Qiang Li PhD, Presenter: Nothing to Disclose

Feng Li MD, PhD, Abstract Co-Author: Nothing to Disclose

Shusuke Sone, Abstract Co-Author: Nothing to Disclose

Kunio Doi PhD, Abstract Co-Author: Nothing to Disclose

To develop a computer-aided diagnostic (CAD) scheme for lung nodules in thin-slice CT and to assess its performance with a cross-validation testing method.

Our database consisted of 112 thin-slice CT scans with 153 nodules obtained at Shinshu University, Japan (80 scans, 91 nodules, including 41 cancers and 50 benign nodules) and the University of Chicago (32 scans, 62 nodules). The database included nodules with a large variation in size (5-30 mm, mean 11 mm), shape, and contrast (solid and ground glass opacity). Our CAD scheme consisted of lung segmentation, selective nodule enhancement, initial nodule detection, and feature extraction and analysis techniques. Automated rule-based classification technique was first used to remove many false positives, and linear discriminant analysis (LDA) was employed to further eliminate false positives. Two testing methods were employed to assess the performance of our CAD scheme. The first was a re-substitution method, in which all nodule candidates were employed for training the rules and LDA classifiers, and were re-substituted into the trained rules and LDA for testing the performance of the CAD scheme. The second was a two-fold cross-validation method, in which the entire dataset was randomly divided into two equal subsets, one for training the rules and LDA classifiers, and the other for testing the trained classifiers. The cross-validation testing method was applied 10 times to the dataset, and the average performance was determined.

For the re-substitution testing method, our CAD scheme detected 90% of nodules (138 out of 153) with a false positive rate of 5.9 per scan. For the cross-validation method, the CAD achieved a comparable false positive rate of 6.5 per scan with a reduced sensitivity of 84%, or a comparable sensitivity of 91% with 21.1 false positives per scan.

With cross-validation testing method, our CAD scheme achieved a low false positive rate and a reasonably high detection rate for nodules with a large variation in size, shape, and contrast.

K.D.: KD is a sharehold of R2 Technology Inc., Los Altos, CA, and Deus Technologies Inc., Rockville, MD.

Li, Q, Li, F, Sone, S, Doi, K, Cross-validation Evaluation of Performance for Computerized Detection of Lung Nodules in Thin-slice CT.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4414218.html