Unsupervised Representation Learning of Cingulate Cortical Folding Patterns

Abstract

The human cerebral cortex is folded, making sulci and gyri over the whole cortical surface. Folding presents a very high inter-subject variability, and some neurodevelopmental disorders are correlated to local folding structures, named folding patterns. However, it is tough to characterize these patterns manually or semi-automatically using geometric distances. Here, we propose a new methodology to identify typical folding patterns. We focus on the cingulate region, known to have a clinical interest, using so-called skeletons (3D representation of folding patterns). We compare two models, β − V AE and SimCLR, in an unsupervised setting to learn a relevant representation of these patterns. We add a decoder to SimCLR to be able to analyse latent space. Specifically, we leverage the data augmentations used in SimCLR to propose a novel kind of augmentations based on folding topology. We then apply a clustering on the latent space. Cluster folding averages, interpolation in the latent space and reconstructions reveal new pattern structures. This structured representation shows that unsupervised learning can help in the discovery of still unknown patterns. We will gain further insights into folding patterns by using new priors in the unsupervised algorithms and integrating other brain data modalities. Code and experiments are available at github.com/neurospin-projects/2021 jchavas lguillon deepcingulate.

Optimizing contrastive learning for cortical folding pattern detection

“…Last, for testing, we train models with the chosen hyperparameters on HCP-2 and evaluate them on ACC-2. We find the AUC of SimCLR on the latent space to be 0.73±0.03 (n=5 models), slightly above the AUC (0.69±0.04, n=5) of the β-VAE model optimized in 8 and applied here to the same dataset (Fig. 6A).…”

“…In the second view, all non-zero voxels inside a random 3D block of the same size are kept, and all non-zero voxels outside this block are removed. 8 In all views, fold bottom pixels are conserved to keep the information on the global longitudinal shape of the sulcus.…”

“…The first is a simple six-layer convolutional network, taken from a β-VAE encoder optimized for cortical skeleton images. 8 The second one is a DenseNet architecture, 22 which gives good results when used on brain MRI data. 23 Its number of dense blocks is reduced to two, considering the smaller image size.…”

“…23 Its number of dense blocks is reduced to two, considering the smaller image size. 8 The third is a PointNet. 24 A PointNet is a neural network that, instead of taking a binary 3D matrix with a fixed size as input, takes point clouds, i.e., a list of 3D coordinates corresponding to the non-zero pixels in the original 3D matrix.…”

“…Recently, β-VAE 6 and SimCLR 7 models have been trained to build a latent space for folding representation. [8][9][10] However, these studies have not used a manually labeled dataset for hyperparameter tuning and evaluation.…”