A cascade-learning approach for automated segmentation of tumour epithelium in colorectal cancer

Abstract Automated segmentation of tumor epithelial tissue from histological images is a fundamental aspiration of digital pathology to improve biomarker assessment and tissue diagnosis. Accurate tumour segmentation is an important step in many automated digital image analysis applications to be used in clinical practice. In particular, segmentation of tumour, non-tumour epithelium and stromal tissue compartments on immunohistochemistry images presents a challenge. Many artifacts, such as staining and/or illumination variations, can confound image analysis. In this paper, we propose a cascade-learning approach which can diminish the impact of these artifacts. It consists of (a) a set of novel invariant features that encodes meaningful information about the appearance and shape of the region of interest and (b) a novel level set formulation where contour evolution is driven by a probabilistic model of the appearance of the region (based on fuzzy c-means). The merit of our approach is that it exploits both appearance and shape information and combines them in the tissue classification framework. We evaluate the performance of our approach on the segmentation of tumour epithelium in colorectal cancer. The experimental results show that our approach is robust to staining differences, additive noise, intensity inhomogeneities, and can cope with a limited number of training samples, when compared to the state-of-the-art tumour epithelial segmentation methods.