BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning

BrainDINO is a self-supervised foundation model trained on approximately 6.6 million unlabeled brain MRI axial slices from 20 diverse datasets, enabling generalizable clinical representation learning across various tasks such as tumor segmentation and brain age estimation. The model employs a frozen encoder with lightweight task heads and demonstrates superior performance compared to existing self-supervised baselines, particularly in scenarios with limited labeled data. This approach underscores the potential for scalable, data-efficient analysis in neuroimaging without the need for extensive pretraining or fine-tuning, making it significant for practitioners in the field.