A novel approach for colon biopsy image segmentation

Colon cancer is one of the leading causes of deaths worldwide. Traditionally, colon cancer is diagnosed using microscopic analysis of colon biopsy images. However, computer based diagnosis involves acquiring a biopsy image, segmenting the image into constituent regions, extracting features, and based on features identifying cancerous and non-cancerous regions. Image segmentation that is the core process in overall diagnosis is extremely challenging due to similar color distribution in various biological regions of histopathological images. Problem gets more complicated for homogenous images or images acquired under different conditions, particularly change in magnification factor. Several segmentation schemes, proposed for colon images, do not address these problems. In this research study, we propose an un-supervised colon biopsy image segmentation scheme that incorporates background knowledge of benign and malignant tissue organization. The scheme detects elliptical epithelial cells in four angular directions of 0°, 45°, 90°, 135°, and divides elliptic constituents into distinct `primitives´. It further makes use of the distribution as well as spatial relations of these `primitives´ to define a homogeneity measure for identifying regions. Contrary to previous ones, the proposed scheme removes dependency on change in magnification and image type. Genetic algorithm (GA) has been employed to optimize several system parameters such as semi-major and semi-minor axis of ellipse, component area threshold to remove smaller components, and merge factor to merge two adjacent and similar regions. Algorithm has been tested on 100 colon biopsy images and improved segmentation accuracy has been observed when compared with segmentation results obtained using circular primitive based techniques.