Rationale and objectives: To achieve high performance in computer-assisted diagnosis (CAD) of polyps with computed tomographic (CT) colonography, the authors (a) developed new gradient concentration and directional gradient concentration (DGC) features for differentiating between the true-positive and false-positive (FP) findings generated by the authors' CAD scheme, and (b) used receiver operating characteristic (ROC) analysis to quantify the differentiation performance of these and other volumetric features.
Materials and methods: CT colonography was performed in 43 patients prone and supine with a helical CT scanner; there were 12 polyps in 11 patients. The polyp candidates generated by the authors' CAD scheme were characterized by nine statistics of six volumetric features, and the resulting 54 feature statistics were combined by a linear or quadratic discriminant classifier. The discrimination performance was measured with round-robin method by ROC analysis and the FP rate of the CAD scheme.
Results: The mean value of shape index (SI) yielded the highest individual ROC performance (area under the curve = 0.92). Among combinations, the mean values of SI and DGC and the variance of CT value yielded a high ROC performance (area under the curve = 0.95). With quadratic classifier, the sensitivity and FP rate of the case-based (data set-based) analysis was 100% (95%) with 2.4 FP findings per patient (1.7 FP findings per data set), respectively.
Conclusion: Combination of the mean values of SI and DGC and the variance of CT value reduced the FP rate substantially without sacrificing sensitivity. These three features are potentially useful in improving the performance of the authors' CAD scheme for detecting polyps with CT colonography.