Brain Tumor Segmentation using the Modified UNET Architecture

Brain tumor is a deadly diseases where the most common and dangerous type of it is glioma, as it may cause death if its grade is high. Patient’s life can be saved if tumor detected in its early stage which thereby starts the treatment procedure sooner. The early detection can be possible by accurately segmenting the brain tumors from MRI scans as they can be easily diagnosed from them. Computer aided diagnostic software can help in automatic segmentation of brain tumors and several methodologies have been designed based on deep learning. In simple U-Net architecture, during encoding or downsampling phase local features are lost which results in constant learning as the model goes deeper. In this paper, we have used the spatial attention-based mechanism to help preserve local features. Additional modules in the encoding path of the redesigned architecture are concatenated during upsampling or decoding path. In the experiments, we have obtained segmentation results for the tumor core (TC), whole tumor (WT), and enhancing tumor (ET) regions of BraTS 2020 dataset. The average dice score were obtained as 0.9027, 0.8868, and 0.9067 for TC, WT, and ET respectively. To show the effectiveness of our methodology we have tested the trained model on 50% of the BraTS 2021 benchmark dataset. The achieved dice scores were 0.8039, 0.7640, and 0.6689 for TC, WT, and ET, respectively.