BrainDINO

BrainDINO is a self-supervised foundation model for brain MRI that learns transferable visual representations without task labels, enabling strong performance across a wide range of neuroimaging problems. Developed by researchers at Emory University, the Georgia Institute of Technology, and Memorial Sloan Kettering Cancer Center, it was introduced in an April 2026 preprint as a general-purpose encoder that can be adapted to clinical tasks with lightweight task heads rather than full network fine-tuning.

The model addresses a persistent bottleneck in medical imaging: high-quality labeled brain MRI is scarce and expensive, while unlabeled scans are abundant. By pretraining on roughly 6.6 million unlabeled axial slices drawn from more than 51,000 3D volumes across 20 heterogeneous datasets, BrainDINO captures anatomy and pathology-relevant features that generalize across populations, scanners, and acquisition protocols. This positions it alongside other recent brain MRI foundation efforts (e.g., BrainIAC and BrainMVP) while leaning on the DINO self-distillation recipe that has proven effective for natural-image representation learning.

BrainDINO is notable for its slice-based 2D design, which avoids the computational cost of volumetric pretraining while still delivering competitive—and often superior—results to both natural-image and MRI-specific baselines, particularly in label-scarce regimes.

#Key Features

• DINOv3-style self-distillation: Uses a teacher–student self-supervised objective on unlabeled slices, removing the need for manual annotations during pretraining.
• Large, diverse pretraining corpus: ~6.6M axial slices from 20 datasets spanning healthy controls, neurodevelopmental and neurodegenerative cohorts, and brain tumor collections.
• Frozen-encoder transfer: Strong performance with the encoder frozen and only lightweight task heads trained, lowering the data and compute burden for adaptation.
• Label efficiency: At 20% labeled data, BrainDINO frequently matches or exceeds baselines trained on 100% of labels, a key advantage for clinical settings.
• Broad task coverage: A single pretrained model supports segmentation, classification, regression, and survival modeling across neuroimaging applications.

#Technical Details

BrainDINO uses a Vision Transformer backbone (ViT-B/16) trained with a DINOv3-style teacher–student self-distillation objective. Slices are processed at 256×256 resolution, with a secondary stage using 1024×1024 upsampled images. The pretraining corpus integrates 20 datasets—including ICBM, ABIDE, PPMI, LONG579, TCGA-GBM, CPTAC-GBM, IvyGAP, REMBRANDT, BraTS collections, and fastMRI-Brain—comprising roughly 6.6 million axial slices from over 51,000 volumes. Evaluated across seven downstream task families, it reported a brain-age mean absolute error of 5.54 years (versus 7.50 for BrainIAC), 0.954 macro-AUC for Alzheimer's classification on ADNI (versus 0.872 for BrainMVP), and 0.927 Dice on BraTS2021 whole-tumor segmentation (versus 0.894 for a DINOv3 baseline). Additional tasks included post-stroke temporal prediction (ATLAS), molecular status prediction (UCSF-PDGM), MRI sequence classification, and Cox-style survival modeling (UPENN-GBM).

#Applications

BrainDINO is intended for researchers and clinical scientists working with brain MRI who need adaptable representations across heterogeneous tasks. Because the frozen encoder transfers well with minimal labeled data, it is well suited to settings where annotation is costly—such as rare-disease cohorts, multi-site studies, and early-stage biomarker discovery. Demonstrated use cases include tumor segmentation, neurodegenerative and neurodevelopmental condition classification, brain age estimation, stroke outcome prediction, glioma molecular-status prediction, MRI sequence identification, and patient survival modeling.

#Impact

BrainDINO contributes to a growing body of work showing that self-supervised foundation models can serve as general-purpose backbones for medical imaging, reducing reliance on large labeled datasets. Its consistent gains over both natural-image and MRI-specific baselines—most pronounced under limited supervision—underscore the value of domain-specific pretraining at scale for neuroimaging. As a notable limitation, the pretrained weights and full self-supervised pretraining pipeline were not publicly released at the time of the preprint (access available on request), which constrains immediate reproducibility and adoption despite the MIT-licensed evaluation code.