Jianwu Xu PhD, Presenter: Nothing to Disclose

Kenji Suzuki PhD, Abstract Co-Author: Nothing to Disclose

A major challenge in the current CADe of polyps in CTC is to reduce the number of false-positive (FP) detections while maintaining a high sensitivity level. Our purpose was to develop a dimensionality reduction technique based on manifold learning for pixel-based classification for improving the specificity of a CADe scheme while keeping a high sensitivity.

Our original CADe consisted of colon segmentation, polyp candidate detection, 3D pattern feature analysis, and classification of the candidates as polyps or non-polyps with quadratic discriminant analysis. The proposed FP reduction technique consisted of Laplacian Eigenaps as a manifold learning method for dimensionality reduction and a support vector machine (SVM) classifier for classification. The Laplacian Eigenmaps operated on pixel values of the regions of interest (ROI) in CTC images directly. As a nonlinear dimensionality reduction technique, the Laplacian Eigenmaps reduced each ROI into a vector of 40 dimensions by preserving the local manifold structure of the original ROI. Then an SVM classifier was applied to the vectors and classified them into polyps and FPs. Our database consisted of CTC scans in both supine and prone positions from 106 patients with a multi-detector-row CT scanner. Each reconstructed CT section had a matrix size of 512×512 pixels, with an in-plane pixel size of 0.5-0.7 mm. Seventeen patients had 29 colonoscopy-confirmed polyps, 15 of which were 5-9 mm and 14 were 10-25 mm in size.

Our original CADe scheme achieved a by-polyp sensitivity of 96.6% (28/29) with 4.6 (489/106) FPs per patient in the data set. By using the SVM classifier coupled with the Laplacian Eigenmaps, 16.0% (78/489) of FPs were removed without any loss of true polyps in a leave-one-lesion-out cross-validation test; thus, our Laplacian Eigenmaps-based CADe achieved a 96.6% by-polyp sensitivity at an FP rate of 3.9 (411/106) per patient.

The SVM classifier coupled with the Laplacian Eigenmaps was able to remove 16% of the FPs produced by our original CADe without sacrificing sensitivity.

CADe with a very low FP rate would be useful for radiologists in improving polyp detection in CTC without impairing the efficiency in their reading.

Xu, J, Suzuki, K, False-Positive Reduction in Computer-aided Detection (CADe) of Polyps in CT Colonography (CTC) with Manifold Learning.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11034413.html