Hongbin Zhu PhD, Abstract Co-Author: Nothing to Disclose

Perry J. Pickhardt MD, Abstract Co-Author: Consultant, Viatronix, Inc Consultant, Medicsight, Inc Consultant, C.B. Fleet Company, Inc Consultant, Covidien AG

Matthew A. Barish MD, Abstract Co-Author: Consultant, Perceptive Informatics, Inc Spouse, Employee, Bracco Group

Zhengrong Liang, Presenter: Nothing to Disclose

Su Wang PhD, Abstract Co-Author: Nothing to Disclose

Yi Fan PhD, Abstract Co-Author: Nothing to Disclose

Erica J. Posniak MD, Abstract Co-Author: Nothing to Disclose

Harris L. Cohen MD, Abstract Co-Author: Nothing to Disclose

Robert S. Richards MD, Abstract Co-Author: Nothing to Disclose

00030490-DMT et al, Abstract Co-Author: Nothing to Disclose

Computer-aided detection of polyps (CADpolyp) has shown promise as an aid to computed tomographic colonography (CTC) for colon screening. This study aims to explore projection-based features of charactering the true positives (TPs) and false positives (FPs) to decrease the number of FPs while maintaining a high sensitivity in detecting TPs.

MATERIALS AND METHODS: Because of the subtle difference in image intensity variation inside a polyp v.s. its outside surrounding normal tissues, features derived by a voxel-by-voxel fashion have been shown useful, but not sufficient for FP reduction. A projection or line integral through a volume is expected to enhance the difference. For each polyp candidate, an optimal local reference frame was built automatically, and projections were sampled along the three orthogonal axes to generate three images in either gray scales or colors. These projected images reflect characteristic patterns exclusively for TPs, such as a denser island, a red core at the central part, etc. Based on these distinguishing patterns, a variety of features were built and fed into the classifier (the supported vector machine) to remove the FPs induced by rectal tubes, adherent stools, prominent folds, ileocecal valves and so on.

The projection-based features were tested on a database of 325 patient studies from 556 CT scans, of which 258 scans have polyps. In these 258 scans, there were 347 clinically significant polyps sized from 5 to 60 mm. Two experiments were performed to show the effectiveness of the new features. One used the well-established features, like the first- and second-order statistics of the shape index, CT values, etc. The other included the new projection features together with well-established features. The output of the classifier shows that at the by-polyp detection sensitivity of 95%, the number of the FPs per scan is 7.8 for the first experiment and reduced to 4.3 for the second.

At the sensitivity level of 95%, we gained an additional 44.9% reduction of FPs if the projection-based features were included. The result is statistically meaningful since our database is rather large and comes from two different institutions (DoD and University of Wisconsin Medical Center).

The projection features would improve the performance of CADpolyp and, therefore, the CTC capability for colon cancer screening.

Zhu, H, Pickhardt, P, Barish, M, Liang, Z, Wang, S, Fan, Y, Posniak, E, Cohen, H, Richards, R, et al, 0, Increasing Detection Specificity in Computer-aided Detection of Colonic Polyps by Projection-based Features for CT Colonography.  Radiological Society of North America 2009 Scientific Assembly and Annual Meeting, November 29 - December 4, 2009 ,Chicago IL. http://archive.rsna.org/2009/8010159.html