Epithelial and stromal tissues are the most common tissue breast cancer pathology images. About 80 percent breast tumors derive from mammary epithelial cells. Therefore,in order to develop computer-aided diagnosis system and analyze the micro-environment of a tumor, it is pre-requisite to segment epithelial and stromal tissues. In this paper, we propose a pixel-wise segmentation based deep convolutional network(CN-PI)model for epithelial and stromal tissues segmentation. The model initially generates two types of training patches whose central pixels are located within annotated epithelial and stromal regions. These context patches accommodate the local spatial dependencies among central pixel and its neighborhoods in the patch. During the testing phase,a square window sliding pixel-by-pixel across the entire image is used to select the context patches. The context patches are then fed to the trained CN-PI model for predicting the class labels of their central pixels. To show the effectiveness of the proposed model, the proposed CN-PI model is compared with 6 patch-wise segmentation based CN models (CN-PA) on two datasets consisting of 106 and 51 hematoxylin and eosin(H&E)stained images of breast cancer,respectively. The proposed model is shown to have F1 classification scores of 90 % and 93 %; accuracy (ACC) of 90 % and 94 %, and Matthews correlation coefficients (MCCS) of 80 % and 88 %, respectively,show improved performances over CN-PA models.%上皮和间质组织是乳腺组织病理图像中最基本的两种组织,约80 % 的乳腺肿瘤起源于乳腺上皮组织.为了构建基于乳腺组织病理图像分析的计算机辅助诊断系统和分析肿瘤微环境,上皮和间质组织的自动分割是重要的前提条件.本文构建一种基于逐像素点深度卷积网络(CN-PI) 模型的上皮和间质组织的自动分割方法.1) 以病理医生标注的两类区域边界附近具有类信息为标签的像素点为中心,构建包含该像素点上下文信息的正方形图像块的训练集.2) 以每个正方形图像块包含的像素的彩色灰度值作为特征,以这些图像块中心像素类信息为标签训练CN模型.在测试阶段,在待分割的组织病理图像上逐像素点地取包含每个中心像素点上下文信息的正方形图像块,并输入到预先训练好的CN 网络模型,以预测该图像块中心像素点的类信息.3) 以每个图像块中心像素为基础,逐像素地遍历图像中的每一个像素,将预测结果作为该图像块中心像素点类信息的预测标签,实现对整幅图像的逐像素分割.实验表明,本文提出的CN-PI 模型的性能比基于图像块分割的CN 网络(CN-PA)模型表现出了更优越的性能.
展开▼
