Liang Meiyan, Wang Xikai, Li Bo, Zhang Shupeng, Javed Muhammad Hamza, Wang Yuxuan, Jia Xiaojun, Wang Lin
Shanxi Key Laboratory of Wireless Communication and Detection, School of Physics and Electronic Engineering, Shanxi University, Taiyuan 030006, China.
Shanxi Key Laboratory of Wireless Communication and Detection, School of Physics and Electronic Engineering, Shanxi University, Taiyuan 030006, China.
Comput Methods Programs Biomed. 2025 Oct;270:108970. doi: 10.1016/j.cmpb.2025.108970. Epub 2025 Jul 16.
Graph-based methods are widely applied in whole-slide histopathology images (WSI) analysis since they can effectively capture spatial relationship between nodes. However, existing methods focus on promoting positive nodes to have similar representations while ignoring the expression of negative samples of each node, failing to fully utilize various diagnostic information for comprehensive analysis.
In this paper, we propose a Dual Collaboration Heterogeneous Graph Convolutional Network (DCH-GCN) framework that considers both positive and negative samples implicit in whole-slide images (WSIs). Specifically, the framework consists of two complementary graphs: a positive edge homogeneous subgraph (PEH-subgraph), constructed using positive samples, and a negative edge heterogeneous subgraph (NEH-subgraph), built from negative samples. These two subgraphs collaboratively capture discriminative patch features within WSIs. The PEH-subgraph encourages spatially adjacent patches to learn similar feature representations, whereas the NEH-subgraph utilizes negative samples to enhance difference for patches exhibiting distinct morphology. In addition, we introduce a negative sample selection principle based on k-DPP and a two-stage instance clustering process to ensure the diversity and rationality of selected negative samples.
Our method was evaluated on three public datasets, achieving an ACC of 0.937, AUC of 0.943, and F1 score of 0.952 on CAMELYON16 for cancer identification; an ACC of 0.923, AUC of 0.965, and F1 score of 0.926 on TCGA-NSCLC for subtype classification; and an ACC of 0.453, AUC of 0.648, and F1 score of 0.445 on TCGA-COAD for cancer staging.
Selecting appropriate negative and positive samples for each patch to construct DCH-GCN can more comprehensively represent the topological information of WSI images and improve the overall prediction performance.
基于图的方法在全切片组织病理学图像（WSI）分析中得到广泛应用，因为它们能够有效捕捉节点之间的空间关系。然而，现有方法侧重于促使正样本具有相似的表示，却忽略了每个节点负样本的表达，未能充分利用各种诊断信息进行综合分析。
在本文中，我们提出了一种双协作异构图卷积网络（DCH - GCN）框架，该框架考虑了全切片图像（WSIs）中隐含的正样本和负样本。具体而言，该框架由两个互补的图组成：一个使用正样本构建的正边缘同质子图（PEH - 子图）和一个由负样本构建的负边缘异质子图（NEH - 子图）。这两个子图协同捕捉WSIs内的判别性图像块特征。PEH - 子图鼓励空间相邻的图像块学习相似的特征表示，而NEH - 子图利用负样本增强具有不同形态的图像块之间的差异。此外，我们引入了基于k - DPP的负样本选择原则和两阶段实例聚类过程，以确保所选负样本的多样性和合理性。
我们的方法在三个公共数据集上进行了评估，在CAMELYON16上进行癌症识别时，准确率（ACC）达到0.937，曲线下面积（AUC）为0.943，F1分数为0.952；在TCGA - NSCLC上进行亚型分类时，ACC为0.923，AUC为0.965，F1分数为0.926；在TCGA - COAD上进行癌症分期时，ACC为0.453，AUC为0.648，F1分数为0.445。
为每个图像块选择合适的负样本和正样本以构建DCH - GCN，可以更全面地表示WSI图像的拓扑信息并提高整体预测性能。
