8 - Multimodal MRI Brain Tumor Segmentation—A ResNet-based U-Net approach

Abstract

This work emphasizes automatic segmentation of brain tumor substructures from 3-D multimodal magnetic resonance imaging (3-D-MMRI) scan images. We have used fully convolutional neural networks to extract the complete, core, and enhancing tumors. The network architecture is designed by combining U-Net and ResNet architectures for increasing the dice score between the segmented image and ground truth image. We have also used weighted cross-entropy and generalized dice loss as loss functions that eliminate data imbalance. The proposed method is trained and tested using the BraTS 2020 training dataset, which contains 305 volumes of high-grade glioma and low-grade glioma 3-D-MMRI scans. The performance of the proposed work is closer to the manually segmented images by experienced neurologists available with BraTS 2020 datasets. We have obtained a mean dice score of complete, core, and enhancing tumor as 0.81, 0.93, and 0.83, respectively.