EXPERIMENTAL EVALUATION OF MULTI-MODAL U-NET FOR BRAIN TUMOR SEGMENTATION ON BRATS 2021 MRI DATASET

Abstract

Brain tumour segmentation from multi-modal MRI data remains a critical issue for computer-assisted diagnosis and therapy planning. Here, using the BraTS 2021 dataset, we suggest a potent multi-modal 2D U-Net architecture for binary tumour segmentation. In order to address the serious class imbalance problem, our model first fuses the T1, T1ce, T2, and FLAIR modalities at the input level. Next, it adopts a hybrid loss technique that combines WBCE and Dice Loss. It is shown that the model converges quantitatively with a testing accuracy of 96.54%. More importantly, the suggested training approach obtains a Dice coefficient of 0.74 and an IoU score of 0.60, significantly enhancing tumor-region localization vs. standard BCE-based training. Results from the study suggest that loss-function design is of paramount importance for medical segmentation performance and that even classical U-Net architectures come close to state-of-the-art when trained by overlap-aware objectives. Our proposed framework is a computationally low-cost and clinically relevant method for segmenting brain tumours.