AnatCL

AnatCL is an anatomically informed representation-learning method for structural brain MRI. The practice uses weakly supervised contrastive learning to develop reusable representations that capture biologically meaningful variation in brain structure. AI enters after MRI preprocessing: a three-dimensional deep neural network is pretrained on structural MRI while incorporating age and anatomical descriptors, such as cortical measurements, as supervisory information. This allows the model to learn representations that are not tied to a single disease label or downstream prediction task. The pretrained representation can subsequently be transferred to different neuroimaging applications through fine-tuning or task-specific prediction. The method was evaluated across 22 downstream phenotyping tasks: 12 diagnostic classification tasks, including Alzheimer’s disease, autism spectrum disorder and schizophrenia, and 10 clinical-assessment score prediction tasks. Its principal output is a reusable pretrained representation for structural MRI that supports subsequent neuroimaging analysis rather than providing a direct clinical diagnosis.

The development of AnatCL required expertise in computer vision, deep learning and neuroimaging, together with access to large structural MRI datasets and standardised preprocessing pipelines. The model was pretrained on structural T1-weighted MRI from the OpenBHB resource and evaluated across several independent neuroimaging cohorts used for neurological and psychiatric phenotyping. A central methodological condition is the availability of anatomical information that can be derived consistently from MRI and used to guide model pretraining. This provides a biologically grounded supervisory signal without requiring complete manual diagnostic annotation for the entire pretraining dataset.

The workflow also depends on computational infrastructure suitable for three-dimensional deep-learning models. Model pretraining is comparatively resource-intensive and requires GPU-supported computation; the authors report experiments run on a cluster of four NVIDIA V100 GPUs, with a single training run taking approximately 10 hours. Downstream reuse is less demanding because pretrained weights are made available. Reproducibility and transfer are supported through the public release of source code, pretrained models and implementation instructions. The method expects MRI data to undergo defined preprocessing before model application, which contributes to consistency across datasets but also creates a dependency on compatible neuroimaging processing workflows. The preprocessing pipeline is based upon open-source tools, which ensures reproducibility.

The main demonstrated outcome is the transferability of the learned representation across heterogeneous neuroimaging tasks. AnatCL was evaluated on 22 downstream tasks, including diagnostic classification for several neurological and psychiatric conditions and prediction of clinical assessment scores. The reported results indicate that incorporating anatomical information during pretraining can improve the robustness and generalisability of representations relative to comparison approaches, with state-of-the-art performance reported on several downstream tasks. Beyond its original publication, AnatCL has been validated by the community: independent benchmarking confirmed that anatomically grounded representations achieve the best stability across scanners (Navarro-Gonzalez et al., 2026). The breadth of evaluation is important because it tests whether the representation retains value beyond a single disease-specific application. The practice is also strong in openness and reuse. Source code, pretrained models and implementation guidance are publicly available, allowing other researchers to use the learned representation without reproducing the full pretraining stage and reducing the computational burden of adoption. The evidence nevertheless remains methodological and benchmark-based. Its principal contribution is a reusable, anatomically informed representation-learning approach for structural brain-MRI research.

Publication: Barbano, C.A., Brunello, M., Dufumier, B., Grangetto, M. and Alzheimer’s Disease Neuroimaging Initiative, 2025. Anatomical foundation models for brain MRIs. Pattern Recognition Letters. https://doi.org/10.1016/j.patrec.2025.11.028

Project / pretrained models: https://github.com/EIDOSLAB/AnatCL

Training code: https://github.com/carloalbertobarbano/contrastive-learning-brain-mri