A Comparative Study of Texture Descriptors for Polyp Detection in Colonoscopy Images

Abstract

Cancer has become the second leading cause of death in the adult population of Ethiopia. Colon cancer is cancer of the large intestine (colon), most cases of colon cancer start as small, noncancerous clumps of cells called adenomatous polyps. Over time some of these polyps can become colon cancers. Polyps may be very small and show few symptoms. Colon is part of the large intestine, and it belongs to our body's digestive system. It reabsorbs large quantities of water and nutrients from undigested food products as they pass through it. Therefore, physicians recommend regular screening test in order to prevent colon cancer by identifying and removing polyps before they become cancer. Colonoscopy is the first method to detect and remove polyps. The accuracy of polyp detection depends on the attentiveness and experience of the endoscopist during the procedure. But computer aided algorithms helps to increase the accuracy of polyp detection .Texture descriptors are one of the methods used to detect colon polyp. Texture is a property that represents the surface and structure of an image. The motivation behind using texture information for polyp detection is that the polyp has different color, shape, size, and appearance. Various texture descriptors are used for polyp detection. However, no work has been reported on the performance comparison of texture descriptors on publicly available polyp dataset and which combines several texture descriptors to improve the accuracy of automatic polyp detection systems even though improvement was reported for natural image datasets. In this research, the performance of the texture descriptors is studied in isolation and by combining multiple of them on recently available mayo clinic large polyp dataset using MATLAB 2018a. The dataset contains 18,500 polyp images with their ground truth image masks. The optimal single descriptor and combination of the texture descriptors which can achieve high classification rate are determined. In this research, the combination of the Wavelet transform, Local binary pattern and grey level co-occurrence matrix gives the highest classification accuracy of 93.74%. The sensitivity and specificity results are 0.934 and 0.9425 while using Support vector machine as a classifier.