Abstract-Early diagnosis and removal of colonic polyps is effective in the elimination of subsequent carcinoma. This paper presents a new approach for computer-aided detection of polyps. The approach mimics the way the radiologists view CT abdomen images and utilizes several geometric attributes obtained from many triples of mutually orthogonal planes. The histogram of the attributes obtained from a sufficiently large number of perpendicular random images serves as a robust signature to represent the shape. We combine the new 3-D pattern recognition with a support vector machine classifier, and show that the number of the false positive detections in the initial polyp detection studies can be substantially reduced. One of the main contributions of this study is the thorough analysis of planar geometrical attributes. When an appropriate combination of planar attributes is used, the false positive rate is reduced by 87 percent beyond that of the initial stage detector, while maintaining a sensitivity level of 95 percent. Using such methods, radiologists should be able to view CTC data much more efficiently and accurately than without CAD. Keywords -Computer aided diagnosis, vector quantization, support vector machine, polyp detection
I. INTRODUCTIONColon cancer is the second leading cause of cancer deaths in the USA [1]. Previous research has shown that adenomatous polyps have a high probability of developing into subsequent colorectal carcinoma [2]. Detection and removal of pre-cancerous polyps can prevent eventual cancer development. As such, a cost-effective and patientcomfortable screening procedure is desirable in order to diagnose the disease in an earlier stage.CT colonoscopy (CTC) is a recently developed, noninvasive screening method that combines spiral CT data acquisition of the air-filled and cleansed colon with 3-dimensional imaging software to create endoscopic images of the colonic surface [9]. While initial results are promising, the method is limited partly due to the extensive amount of radiologist time involved in the interpretation process. Therefore, an automated computer-aided detection method for polyps is necessary to increase efficiency prior to the widespread use of CTC for screening.Automated polyp detection is a new, but rapidly growing area of research. The problem of identifying colonic polyps is very challenging because they come in various sizes and shapes, and because thickened folds and retained stool may mimic their shape and density. Fig. 1 that normals to the colon surface will intersect with neighboring normals depending on the local curvature features of the colon. In [7], Gokturk and Tomasi designed a method where a sphere is fit locally to the isodensity surface passing through every CT voxel in the wall region and densely populated nearby sphere centers are considered as polyp candidates.Due to the large number of false positive detections, all of the methods mentioned above can be considered more as polyp candidate detectors than polyp detectors. This paper presents a sta...