Two-step segmentation of Hematoxylin-Eosin stained histopathological images for prognosis of breast cancer

Accurately predicting the risk of cancer recurrence and metastasis is critical to cancer individualized treatment. Currently, physicians commonly use histological grade, which is determined by pathologists via performing a semi-quantitative analysis of three histological and cytological features on Hematoxylin-Eosin (HE) stained histopathological images, to assess the prognosis of a breast cancer patient and the treatment option. In order to efficiently and objectively make full use of the underlying invaluable information in HE stained histopathological images, this work proposes a computational method to extract the potential morphological information as features to establish an classification model for the prognosis of cancer. Firstly, we propose a method based on the pixel-wise support vector machine (SVM) classifier for segmenting tumor nests-stroma and a method based on the marker-controlled watershed for segmenting cell nucleus, then we subclassify all image objects and extract a rich set of predefined quantitative morphological features. Secondly, a classification model based on these measurements is used to predict the binary patients´ outcome of 8-year disease free survival (8-DFS). Finally, the predict model is tested on two independent cohorts of breast cancer patients. Experimental results demonstrate the efficiency and effectiveness of the proposed method, providing valuable and reasonable prognosis information for breast cancer.