Attention and boundary guided network for colorectal gland segmentation
Automatic and accurate segmentation of gland contour from colorectal histopathology is of great help to colorec-tal pathological diagnosis.However,due to the narrow gaps between glands and the morphological variability of dif-ferent grades of glands,accurately segmenting each gland instance is a significant challenge.In this paper,an at-tention and boundary guided network is proposed for gland segmentation.The boundary branch is supervised by the ideal boundary map,and the global feature integration module is employed to extract the gland boundary,which is then input to the subsequent decoding stage to assist gland segmentation.A multi-scale attention module is intro-duced to extract multi-scale context information and increase the model's receptive field.Finally,the boundary-at-tention fusion module is proposed to supplement boundary details,further refine the segmentation results,and ob-tain the final gland segmentation map.The effectiveness of the proposed model is validated on a publicly available colorectal adenocarcinoma dataset GlaS,achieving superior performance over other networks.
万方数据
维普
