BrainIAC AI model predicts brain age, dementia risk and cancer survival

Harvard-affiliated researchers report BrainIAC, a self-supervised AI model trained on nearly 49,000 brain MRIs and validated on 48,965 scans, that estimates brain age, assesses dementia risk, detects tumor mutation types and predicts brain cancer survival; results are published in Nature Neuroscience.

Researchers at Harvard-affiliated Mass General Brigham have developed a foundation AI model called BrainIAC that extracts multiple disease-related signals from routine brain MRI scans. The model was pretrained using self-supervised learning on nearly 49,000 brain MRIs and was validated on 48,965 diverse scans across seven tasks. Reported applications include estimating a person's "brain age", predicting dementia risk, detecting brain tumor mutation types, and forecasting survival after brain cancer. The team reported that BrainIAC outperformed several task-specific models and performed well when training data were limited. Key findings: - BrainIAC was pretrained with self-supervised learning on nearly 49,000 brain MRI scans and validated on 48,965 diverse scans across seven tasks. - Reported capabilities include estimating brain age, assessing dementia risk, identifying brain tumor mutation types, and predicting survival from brain cancer. - The model generalized across healthy and abnormal images and succeeded on tasks ranging from scan classification to mutation detection. - BrainIAC outperformed three conventional, task-specific AI frameworks in the reported comparisons. - Performance advantages were largest when annotated training data were scarce or when task complexity was high. - Authors noted further research is needed to test the framework on additional brain imaging methods and larger datasets; the study received partial support from the National Institutes of Health/National Cancer Institute and the Botha‑Chan Low Grade Glioma Consortium. Summary: The investigators report that BrainIAC generalized across imaging types and outperformed conventional models, which the authors say could advance biomarker discovery and diagnostic research. Undetermined at this time.