models for neuroimaging using routine health system data. By training a three-dimensional visual foundation model on 5.24 million routine clinical computed tomography (CT) and magnetic resonance imaging (MRI) image series, the model learns a shared representation of neuroanatomy and disease.
This method outperformed foundation models trained on public Internet and medical data, enabling preliminary report generation and real-time triage in health systems.
The study, titled "Learning neuroimaging models from health system-scale data," was published in Nature Biomedical Engineering (2026). The researchers, led by A.K. and T.H., utilized ChatGPT to assist in preparing their contribution to the Research Briefing. The work builds upon previous research in the field, including the development of contrastive image-text pretraining (CLIP) for vision-language models by Radford et al.
(2021), and the demonstration of large-scale health system data supporting clinical prediction models by Jiang et al. (2023).
Written by urgent.news from Nature Medicine's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
