A method for detecting colonic polyps using curve fitting from 3D abdominal CT images

Abstract

This paper proposes a method for detecting colonic-polyp candidates from 3-D abdominal CT images based on curvatures calculated from curve fitting results. The proposed method can detect polyp candidates with a very low false-positive rate. Development of a computer aided detection (CAD) system for colonic polyps is expected to be continue due to significant increase of colonic cancers in Japan. Many research groups have reported methods for detecting colonic polyp candidates based on curvatures on colonic walls. However, because they approximated the first and second-order derivatives, which are required for computing curvatures, by using intensity differentiation, detection results were significantly influenced by noise and included many false-positives. To reduce false-positives, we propose a method for geometrically calculating curvatures based on curve-fitting to iso-intensity points. First the colonic wall region is segmented by using a region-growing method from original images. For each point on the colonic wall (target point), we find iso-intensity points around the given processing area by using linear-interpolation. Curve fitting to the obtained points is performed by using a least squared error method, and the size of the processing area is automatically adjusted during the curve fitting process. The curvature of the target point is calculated from the first and second derivatives of the obtained curve, after which the shape index and curvedness are calculated. We extract points whose shapes are classified as convex and within the predefined curvedness. Colonic polyp candidates are obtained by performing connected component analysis including small component elimination. The proposed method was applied to a noisy artificial-figure image and abdominal CT images. Experimental results indicate that the proposed method detected a few false-positives while maintaining 100% true positive rate, whereas the previous method generated 10 FPs under the same experimental conditions.