Abstract:
BRaTS’20 dataset aims for better understanding and developing an AI-based approach with novelty for multimodal segmentation of brain tumor using MRI images that are already in use since 2015 for better and accurate diagnosis of brain tumor. Pre-operative multimodal MRI scans of glioblastoma (GBM/HGG) and lower grade glioma (LGG), with pathologically confirmed diagnosis are available for each year where AI students are welcomed for challenges to develop novel models. These datasets contain training, validation and testing data for respective year’s BraTS challenge. Our study involve automated segmentation using SegResNet model for 3T multimodal MRI scans of recently provided BraTS dataset 2020. Our model has been designed based on the encoder-decoder structure and is able to achieve a 0.90 mean dice score on training set and 0.87 on the validation set. Experimental results on the testing set demonstrate no over or under fitting and is able to achieve average dice scores of 0.9000, 0.8911 and 0.8426 for the tumor core, whole tumor and enhancing tumor respectively. The proposed BraTS model underwent through some specific modifications that created novelty comparing datasets and models of previous benchmarks.Our approach has surpassed the previous models of BraTS’20 dataset in many ways giving highest dice scores for tumor core and enhancing tumor while second highest for whole tumor.
