Brain tumor pathological image segmentation based on residual U-net ++ network

Abstract

Abstract The incidence of brain tumors has been on the rise, highlighting the need for early and accurate diagnosis. Deep learning in image segmentation can assist in automated diagnosis and diagnostic efficiency. This study proposes an improved U-Net + + network with a jump short connection instead of a long connection to be more sensitive to edge information. The residual structure replaces the convolutional blocks to prevent the model from degradation, while the residual blocks are batch normalized for increasing convergence speed. The improved U-Net + + was validated on the Tumor Imaging Archive dataset, with a Dice factor of 0.927, Accuracy of 0.995, and Sensitivity of 0.913. The improved U-Net + + network can better fuse features of different depths and update the weights of the network in time. This allows the network to approach the true value without degradation during the training process, and the segmentation accuracy of the training model is also increased.