An Automatic Segmentation Model for Brain Tumorous and its Sub-Regions in Magnetic Resonance Images

The brain tumor segmentation plays a pivotal role in the field of medical image processing. Ensuring precise segmentation is crucial for enhancing therapy alternatives and increasing patient survival rates. Determining the precise boundaries of brain tumors in several Magnetic Resonance Images (MRI) is very intricate and time-intensive. The segmentation of malignant areas in brain tumors has been a considerable obstacle in the clinical diagnosis and therapy of cancer. This study presents an automated approach for segmenting brain tumors and their related subregions, including whole tumor (WT), enhancing tumor (ET), and tumor core (TC). The work utilized DeepLabv3 for segmentation and the U-Net architecture as its backbone structure. Deformable convolution layers have substituted the convolutional layers of U-Net. To improve the segmentation of the tumor and its subregions, the whole architecture has been integrated using Atrous Spatial Pyramid Pooling (ASPP). The proposed framework has shown exceptional performance metrics on the test dataset, with a test accuracy of 99.45 % and a precision of 99.46%. The suggested model exhibits a categorical cross-entropy loss of 0.016, accompanied by a mean dice coefficient score (DSC) of 73.37% and an Intersection over Union (IoU) of 82.89%. Furthermore, the model has a sensitivity of 99.31 % and a specificity of 99.82%.