BrainIAC foundation model for human brain MRI published

Researchers have published Brain Imaging Adaptive Core, or BrainIAC, a foundation model for brain magnetic resonance imaging. The model applies self-supervised learning and pretraining to learn generalized representations from unlabeled data for adaptation to downstream tasks. BrainIAC was trained and validated on 48,965 brain MRIs across broad demographics and medical settings. Pretraining used contrastive self-supervised learning on 32,015 multiparametric MRIs from 16 datasets that spanned ten medical conditions. The model was adapted to seven downstream tasks. These tasks were MRI sequence classification, brain age prediction, detection of isocitrate dehydrogenase mutation, survival prediction for brain tumors, early dementia prediction through mild cognitive impairment versus healthy control, time-to-stroke prediction, and adult glioma segmentation. BrainIAC was compared with supervised training from scratch, the MedicalNet pretrained model, and the BrainSegFounder segmentation model. It showed stronger performance than the benchmarks, particularly in low-data and few-shot settings with one or five samples per class as well as in high-difficulty prediction tasks. The model also demonstrated resilience under simulated imaging perturbations that mimicked real-world scanner variability.