Fig. 1: Overview of the study.

BrainIAC is a general-purpose foundation model for brain MRI analysis, trained using a contrastive SSL approach and validated on seven diverse downstream applications (magnetic resonance sequence classification, time-to-stroke prediction, brain age estimation, mild cognitive impairment (MCI) classification, overall survival prediction for brain tumors, classification of IDH mutational status and brain tumor segmentation). BrainIAC outperforms supervised training (Scratch) and finetuning from publicly available models suitable for brain MRI (MedicalNet, BrainSegFounder). BrainIAC serves as a vision encoder for 3D Brain MRI scans generating robust latent feature representations that can be adapted easily to downstream applications. a, Datasets used in the study: a pool of 34 datasets ranging across ten neurological conditions and four sequences totaling 48,965 brain MRI scans was curated and preprocessed. GBM, glioblastoma; HGG, high-grade glioma; PLGG, pediatric low-grade glioma. b, BrainIAC was trained using contrastive learning-based SSL approach SimCLR on 32,015 Brain MRI scans. Full 3D brain magnetic resonance volumes were first decomposed into several randomly cropped, intensity-augmented patches. Differently augmented views of the same anatomical patch form a positive pair, whereas all other patches act as negative examples. Optimizing a SimCLR contrastive loss attracts positive pairs in latent space while repelling negatives. BrainIAC was further evaluated in downstream settings of classification, regression and segmentation. c, BrainIAC outperforms other approaches (Scratch, MedicalNet, BrainSegFounder) for downstream application at highly limited data availability in few-shot settings with one sample per class (K = 1), five samples per class (K = 5) and linear probing. Left: five tasks (excluding brain age prediction and time-to-stroke prediction); right: all seven tasks.

Source data