Optimizing contrastive learning for cortical folding pattern detection

TL;DR

This work tackles the problem of detecting cortical folding patterns, which exhibit substantial inter-subject variability, by leveraging self-supervised contrastive learning on cortical skeletons derived from MRI. The authors train a SimCLR framework on large HCP and UKBiobank datasets and evaluate via a linear classifier on a manually labeled ACC dataset for the double-parallel cingulate pattern, using topology-based augmentations tailored to cortical graphs. The best configuration—a ConvNet with a 10-dimensional latent space and branch-clipping augmentation—achieves an AUC around $0.73 \pm 0.03$ on ACC-1, with performance improving to $0.76$ when trained on the larger UKBiobank dataset and showing reduced variability with more data. The study demonstrates the feasibility and robustness of self-supervised representations for folding pattern detection and suggests extending the approach to other brain regions and multimodal fusion to discover additional biomarkers.

Abstract

The human cerebral cortex has many bumps and grooves called gyri and sulci. Even though there is a high inter-individual consistency for the main cortical folds, this is not the case when we examine the exact shapes and details of the folding patterns. Because of this complexity, characterizing the cortical folding variability and relating them to subjects' behavioral characteristics or pathologies is still an open scientific problem. Classical approaches include labeling a few specific patterns, either manually or semi-automatically, based on geometric distances, but the recent availability of MRI image datasets of tens of thousands of subjects makes modern deep-learning techniques particularly attractive. Here, we build a self-supervised deep-learning model to detect folding patterns in the cingulate region. We train a contrastive self-supervised model (SimCLR) on both Human Connectome Project (1101 subjects) and UKBioBank (21070 subjects) datasets with topological-based augmentations on the cortical skeletons, which are topological objects that capture the shape of the folds. We explore several backbone architectures (convolutional network, DenseNet, and PointNet) for the SimCLR. For evaluation and testing, we perform a linear classification task on a database manually labeled for the presence of the "double-parallel" folding pattern in the cingulate region, which is related to schizophrenia characteristics. The best model, giving a test AUC of 0.76, is a convolutional network with 6 layers, a 10-dimensional latent space, a linear projection head, and using the branch-clipping augmentation. This is the first time that a self-supervised deep learning model has been applied to cortical skeletons on such a large dataset and quantitatively evaluated. We can now envisage the next step: applying it to other brain regions to detect other biomarkers.

Figures (6)

• Figure 1: Examples of cortical patterns.A. The power button sign pattern in the central region is linked to epilepsy mellerio_power_2015. B. The anterior cingulate cortex (ACC) has two known patterns, a double-parallel pattern and a single-fold pattern cachia_shape_2014. We use the detection of these two ACC patterns to evaluate our models.
• Figure 2: Dataset splits used for parameter tuning and evaluation.
• Figure 3: Preprocessing the inputs of the deep learning algorithm.Left. The cortical skeleton (in blue) represents the cortical folds for the right hemisphere. Middle. The cingulate crop with branch labels (one color per branch label): each branch represents a simple surface, a junction line, or a bottom line of a fold. Right. From the cingulate crop, we construct random views, using either the cutout augmentation, in which only bottom branches (represented in red) are kept inside the cutout (top right), or the branch-clipping augmentation, in which branches are randomly removed, and all bottom branches are removed (bottom right) .
• Figure 4: Preprocessing and training pipeline. We train the SimCLR model either on the HCP or on the UKBioBank dataset (top) and evaluate the model on the ACC dataset, which has been manually labeled for the double-parallel pattern in the cingulate region (bottom).
• Figure 5: Parameter optimization results. Training is done on HCP-1 (half of the HCP dataset). Evaluation is done on ACC-1 (half of the ACC dataset). Each small marker represents a different trained model. A. AUC score as a function of the latent space size for the ConvNet backbone. Black lines with squares and dashed lines with down triangles stand respectively for the branch-clipping and the cutout augmentations. The thinner solid black line corresponds to a Principal Component Analysis (PCA) model. B. and C. are similar plots, respectively, for the DenseNet and the PointNet backbones.