A Two-Stage Special Feature Deep Fusion Network with Spatial Attention for Hippocampus Segmentation

This paper proposes a method for automatically segmenting the hippocampus by a two-stage special feature deep fusion network based on spatial attention. The method extracts specific features of different resolutions to supplement the image detail information which are important for semantic segmentation and are partially lost in the downsampling process. At the same time, the spatial attention mechanism is used to solve the imbalance between the hippocampus and the background. The verification results on the NITRC dataset show that the method has achieved higher quality hippocampal segmentation results than U-Net++.