Post

New research · Radiology
Nature medicine · 23h
StudyNature medicine · 2026

Health system learning enables generalist neuroimaging models.

Akhil Kondepudi, Akshay Rao, Chenhui Zhao … Todd Hollon
Read paper
RadiologyStudy

A new artificial intelligence model for neuroimaging was trained on over 5 million clinical scans.

Health system learning enables generalist neuroimaging models.

Akhil Kondepudi et al. · Nature medicine · 2026
Background

Frontier artificial intelligence (AI) models have advanced rapidly through training on internet-scale public data, yet such systems lack access to private clinical data.

5.24
Results
million
the number of brain scans used to train the new artificial intelligence model
n = 5.24 million clinical MRI and CT
More results

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.

More results

NeuroVFM reduces hallucinated findings and critical errors, offering safer clinical decision support.

“
Conclusion

These results establish health system learning as a paradigm for building generalist medical AI and provide a scalable framework for clinical foundation models.

Read paper
0 comments

No comments yet. Be the first.

Related papers

LatestFoundational
Review
10-60 mGy
the range of radiation exposure to the baby from belly or pelvis CT scans
Cross-sectional
56.5%
More than half of computed tomography reports did not mention tumor deposits.
Cohort Study
6.7%
Among hereditary hemorrhagic telangiectasia patients re-imaged, 6.7% had clinically significant liver arteriovenous malformations.
Cohort Study
kappa = 0.819-0.838
adjudicated CT reports showed excellent agreement with the final diagnosis
AI / Informatics
0.87 million
the AI model learned from nearly a million image-report pairs
Guideline
“In those instances where peer reviewed literature is lacking or equivocal, experts may be the primary evidentiary source available to formulate a recommendation.”
AI / Informatics
106 anatomical structures
whole-body reference charts were created for this many different body parts
AI / Informatics
OR = 6.15
Large language models were much more likely to answer easier questions correctly than harder ones