Using contextual information to classify nuclei in histology images

Nucleus classification is a central task in digital pathology. Given a tissue image, our goal is to classify detected nuclei into different types, for example nuclei of tumor cells, stroma cells, or immune cells. State-of-the-art methods achieve this by extracting different types of features such as morphology, image intensities, and texture features in the nucleus regions. Such features are input to training and classification, e.g. using a support vector machine. In this paper, we introduce additional contextual information obtained from neighboring nuclei or texture in the surrounding tissue regions to improve nucleus classification. Three different methods are presented. These methods use conditional random fields (CRF), texture features computed in image patches centered at each nucleus, and a novel method based on the bag-of-word (BoW) model. The methods are evaluated on images of tumor-burdened tissue from H&E-stained and Ki-67-stained breast samples. The experimental results show that contextual information systematically improves classification accuracy. The proposed BoW-based method performs better than the CRF-based method, and requires less computation than the texture-feature-based method.