3D-Neuro-SimCLR

3D-Neuro-SimCLR is a self-supervised foundation model for volumetric brain MRI analysis, developed by researchers at McGill University and Mila - Quebec Artificial Intelligence Institute and introduced in a September 2025 preprint. Most deep-learning models for brain MRI are trained from scratch on a single task and a single disease cohort, which limits their ability to generalize across neurological conditions, scanners, and patient populations. 3D-Neuro-SimCLR addresses this by learning general-purpose representations directly from large, heterogeneous collections of unlabeled scans.

The model adapts SimCLR-style contrastive pretraining to 3D T1-weighted structural MRI. Rather than relying on disease labels, it learns to pull together augmented views of the same scan and push apart views of different scans, building an embedding space that captures broadly transferable anatomical features. This pretrained encoder can then be fine-tuned for a wide range of downstream clinical prediction tasks with relatively little labeled data.

By aggregating 11 publicly available datasets spanning Alzheimer's disease, Parkinson's disease, frontotemporal dementia, stroke, and healthy aging, the work contributes an accessible, broadly applicable foundation model for clinical brain MRI, with pretrained weights released openly for the neuroimaging community.

#Key Features

• Disease-agnostic pretraining: Contrastive self-supervision across 11 datasets and many neurological conditions yields representations that transfer across diagnostic tasks rather than being tuned to one cohort.
• Volumetric 3D modeling: The encoder operates on full 3D T1-weighted volumes (preprocessed to MNI152 ICBM space), preserving spatial context that slice-based approaches discard.
• Label-efficient fine-tuning: The model matches or exceeds baselines on Alzheimer's classification even when fine-tuned on only 20% of the labeled training samples.
• In- and out-of-distribution generalization: Evaluation spans both held-in datasets and entirely unseen cohorts, testing robustness to scanner and population shift.
• Open release: Code and pretrained foundation weights are distributed under an MIT license, lowering the barrier to reuse in academic and clinical research.

#Technical Details

The encoder is a 3D ResNet-18 convolutional network paired with a 64-dimensional projection head for contrastive learning, implemented in PyTorch with components adapted from SimCLR and MONAI. Pretraining used 44,958 scans from 18,759 patients drawn from ADNI, AOMIC, CoRR, DLBS, GSP, HABS-HD, MCSA, NIFD, PPMI, SALD, and SOOP, with scans preprocessed via TurboPrep and registered to the MNI152 ICBM template. The authors evaluated four downstream tasks: Alzheimer's disease classification (AIBL, AUC 0.929), sex classification (IXI, AUC 0.991), brain age regression (IXI, MAE 4.35 years), and stroke-scale regression (SOOP, MAE 5.37). Across all four tasks, the fine-tuned SimCLR model outperformed both a Masked Autoencoder (ViT-Tiny) self-supervised baseline and fully supervised models trained from scratch.

#Applications

The model is intended as a reusable backbone for clinical and research neuroimaging pipelines. Researchers studying neurodegenerative and cerebrovascular disease can fine-tune it for classification, regression, or biomarker prediction tasks without assembling large labeled datasets, which is especially valuable for rare conditions or small single-site studies. Its label efficiency makes it well suited to settings where expert annotation is scarce, and its broad pretraining distribution supports deployment across cohorts with differing scanners and demographics.

#Impact

3D-Neuro-SimCLR contributes to a growing effort to bring the foundation-model paradigm to 3D medical imaging, where data scale and annotation cost have historically constrained deep learning. By demonstrating that contrastive pretraining across diverse neurological diseases produces a single transferable brain MRI encoder, and by releasing the weights openly, the work provides a practical starting point for downstream neuroimaging research. As a recent preprint its long-term adoption remains to be established, and like other structural-MRI models its evaluation is bounded by the modalities and populations represented in its public training datasets.