Brain Tumor Image Segmentation Method Based on Multi-scale and Attention

Abstract

Brain tumor, as a high-risk disease of the brain, has been a threat to human life and health. In order to help doctors diagnose some parts of brain tumor accurately in hospitals, multi-scale fusion brain tumor image segmentation network has shown strong feature extraction ability and image segmentation accuracy improvement. In the original Unet network, only the feature information of the current layer is used in the jump connection layer, and the relevant feature information of the shallow network is ignored, so the segmentation accuracy will be affected accordingly. We use an improved segmentation network to solve this problem. Firstly, the multi-scale feature fusion module MFF is added to the encoder to fuse the features of different scales to improve the segmentation ability of the network. Secondly, the attention module ResCBAM is added to the jump connection layer of the encoder and decoder to guide the encoder to adaptively learn the important feature information in the jump connection. The BraTS2020 dataset in MICCAI competition was used for ablation experiments and contrast experiments, and Dice coefficient and HD95 were used as evaluation indicators. Through the experimental results, it can be seen that the improved network can extract more features in the whole tumor, tumor core and enhanced tumor region, and the segmentation effect of brain tumors is good. At the same time, the model parameters and the number of iterations are reduced.