Segmentation of Glioblastoma Multiforme Via - Attention Neural Network

Abstract

Segmenting brain tumors automatically is a challenging task in medical image processing. Glioblastoma Multiforme (GBM) is difficult to identify precisely and quickly because it lacks a well-defined mass with distinct borders and is occasionally star-shaped. Developing a computational model capable of disease detection, treatment planning, and monitoring would be extremely advantageous to clinicians. U-Net is a frequently used deep learning architecture for medical image segmentation but has limitations in extracting some of the more complicated characteristics. The U-Net is a convolutional neural network (CNN) architecture that is used for image segmentation purposes. In this research, a novel CNN based U-Net architecture with a Self-Attention Module is proposed for highlighting the spatial important features from high level features. Standard metrics such as Dice Score, Jaccard Index, Hausdorff Distance, Hausdorff-95 Distance, Precision, Recall, Sensitivity, and Specificity are used to assess the performance of our proposed model. Except for Hausdorff Distance and Hausdorff-95 Distance, larger values indicate greater performance and lower values indicate worse performance for all of these metrics. A study of the means train and test results for all measures utilized in this paper on the Brats- 2019 dataset, indicates that WT segmentation outperforms TC and ET for GBM segmentation. Our technique is tested on the BRATS 2019 challenge's public benchmark for the task of segmenting malignant brain tumors.