Comparative Performance of State-of-the-Art Classifiers in Computer-Aided Detection for CT Colonography

Several effective machine learning and pattern recognition schemes have been developed for medical imaging. Although many classifiers have been used with computer-aided detection (CAD) for computed tomographic colonography (CTC), little is known about their relative performance. This pilot study compares the performance of several state-of-the-art classifiers and feature selection methods in the classification of lesion candidates detected by CAD in CTC. There were four classifiers: linear discriminant analysis (LDA), radial basis function support vector machine (RBF-SVM), random forests (RF), and gradient boosting machine (GBM). There were five feature selection methods: sequential forward inclusion (SFI) of principal components (PCs), univariate filtering (UF), UF of PCs, recursive feature elimination (RFE), and RFE of PCs. A strategy of using all available features was tested also. For evaluation, 232,211 detections by a CAD system on 1,211 patients were subsampled randomly to create 10 different populations of 500 true-positive (TP) and 500 false-positive (FP) detections. The classifier performance was evaluated by use of the area under the receiver operating characteristic curve of 3 repeated 10-fold cross-validations. According to the result, the discrimination performance of the RBF-SVM classifier with feature selection by the RFE of PCs compared favorably with other methods, although no single classifier outperformed other classifiers under all conditions and feature selection schemes.