A colon cancer grade prediction model using texture and statistical features, SMOTE and mRMR

Colon cancer is usually diagnosed by the visual examination of histopathological tissue specimens using microscope. Histopathologists not only identify the healthy or malignant tissue under microscope, but also determine the quantitative grade of cancer in malignant subjects. This manual process is slow, laborious and affected by the experience and work-load of the histopathologist. Therefore, researchers are working to develop automatic colon cancer grade prediction systems, which can provide a supporting second opinion to the histopathologists. This work presents a computer-aided diagnostic system utilizing Haralick texture features and statistical moments of intensity histogram based features for prediction of colon cancer grades from biopsy images. Further, Synthetic Minority Oversampling Technique (SMOTE) is applied to create a balanced dataset in terms of equal representation of all the classes. The discerning features are selected from the oversampled feature space by using minimum Redundancy Maximum Relevance (mRMR) feature selection methodology. Radial basis function (RBF) kernel of support vector machine (SVM) is applied for classification of samples into three grades of colon cancer. The texture and statistical features coupled with data balancing and feature selection enables the SVM classifier to yield good performance. The individual effect of SMOTE and mRMR on performance enhancement has also been analyzed in detail. The proposed cancer grade prediction system yields better performance compared to existing techniques on a colon cancer grading dataset in terms of various performance measures such as accuracy, sensitivity, specificity, receiver operating characteristics (ROC) curves, and Kappa statistics.