Redundancy Reduction in Semantic Segmentation of 3D Brain Tumor MRIs

Another year of the multimodal brain tumor segmentation challenge (BraTS) 2021 provides an even larger dataset to facilitate collaboration and research methods, which are necessary for disease analysis treatment planning. A large size BraTS advent modern GPUs provide a better opportunity deep-learning based approaches learn representation from data. In this work, we maintained encoder-decoder network, but focused on modification network training process that minimizes redundancy under perturbations. Given set trained networks, further introduce confidence ensembling techniques improve performance. We evaluated method 2021, in terms dice enhanced core, core whole tumor, achieved 0.8600, 0.8868 0.9265 average validation set, 0.8769, 0.8721, 0.9266 testing set. Our team (NVAUTO) submission was top performing ET TC scores, using Brats ranking system (based Hausdorff distance per case) 2nd place 4th