A feasibility study of high order volumetric texture features for computer aided diagnosis of polyps via CT colonography

Abstract

Differentiating pathological stages, i.e., hyperplastic (H), tubular adenoma (Ta), tubulovillous adenoma (Va) and adenocarcinoma (A), of detected colon lesions is a main task for computer aided diagnosis (CADx) of polyps for computed tomography colonography (CTC). In this paper, we propose a virtual pathological model of differentiating the polyp types based on the Haralick texture analysis model, which computes various correlations of the image density distribution inside each polyp volume. Our model explores the utility of texture features from higher order differentiations or amplification, i.e., gradient and curvature, of the image density distribution, mimicking the amplification in pathology. The first set of texture features is extracted from the gradient of the image density distribution. The second set of texture features is derived from the curvature of the image density distribution. The gain of these two sets of newly developed higher order texture features was measured using the area under the receiver operating characteristic (ROC) curve (AUC) from a database of 124 lesions (polyps and masses, confirmed by both optical colonoscopy (OC) and CTC). Support vector machine (SVM) is employed for classification. The gain by the two sets new features over the original Haralick texture model is noticeable, i.e., by 15% of improvement of the average AUC by including first set and second set of new texture features for group HvsRest and 11% for group H&TavsRest than the basic Haralick texture features.