Automated Colon Cancer Detection Using Structural and Morphological Features

Traditionally, colon cancer is diagnosed using microscopic analysis of histopathological colon samples. The manual examination of tissue specimens is not only time consuming, but is also subjective, and depends upon a few factors such as experience and work-load of the histopathologist. Therefore, research community is constantly putting efforts in developing automated colon cancer diagnostic systems, which can provide reliable second opinion to the histopathologists. Colon biopsy image based classification is one of such computer-aided diagnostic technique that can help in quantifying the differences in the structure of normal and malignant colon tissues without needing the subjective involvement of histopathologists. In this work, we propose a computer-aided diagnostic technique that models the differences in the regular organization of normal colon tissues and irregular structure of malignant colon tissues in terms of a few features such as least square distances, elliptic Fourier descriptors (EFDs) and morphological features. These features are extracted from each colon biopsy image, and are given as input to classifier. An ensemble of SVM kernels, based on majority voting, has been developed for classification of samples into normal and malignant classes. Features are also combined to develop an information rich hybrid feature vector, which is also used for the classification. The proposed method has been tested on a colon biopsy image based dataset, and performance has been observed in terms of accuracy, sensitivity, specificity, receiver operating characteristics (ROC) curves, area under the curve (AUC), and Kappa statistics. It has been observed that each feature type performs reasonably well. Further, ensemble classifier and the hybrid feature vector have shown better performance compared to individual features and the individual classifiers, respectively.