结合卷积神经网络与图卷积网络的乳腺癌病理图像分类研究  

Research on breast cancer pathological image classification combined with convolution neural network and graph convolution network

作　　者：汪琳琳 施俊[1] 韩振奇 刘立庄[2] WANG Linlin;SHI Jun;HAN Zhenqi;LIU Lizhuang(School of Communication and Information Engineering,Shanghai University,Shanghai 200444;Shanghai Advanced Research Institute,Chinese Academy of Sciences,Shanghai 201210)

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]中国科学院上海高等研究院,上海201210

出　　处：《北京生物医学工程》2021年第2期130-138,共9页Beijing Biomedical Engineering

基　　金：国家自然科学基金重点项目(81830058);上海市科学技术委员会科技创新行动计划(17411953400、18010500600、18411967400)资助。

摘　　要：目的乳腺癌的精确诊断对于后续治疗具有重要临床意义,组织病理学分析是肿瘤诊断的金标准。卷积神经网络(convolution neural network,CNN)具有良好的局部特征提取能力,但无法有效捕捉细胞组织间的空间关系。为了有效利用这种空间关系,本文提出一种新的结合CNN与图卷积网络(graph convolution network,GCN)的病理图像分类框架,应用于乳腺癌病理图像的辅助诊断。方法首先对病理图像进行卷积及下采样得到一组特征图,然后将特征图上每个像素位置的特征向量表示为1个节点,构建具有空间结构的图,并通过GCN学习图中蕴含的空间结构特征。最后,将基于GCN的空间结构特征与基于CNN的全局特征融合,并同时对整个网络进行优化,实现基于融合特征的病理图像分类。结果本研究在提出框架下进行了3种GCN的比较,其中CNN-sGCN-fusion算法在2015生物成像挑战赛乳腺组织学数据集上获得93.53%±1.80%的准确率,在Databiox乳腺数据集上获得78.47%±5.33%的准确率。结论与传统基于CNN的病理图像分类算法相比,本文提出的结合CNN与GCN的算法有效融合了病理图像的全局特征与空间结构特征,从而提升了分类性能,具有潜在的应用可行性。Objective The accurate diagnosis of breast cancer is of great clinical significance for subsequent treatment,and histopathological analysis is the gold standard for tumor diagnosis.Convolution neural network(CNN)has good local feature extraction capabilities,yet it cannot effectively capture the spatial relationship between cells and tissues.In order to effectively utilize this spatial relationship,this paper proposes a new pathological image classification framework combining CNN and graph convolution network(GCN)for the auxiliary diagnosis of breast cancer pathological images.Methods Firstly,the pathological image is convoluted and subsampled to get a group of feature maps.Then,the feature vector of each pixel position on the feature maps is represented as a node to construct the graph with spatial structure,and the spatial structure features contained in the graph are learned by GCN.Finally,the spatial structure features based on GCN are fused with the global features based on CNN,and the whole network is optimized at the same time to achieve pathological image classification based on fusion features.Results This research compares three types of GCN under the proposed framework.Among them,the CNN-sGCN-fusion algorithm achieves 93.53%±1.80%accuracy on Bioimaging Challenge 2015 Breast Histology Dataset,and 78.47%±5.33%accuracy on Databiox breast dataset.Conclusions Compared with the traditional pathological image classification algorithms based on CNN,the algorithm proposed in this paper combines the global and spatial structure features of pathological images effectively,which improves the classification performance and has potential application feasibility.

关 键 词：乳腺癌 病理图像分类 图卷积网络 卷积神经网络 空间相关性 

分 类 号：R318.04[医药卫生—生物医学工程]