Abstract:

There are two problems in traditional Convolutional Neural Network（CNN） classification of histopathological images, on the one hand, limited by the size of CPU, CNN can not directly train high-resolution pathological images, which inevitably destroys the spatial structure information between cells and cannot learn global feature information. On the other hand, normal cells and cancerous cells in pathological images have their own pathological image features and have certain spatial correlation, but they cannot be fully expressed in structured two-dimensional array images. This paper proposes a histopathological images classification method based on Cell Graph Convolutional Network（C-GCN）. The high-resolution pathological image is transformed into graph structure. The GraphSAGE（Graph SAmple and aggregate） module is combined with graph pooling in the traditional GCN to extract more representative general features, so that C-GCN can directly learn features from high-resolution histological images and improve the robustness of the model.

Key words: histopathological image, graph convolutional, GraphSAGE（Graph SAmple and aggregate）, graph pooling

摘要：

针对传统CNN（Convolutional Neural Network）在组织病理图像分类中存在的两个问题：其一，受限于内存大小，CNN无法对高分辨率的病理图像直接进行训练，这不可避免地破坏了细胞之间的空间结构信息，且无法学习全局的特征信息；其二，病理图像中的正常细胞和癌变细胞均有自身的病理学图像特征并且在空间上具有一定的关联性，但在结构化的二维阵列图像中无法被充分的表达。提出一种基于细胞图卷积（Cell-Graph Convolutional Network，C-GCN）的组织病理图像分类方法，将高分辨率的病理图像转换为图结构，在传统的GCN中将GraphSAGE（Graph SAmple and aggregate）模块与图池化相结合，提取出更具有代表性的一般性特征，使得C-GCN可以直接在高分辨率的组织学图像中学习特征，提高了模型的鲁棒性。

关键词: 病理图像, 图卷积, GraphSAGE, 图池化