BrainDINO: Self-Supervised Foundation Model for Brain MRI Generalizes Across Clinical Tasks
BrainDINO is a self-distilled foundation model that has been trained on around 6.6 million unlabeled axial slices sourced from 20 different datasets. This model illustrates that a singular self-supervised representation can effectively generalize across various brain MRI endpoints. It leverages a frozen encoder paired with lightweight task heads, facilitating transfer for tasks such as tumor segmentation, classification of neurodegenerative and neurodevelopmental disorders, brain age estimation, post-stroke temporal forecasting, molecular status prediction, MRI sequence classification, and survival modeling. Notably, BrainDINO consistently matched or surpassed both natural-image and MRI-specific self-supervised benchmarks across multiple tasks and supervision frameworks.
Key facts
- BrainDINO is a self-distilled foundation model for brain MRI.
- Trained on approximately 6.6 million unlabeled axial slices.
- Data from 20 datasets with broad variation in population, disease, and acquisition setting.
- Uses frozen encoder with lightweight task heads.
- Supports transfer across tumor segmentation, neurodegenerative and neurodevelopmental conditions classification, brain age estimation, post-stroke temporal prediction, molecular status prediction, MRI sequence classification, and survival modeling.
- Consistently equaled or exceeded natural-image and MRI-specific self-supervised baselines.
- Published on arXiv with ID 2604.27277.
Entities
Institutions
- arXiv