An Investigation of Texture Features Based on Polyp Size for Computer-Aided Diagnosis of Colonic Polyps

Abstract

Computer-aided diagnosis of polyps has played an important role in advancing the screening capability of computed tomographic colonography (CTC). Texture features, including intensity, gradient, and curvature, are the essential components in differentiation of neoplastic and nonneoplastic polyps. Clinical study has shown that the malignancy rates of polyps increase in correlation to their size. In this chapter, we present a study to investigate the effect of separating polyps based on size on the performance of machine learning. First, the volume of interest of each polyp was extracted and further confirmed by Radiologists. All polyp masses in this study have a diameter size ranging from 6 to 30 mm. Then, we group polyps into three groups based on their sizes: 6–9 mm, 10–30 mm, and a combined group of 6–30 mm. The corresponding malignancy risks of polyps were also recorded. From each polyp volume, we extracted the traditional 14 Haralick texture features plus 16 additional features with a total 30 texture features. Those features were further grouped as descriptors for intensity, gradient, and curvature characteristics. Finally, we employed the Random Forest classifier to differentiate neoplastic and nonneoplastic polyps. The proposed texture feature analysis was studied using 228 polyp masses. We generated a measure of area under the curve (AUC) values of the receiver operating characteristics (ROC) curve. The performance was determined by assessing the sensitivity and specificity values. Experimental results demonstrated that gradient and curvature features were ideal for differentiation of the malignancy risk for medium-sized polyps, whereas intensity feature was better for smaller-sized polyps.