Exploring Compound Loss Functions for Brain Tumor Segmentation

In this study, we introduce a modified 3D U-Net framework tailored for the BraTS 2023 Segmentation - Adult Glioma challenge. Alongside conventional techniques such as data augmentation, post-processing, and Monte Carlo dropout, we investigate the efficacy of compound loss functions with a primary focus on mitigating class imbalance. In particular, we investigate various combinations of cross-entropy, boundary, and dice loss functions to identify the most suitable loss for the given data distribution. By engineering the baseline U-Net model with these modifications, we have determined that the combination of dice and cross-entropy loss yields encouraging results, exemplified by lesion-wise dice scores of 0.753, 0.791, and 0.886. Our analysis justifies the use of specially designed loss functions for the underlying data distribution at hand.