Imbalance-Aware Self-supervised Learning for 3D Radiomic Representations

Radiomics can quantify the properties of regions interest in medical image data. Classically, they account for pre-defined statistics shape, texture, and other low-level features. Alternatively, deep learning-based representations are derived from supervised learning but require expensive annotations often suffer overfitting data imbalance issues. In this work, we address challenge representation a 3D an effective quantification under imbalance. We propose self-supervised framework to learn high-level features volumes as complement existing radiomics Specifically, demonstrate how fashion using Siamese network. More importantly, deal with by exploiting two unsupervised strategies: a) sample re-weighting, b) balancing composition training batches. When combining learned feature traditional radiomics, show significant improvement brain tumor classification lung cancer staging tasks covering MRI CT imaging modalities. Codes available https://github.com/hongweilibran/imbalanced-SSL.