A deep conventional neural network model for glioma tumor segmentation

Abstract

AbstractGlioma represents one of the most aggressive cancers, which can develop in the brain. The automatic tumor segmentation and its sub‐regions represent a challenging task owing to their considerable structural variation. It can appear in different ways and with several shapes, which makes tissue identification a crucial task. Reliable and accurate segmentation presents an important component in tumor treatment and diagnosis planning. To overcome these drawbacks, various Deep Learning (DL) schemes are proposed to aid doctors. In this paper, a novel Convolutional Neural Network (CNN) scheme for glioma segmentation was proposed. Our suggested technique consists of three phases. First, we use intensity normalization to ameliorate the image quality as a preprocessing step. Second, an automatic segmentation technique based on CNN has been proposed. The new model has several layers. Finally, and with the goal of refining the segmentation results, we employ a post‐processing approach. We use the public benchmark BraTS datasets from 2018 and 2020 with low‐grade and high‐grade glioma tumors to test the suggested framework. These datasets contain about 285 and 369 patients, respectively. The four modalities are exploited. Each patient has about 155 2D images from every modality. All the images have the same size (240 × 240 pixels). Our technique performs well compared to new methods, with Dice scores of 0.88 for the Whole Tumor, 0.84 for Tumor Core, and 0.71 for Enhancing Tumor based on the first dataset. According to the second dataset, the three regions had an average of 0.88, 0.9, and 0.75, respectively. The Jaccard indexes for the first data set are 0.8, 0.73, and 0.56 for the three regions, respectively. The second data set attains 0.8, 0.82, and 0.6 for the three regions. The results show that the suggested framework is an excellent way to segment data, especially compared to other methods.