Segmentation of Hematoxylin-Eosin stained breast cancer histopathological images based on pixel-wise SVM classifier  

作　　者：QU AiPing CHEN JiaMei WANG LinWei YUAN JingPing YANG Fang XIANG QingMing MASKEY Ninu YANG GuiFang LIU Juan LI Yan 

机构地区：[1]State Key Laboratory of Software Engineering, School of Computer, Wuhan University [2]Center of High Performance Computing, Huaihua University [3]Department of Oncology, Zhongnan Hospital of Wuhan University, Hubei Key Laboratory of Tumor Biological Behaviors & Hubei Cancer Clinical Study Center [4]Department of Pathology, Zhongnan Hospital of Wuhan University

出　　处：《Science China(Information Sciences)》2015年第9期52-64,共13页中国科学（信息科学）（英文版）

基　　金：supported by National Natural Science Foundation of China(Grant No.81230031/H18);National Science Foundation of China(Grant No.61272274);Program for New Century Excellent Talents in Universities(Grant No.NCET-10-0644);Open Research Fund of State Key Laboratory of Hybrid Rice(Wuhan University)(Grant No.KF201301);Hubei Provinces Outstanding Medical Academic Leader Program and Natural Science Foundation of Hubei Province(Grant No.2013CFB374)

摘　　要：Hematoxylin-Eosin （HE） staining is the routine diagnostic method for breast cancer （BC）, and large amounts of HE stained histopathological images are available for analysis. It is emergent to develop com- putational methods to efficiently and objectively analyze these images, with the aim of providing potentially better diagnostic and prognostic information for BC. This work focuses on analyzing our in-house HE stained histopathological images of breast cancer tissues. Since tumor nests （TNs） and stroma morphological charac- teristics can reflect the biological behaviors of breast invasive ductal carcinoma （IDC）, accurate segmentation of TNs and the stroma is the first step towards the subsequent quantitative analysis. We first propose a method based on the pixel-wise support vector machine （SVM） classifier for segmenting TNs and the stroma, then extract four morphological characters related to the TNs from the images and investigate their relationships with the patients＇ 8-year disease free survival （8-DFS）. The evaluation result shows that the classification based segmentation method is able to distinguish between TNs and stroma with 87.1% accuracy and 80.2% precision, suggesting that the proposed method is promising in segmenting HE stained IDC histopathological images. The Kaplan-Meier survival curves show that three morphological characters （number of TNs, total perimeter, and average area of TNs） in the images have statistical correlations with 8-DFS of the patients, illustrating that the segmented images can help to identify new morphological factors in IDC TNs for the prediction of BC prognosis.Hematoxylin-Eosin （HE） staining is the routine diagnostic method for breast cancer （BC）, and large amounts of HE stained histopathological images are available for analysis. It is emergent to develop com- putational methods to efficiently and objectively analyze these images, with the aim of providing potentially better diagnostic and prognostic information for BC. This work focuses on analyzing our in-house HE stained histopathological images of breast cancer tissues. Since tumor nests （TNs） and stroma morphological charac- teristics can reflect the biological behaviors of breast invasive ductal carcinoma （IDC）, accurate segmentation of TNs and the stroma is the first step towards the subsequent quantitative analysis. We first propose a method based on the pixel-wise support vector machine （SVM） classifier for segmenting TNs and the stroma, then extract four morphological characters related to the TNs from the images and investigate their relationships with the patients＇ 8-year disease free survival （8-DFS）. The evaluation result shows that the classification based segmentation method is able to distinguish between TNs and stroma with 87.1% accuracy and 80.2% precision, suggesting that the proposed method is promising in segmenting HE stained IDC histopathological images. The Kaplan-Meier survival curves show that three morphological characters （number of TNs, total perimeter, and average area of TNs） in the images have statistical correlations with 8-DFS of the patients, illustrating that the segmented images can help to identify new morphological factors in IDC TNs for the prediction of BC prognosis.

关 键 词：breast cancer histopathology images segmentation image analysis support vector machine survival analysis 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] R737.9[自动化与计算机技术—计算机科学与技术]