Automated segmentation of brain tumor based on improved U-Net with residual units

The study presents a new approach to automate segmentation of clinically significant brain tumor and, to a certain extent, addresses two major issues associated with brain tumor segmentation, namely, structural complexity and class imbalance. By combining the constructed new residual learning model and the structural advantages of U-Net convolutional neural network (CNN), a novel residual structural network was developed. First, the residual unit is constructed to further construct the entire network architecture, which can simplify the training process of the convolutional network and enrich the skip connections of the network to improve the feature extraction effect. Second, the fusion of basic classifiers dynamically trained with samples of different categories can achieve greater flexibility and higher accuracy. First, maximum pooling and batch normalization were employed to improve regularization; thus, the model includes fewer network parameters but exhibits superior performance. The method was subsequently verified on the BRATS 2018 database. For High Grade Gliomas (HGG) and Low Grade Gliomas (LGG) mixed data, the method achieved Dice coefficient metric of 95.36, 92.86, and 89.93; Recall metric of 95.14, 92.04, and 89.69; and Intersection over union metric of 91.17, 86.79, and 81.93 for the complete, core, and enhancing regions, respectively.