Segmentation of Brain Tumours Using Optimised U-Net Architecture

Recent research has shown that convolutional networks can be much deeper, more accurate, and easier to train if they have dumpier links between layers close to the input and the output. In this paper, we investigate the optimal depth, number of convolutional channels, and post-processing technique for the “brain tumour segmentation” task in the BraTS21 challenge by testing for profound supervision loss, Focal loss, decoder attention, drop block, and residual connections. We tested our suggested architecture with deep supervision, drop block, and Focal loss, together with numerous U-Net network variations, including the initially proposed U-Net, the UNETR, Residual U-Net, and Attention U-Net, in order to find the optimal training sequence.