Automated identification of colorectal glands morphology from benign images

Abstract

Adenocarcinoma is a common and prevalentform of cancer -which arises from glandular structures in epithelium. Glands form an important histological constituent of most organs which are utilized for secreting proteins and carbohydrates. The pathological variations of the glands are routinely accessed by pathologists to measure the degree of malignancy of adenocarcinomas in lung, breast, prostate and colon. The identification of these glands is tedious and time consuming work so automated analysis of identifying a glands morphology plays a vital role in mass screening the patients. To the best of our knowledge, no work has been done until now to automatically identify the sparse variations of glandular pathology. So, in this paper, we have automatically extracted the glandular pathology, classified its sub structures and used that information to detect various pathological conditions of glands morphology. To segment various glands from the candidate scan, we have computed super-pixels where similar super-pixels are fused together using robust density-based spatial clustering of application with noise (DBSCAN) algorithm. After that, we extracted 10 distinct textural and histogram features for the automated classification of glandular morphology. The proposed classification system is based on supervised support vector machines (SVM) classifier, trained on hematoxylin and eosin (H&E) stained slide images, consisting numerous histologic grades of colon cancer. The dataset from recently held M1CCA1 GLAS challenge competition is used in this research where we utilized 80 benign tumorous images depicting various glandular structures. The proposed system has correctly identified glandular morphology from 78 images achieving the accuracy of 97.5%.

Keywords: Adenocarcinoma, Benign, Glands, Super-Pixels, Support Vector Machines (SVM)