Prediction of Normal & Grades of Cancer on Colon Biopsy Images at Different Magnifications Using Minimal Robust Texture & Morphological Features

Classification of colon biopsy images to normal and various cancer grades is a pivotal task for histopathologists as it involves visual analysis under the microscope at different magnifications and hence may give rise to observational inconsistency. This paper emphasis on categorization of colon biopsy images into normal, well, moderate and poor classes thereby analyzing the best magnification and classifier suited for classification. A hybrid feature set consisting of morphological and texture features are obtained from images followed by class balancing to overcome imbalancing problem and then optimized feature selection. Classifiers such as SVM, Random Forest, Multilayer Perceptron and Naive Bayes are experimented for classification. The proposed model is evaluated with colon biopsy images acquired from Aster Medcity, Kochi, India at different magnifications 10X, 20X and 40X where all the magnifications performed well, but 20X gave an improved accuracy of 94.27% with the Random Forest classifier. Advance measures based on entropy triangle are used to rank classifiers apart from the standard performance measures, where Random Forest classifier is best for the proposed model for all magnifications.