UV-Nets: Semantic Deep Learning Architectures for Brain Tumor Segmentation

Semantic segmentation in medical imaging is a delicate computational task, which aims to categorize each pixel in an image into one of several specified categories or classes. Deep learning, particularly convolutional neural networks (CNNs), has revolutionized the field of semantic segmentation, then, many popular architectures have been developed for this purpose and they performed well in terms of image segmentation. In this paper, we proposed a comparison study between two typical deep learning architectures: U-Net for 2D image segmentation and the volumetric ConvNets V-Net which designed for 3D medical image segmentation. The models were evaluated on the BraTS 2020 dataset to segment different sub-regions of brain tumor. The comparison demonstrates the effectiveness and the powerful of each architecture in terms of quality and quantity of results besides their time execution. For 2D segmentation, U-Net achieved the top score of 64% in terms of mean Intersection over Union (mIoU) and 94% of Accuracy, on the other side, V-Net gets the best result for 3D segmentation by achieving 75% of mIoU and 98% of Accuracy.