基于弱监督学习的双分支结直肠病理图像腺体分割-《中国医学物理学杂志》

文章信息/Info

Title:: Gland segmentation in colorectal pathological image using dual-branch network based on weakly supervised learning

作者:: 李子成1; 贾伟1; 2; 赵雪芬1; 高宏娟1; 2; 1.宁夏大学信息工程学院，宁夏银川 750021； 2.宁夏“东数西算”人工智能与信息安全重点实验室，宁夏银川 750021

Author(s):: LI Zicheng1; JIA Wei1; 2; ZHAO Xuefen1; GAO Hongjuan1; 2; 1. School of Information Engineering, Ningxia University, Yinchuan 750021, China 2. Ningxia Key Laboratory of Artificial Intelligence and Information Security for Channeling Computing Resources from the East to the West, Yinchuan 750021, China

Keywords:: Keywords: weakly supervised learning colorectal pathological image gland segmentation pseudo-label class activation map

摘要:: 现有弱监督分割方法难以获得结直肠病理图像的细粒度腺体特征，导致无法生成高质量伪标签的问题，影响腺体分割的效果。为了解决上述问题，提出一种基于弱监督学习的双分支结直肠病理图像腺体分割方法。首先，将patch级结直肠病理图像输入到第一个分支网络中，通过特征交互模块和亲和度注意力融合模块实现patch级图像的局部和全局特征的交互和融合，并获得细粒度腺体特征。然后，将图像级结直肠病理图像输入到第二个分支网络中，利用局部类激活注意力模块定位腺体位置，并获得粗粒度类激活图。最后，通过细粒度腺体特征和粗粒度类激活图，得到高质量伪标签，并在分割网络中经过跨尺度连接空间感知模块，实现腺体分割。实验结果表明，将所提方法在GlaS和CRAG两个结直肠病理图像数据集中进行实验，与其他分割方法相比取得较好的分割效果，验证所提方法的有效性。

Abstract:: Abstract: To address the issue that the existing weakly supervised segmentation methods have difficulties in obtaining fine-grained glandular features from colorectal pathological images, leading to the inability to generate high-quality pseudo-labels and compromising the gland segmentation performance, a dual-branch network based on weakly supervised learning is proposed for gland segmentation in colorectal pathological image. The patch-level colorectal pathological images are input into the first branch network, where the interaction and fusion of local and global features of patch-level images are achieved through the feature interaction module and affinity attention fusion module, and fine-grained glandular features are obtained. Subsequently, image-level colorectal pathological images are input into the second branch network, where the gland locations are located using the partial class activation attention module, and coarse-grained class activation maps are obtained. Finally, high-quality pseudo-labels are derived from the fine-grained glandular features and coarse-grained class activation maps, and gland segmentation is realized in the segmentation network through the cross-scale connected spatial perception module. The tests on two colorectal pathological image datasets (GlaS and CRAG) reveal that the proposed method is superior to other segmentation methods in segmentation performance, confirming its effectiveness.