Health system learning enables generalist neuroimaging models
10 affiliations
- 1Machine Learning in Neurosurgery Lab, University of Michigan, Ann Arbor, MI, USA
- 2University of Michigan Computational Medicine and Bioinformatics, Ann Arbor, MI, USA
- 3University of Michigan Computer Science and Engineering, Ann Arbor, MI, USA
- 4University of Michigan Neurosurgery, Ann Arbor, MI, USA
- 5University of Cologne Neurosurgery, Cologne, Germany
- 6University of Michigan Radiology, Ann Arbor, MI, USA
- 7Machine Learning in Neurosurgery Lab, University of Michigan, Ann Arbor, MI, USA. [email protected]
- 8University of Michigan Computational Medicine and Bioinformatics, Ann Arbor, MI, USA. [email protected]
- 9University of Michigan Computer Science and Engineering, Ann Arbor, MI, USA. [email protected]
- 10University of Michigan Neurosurgery, Ann Arbor, MI, USA. [email protected]
Abstract
Frontier artificial intelligence (AI) models have advanced rapidly through training on internet-scale public data, yet such systems lack access to private clinical data. Neuroimaging is underrepresented in the public domain due to identifiable facial features within magnetic resonance imaging (MRI) and computed tomography (CT) scans, restricting model performance in clinical medicine. Here we show that frontier models underperform on neuroimaging tasks and that learning directly from uncurated data generated during routine clinical care at health systems, a paradigm we call 'health system learning', yields high-performance, generalist neuroimaging models. We introduce NeuroVFM, a visual foundation model trained on 5.24 million clinical MRI and CT volumes using a scalable volumetric predictive architecture. NeuroVFM learns comprehensive representations of brain anatomy and pathology, achieving state-of-the-art performance across multiple clinical tasks, including radiologic diagnosis and report generation. The model embeds MRI and CT scans into a shared neuroanatomic latent space and grounds diagnostic findings. When paired with open-source language models, NeuroVFM generates radiology reports that surpass frontier models in accuracy, clinical triage and expert preference. NeuroVFM reduces hallucinated findings and critical errors, offering safer clinical decision support. These results establish health system learning as a paradigm for building generalist medical AI and provide a scalable framework for clinical foundation models.
Related
- Neuroimaging Inequities in Africa: Health system readiness, diagnostic capacity, and an embedded case study of portable ultra-low-field MRI implementation in Ethiopia
- Endocannabinoid System Genes in Dementia: A Systematic Mendelian Randomization Study of Functional Variants and Tissue-Specific Expression Across Clinical, Molecular, and Neuroimaging Phenotypes
- Learning to integrate? Multimorbidity, chronic care and health system adaptation in post-USAID Zimbabwe
- Multimodal neuroimaging signature of sleep problems captures preadolescent mental health trajectories
- From brain networks to the patient-environment system in clinical neuroimaging
BibTeX
@article{kondepudi2026health,
title = {{Health system learning enables generalist neuroimaging models}},
author = {Akhil Kondepudi and Akshay T Rao and Chenhui Zhao and Yiwei Lyu and Edward Samir Harake and Soumyanil Banerjee and Jacob D. Ogle and Rushikesh Sanjeev Joshi and Anna-Katharina Meißner and Xinhai Hou and Cheng Jiang and Asadur Chowdury and Ashok Srinivasan and Brian D. Athey and Vikas Gulani and Aditya Pandey and Honglak Lee and Todd Charles Hollon},
journal = {Nature Medicine},
year = {2026},
month = jul,
doi = {10.1038/s41591-026-04497-1},
url = {https://doi.org/10.1038/s41591-026-04497-1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Health system learning enables generalist neuroimaging models
AU - Akhil Kondepudi
AU - Akshay T Rao
AU - Chenhui Zhao
AU - Yiwei Lyu
AU - Edward Samir Harake
AU - Soumyanil Banerjee
AU - Jacob D. Ogle
AU - Rushikesh Sanjeev Joshi
AU - Anna-Katharina Meißner
AU - Xinhai Hou
AU - Cheng Jiang
AU - Asadur Chowdury
AU - Ashok Srinivasan
AU - Brian D. Athey
AU - Vikas Gulani
AU - Aditya Pandey
AU - Honglak Lee
AU - Todd Charles Hollon
PY - 2026
DA - 2026/07/10
JO - Nature Medicine
DO - 10.1038/s41591-026-04497-1
UR - https://doi.org/10.1038/s41591-026-04497-1
AB - Frontier artificial intelligence (AI) models have advanced rapidly through training on internet-scale public data, yet such systems lack access to private clinical data. Neuroimaging is underrepresented in the public domain due to identifiable facial features within magnetic resonance imaging (MRI) and computed tomography (CT) scans, restricting model performance in clinical medicine. Here we show that frontier models underperform on neuroimaging tasks and that learning directly from uncurated data generated during routine clinical care at health systems, a paradigm we call 'health system learning', yields high-performance, generalist neuroimaging models. We introduce NeuroVFM, a visual foundation model trained on 5.24 million clinical MRI and CT volumes using a scalable volumetric predictive architecture. NeuroVFM learns comprehensive representations of brain anatomy and pathology, achieving state-of-the-art performance across multiple clinical tasks, including radiologic diagnosis and report generation. The model embeds MRI and CT scans into a shared neuroanatomic latent space and grounds diagnostic findings. When paired with open-source language models, NeuroVFM generates radiology reports that surpass frontier models in accuracy, clinical triage and expert preference. NeuroVFM reduces hallucinated findings and critical errors, offering safer clinical decision support. These results establish health system learning as a paradigm for building generalist medical AI and provide a scalable framework for clinical foundation models.
SN - 1078-8956
SN - 1546-170X
SN - 1744-7933
ER -