MixUp Brain-Cortical Augmentations in Self-supervised Learning

Learning biological markers for a specific brain pathology is often impaired by the size of the dataset. With the advent of large open datasets in the general population, new learning strategies have emerged. In particular, deep representation learning consists of training a model via pretext tasks that can be used to solve downstream clinical problems of interest. More recently, self-supervised learning provides a rich framework for learning representations by contrasting transformed samples. These methods rely on carefully designed data manipulation to create semantically similar but syntactically different samples. In parallel, domain-specific architectures such as spherical convolutional neural networks can learn from cortical brain measures in order to reveal original biomarkers. Unfortunately, only a few surface-based augmentations exist, and none of them have been applied in a self-supervised learning setting. We perform experiments on two open source datasets: Big Healthy Brain and Healthy Brain Network. We propose new augmentations for the cortical brain: baseline augmentations adapted from classical ones for training convolutional neural networks, typically on natural images, and new augmentations called MixUp. The results suggest that surface-based self-supervised learning performs comparably to supervised baselines, but generalizes better to different tasks and datasets. In addition, the learned representations are improved by the proposed MixUp augmentations. The code is available on GitHub (https://github.com/neurospin-projects/2022 cambroise surfaugment).