Enhancing Slice-Wise Brain MRI Tasks using Self-Supervised and Auxiliary Learning

A recurrent challenge in medical image processing using deep learning is the need for large labeled datasets to solve clinically relevant tasks. Such data is difficult to obtain because it often requires a high level of professional specialization and significant time investment. Self-supervised learning (SSL) allows models to learn task-agnostic feature representations from unlabeled data prior to the use of task-specific labeled samples to solve downstream tasks. In this work, we focus on the tasks of craniopharyngioma recognition (CPGR) and detection of hypothalamic involvement (DHI). We compare the results obtained by a 2D convolutional network in these tasks using supervised learning, to those obtained using SSL plus fine-tuning. Three SSL methods are tested: SimCLR, Decoupled Contrastive Learning (DCL), and Variance-Invariance-Covariance Regularization (VICReg), using slices from structural brain magnetic resonance imaging (MRI). We also introduce Slice-Wise Regularization (SWR), a novel auxiliary learning task designed to prevent the model from decorrelating representations of contiguous slices when solving tasks using brain MRI datasets. A relevant aspect of the proposed auxiliary method is that it does not require any extra human-made annotations, and it leverages intrinsic structural properties of MRI. We compare the performances of different configurations of our proposed method using SSL and SWR on both downstream tasks with those obtained using supervised learning. We obtained statistically significant improvements with our method (SSL + SWR), achieving F1-scores of 80.3 $\pm$ 2.4 for CPGR and 82.8 $\pm$ 5.0 for DHI, in comparison to 74.4 $\pm$ 4.9 for CPGR and 65.4 $\pm$ 6.5 for DHI, when using supervised learning.