Feature-based analysis of cell nuclei structure for classification of histopathological images

Pathological examination of histopathological images remains the main method of cancer diagnosis. We proposed a novel automatic feature-based analysis scheme for classification of the histopathological images. An animal model of intestinal carcinogenesis-multiple intestinal neoplasia mouse model was used to evaluate the feasibility of the nuclear structure feature for detecting early-stage carcinogenesis and assessing cancer risk. Firstly, the cell nuclei are segmented based on collaborate cell localization and the improved morphology method. Then several types of features, including shape features, statistical features and textual features (Gabor and Markov random field features) are extracted. Feature selection methods including wrapper, filter and the maximum relevance-minimum multicollinearity are applied to obtain the optimal feature set. Experimental results show that the proposed segmentation method can automatically segment histopathological images and has effective segmentation results. The maximum relevance-minimum multicollinearity method outperformed all other methods in term of classification accuracy. The textual features can effectively improve the characterization of cell nuclei structure and feature selection methods can get better classification results.