Brain Tumor Image Segmentation Based on Global-Local Dual-Branch Feature Fusion

Accurate segmentation of brain tumor medical images is important for confirming brain tumor diagnosis and formulating post-treatment plans. A brain tumor image segmentation method based on global-local dual-branch feature fusion is proposed to improve brain tumor segmentation accuracy. In target segmentation, multi-scale features play an important role in accurate target segmentation. Therefore, the global-local dual-branch structure is designed. The global branch and local branch are deep and shallow networks, respectively, to obtain the semantic information of brain tumor in the deep network and the detailed information in the shallow network. In order to fully utilize the obtained global and local feature information, an adaptive feature fusion module is designed to adaptively fuse the global and local feature maps to further improve the segmentation accuracy. Based on various experiments on the Brats2020 dataset, the effectiveness of the composition structure of the proposed method and the advancedness of the method are demonstrated.