AnatCL

AnatCL is a family of anatomical foundation models for structural brain MRI, developed by EIDOSLAB at the University of Turin together with CEA NeuroSpin (Paris-Saclay) and first released as a preprint in August 2024. The work addresses a recurring problem in neuroimaging: deep models trained on small, single-site clinical cohorts tend to overfit and generalize poorly across scanners, sites, and conditions. AnatCL instead learns general-purpose representations of brain anatomy on a large pool of healthy-control scans, producing an encoder that can be reused across many downstream diagnostic and predictive tasks.

The central idea is to inject explicit anatomical knowledge into a contrastive pretraining objective. Rather than treating subjects as interchangeable or relying only on chronological age, AnatCL defines how "similar" two brains are using region-wise anatomical measurements—cortical thickness, gray matter volume, and surface area—and aligns the learned representation space accordingly. This anatomically informed weak supervision yields embeddings that capture clinically meaningful structure beyond what a purely self-supervised or age-only objective recovers.

AnatCL fits into the line of weakly supervised contrastive methods for neuroimaging (notably y-Aware and prior EIDOSLAB brain-age work) and extends them with a richer, anatomy-derived notion of positive pairs. Pretrained weights are openly released under the MIT license, making it a practical backbone for brain MRI research groups without large labeled datasets.

#Key Features

• Anatomically informed contrastive objective: Extends the y-Aware kernel framework by defining a continuous "degree of positiveness" between scans from region-wise cortical measurements rather than discrete labels or age alone.
• Multiple anatomical targets: Pretraining uses cortical thickness (CT), gray matter volume (GMV), and surface area (SA), letting the encoder capture complementary aspects of brain morphology.
• Broad downstream validation: Evaluated on 12 diagnostic tasks (including Alzheimer's disease, autism spectrum disorder, and schizophrenia) and the prediction of 10 clinical assessment scores.
• Transferable, robust representations: Pretraining on healthy controls produces features that generalize across sites and scanners, improving robustness compared to task-specific training from scratch.
• Open weights: Pretrained models are released under an MIT license and installable directly from the project repository.

#Technical Details

AnatCL uses a 3D ResNet-18 convolutional backbone applied to volumetric T1-weighted MRI; the authors found that larger backbones such as 3D ResNet-50 offered no consistent advantage and retained the smaller, more deployable model. Inputs are voxel-based morphometry volumes of size 121×128×121 preprocessed with the CAT12 toolbox. Pretraining is performed on OpenBHB, an aggregation of roughly 4,000 healthy-control T1 MRIs drawn from public cohorts including ABIDE I/II, CoRR, GSP, and IXI. The contrastive loss replaces hard positive/negative labels with kernel-weighted anatomical similarity, pulling representations together in proportion to how alike two brains are across the chosen anatomical features. The authors release both global and local (region-wise) variants of the encoder, reporting state-of-the-art or competitive performance across the diagnostic and clinical-score benchmarks. The peer-reviewed version appeared in Pattern Recognition Letters (2026).

#Applications

AnatCL serves as a pretrained backbone for structural brain MRI analysis, particularly useful for groups with limited labeled clinical data. Typical uses include fine-tuning or linear probing for neurodegenerative and neuropsychiatric disease classification (Alzheimer's, autism, schizophrenia), regression of clinical assessment scores, and brain-age estimation as a biomarker of atypical aging. Because the encoder is trained to respect anatomical similarity, its embeddings are also suitable for exploratory analyses such as cohort stratification and as feature extractors in larger neuroimaging pipelines.

#Impact

AnatCL demonstrates that encoding explicit anatomical priors into contrastive pretraining produces more robust and generalizable neuroimaging representations than age-only or purely self-supervised baselines, advancing the case for foundation models in structural MRI. By releasing open MIT-licensed weights and an installable package, it lowers the barrier for clinical neuroimaging labs to adopt transfer learning rather than training from scratch on small cohorts. Its progression from preprint to publication in Pattern Recognition Letters reflects validation by the pattern-recognition community. Limitations include reliance on CAT12-based VBM preprocessing and pretraining restricted to healthy controls, which may bound how much pathology-specific structure the frozen encoder captures.