A novel compound-based loss function for glioma segmentation with deep learning

Abstract

The glioma segmentation from the Magnetic Resonance Imaging (MRI) is known to be a tedious task because of the variability in the tumor’s morphology, extent and localization. The commonly used deep learning loss functions need advancement to segment the extremely small and multiple objective areas present in a single MRI. Dice loss is well known for segmenting imbalance classes, but it removes background details that represent a lot of details. Also, in binary cross-entropy (BCE), the segmentation classes are equally valued, eliminating the features of smaller regions. So, this work proposes a new compound-based loss function that incorporates background details and enhances the segmentation accuracies by predicting even small and multiple tumor regions. This loss function adds a negative logarithm of the dice coefficient, which can be recognized as implicit regularization and optimizes the training process. In addition, a dual-modal system with the proposed loss function is utilized to highlight the modality correlation with Fluid-attenuated inversion recovery (FLAIR) and T2. It achieves the highest values of evaluation metrics on the test set, indicating better generalization capabilities of the proposed loss function. The accuracy of the proposed segmentation approach has been validated on Multimodal brain tumor segmentation (BraTS) challenge 2018 and BraTS 2019 datasets. The experiments showed that the proposed approach outperforms the other state-of-the-art segmentation algorithms while achieving mean dice coefficient and mean Hausdorff distance as 0.960 and 2.30 respectively for BraTS 2018 whereas 0.962 and 2.29 for BraTS 2019.