基于边界分布注意引导的结直肠腺体分割网络

Attention and boundary guided network for colorectal gland segmentation

从结直肠组织病理图中自动精确地分割腺体轮廓对结直肠病理诊断有极大帮助,然而,腺体之间间隙狭小且不同等级的腺体具有形态变异性,准确地分割出每个腺体实例具有很大的挑战。为此,本文提出了一种基于注意力的边界引导网络用于腺体分割。具体来说,本方法在边界分支中使用理想的边界图进行监督,引入全局特征整合模块提取腺体边界,并输入后续的解码阶段辅助腺体分割。通过多尺度注意力模块提取多尺度上下文信息,增大模型的感受野。提出边界注意融合模块补充边界细节信息,进一步细化分割结果,得到最终的腺体分割图。所提出模型的有效性在公开的结直肠腺癌数据集GlaS上得到了验证,取得了优于其他网络的性能。

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.