Multiregional Segmentation of High-Grade Glioma Using Modified Deep UNET Model with Edge-Detected Multimodal MRI Images

Abstract

With newly emerging technologies in the field of computer science, there is rising awareness about its applications in the medical sciences. One of such important applications is early and accurate diagnosis of cancer tumor. Brain tumor is a deadly disease and needs to be diagnosed accurately on time. Among many types of brain tumors, high-grade glioma (HGG) is the most belligerent type, and the aggressive nature of tumors affects the survival outcomes of patients. Image processing and deep learning techniques have helped a lot in this endeavor. The proposed work deals with a hybrid methodology that combines simple edge detection technique with deep convolutional neural network to achieve state-of-the-art results. The Brain Tumor Segmentation (BRATS) 2018 dataset is used, which is provided under BRATS contest organized by MICCAI international conference (Medical Image Computing and Computer-Assisted Intervention). The Sobel operator is used for edge detection, and such edge-detected images are further trained using the modified deep UNET (md-UNET)-based model to segment glioma tumor regions into three classes namely—enhancing tumor, non-enhancing necrotic tumor, and edema. The data of 210 patients is used for training the proposed model using train-validate-test split. The model has achieved an accuracy of 98.94%, dice score of 99.03%, and dice loss of 0.0096 for separate test cohort of 6720 multi-modal MRI images of 40 glioma patients.