Evaluation of classifiers for computer-aided detection in computed tomography colonography

Computer-aided detection (CAD) is an emerging technique which provides an optimal method for automated detection of colonic polyps in computed tomography colonography (CTC). Differentiating true-positives (TPs) from false-positives (FPs) is one of the main tasks of CAD. One major challenge for the differentiation task is how to classify the very unbalanced datasets. Many classifiers have been introduced to perform the differentiation task and some are proved to be useful. However, there has so far been no comparative study to evaluate the effectiveness of these classifiers. In this paper, we present a comparative study, which quantitatively assesses the most commonly used classifiers, e.g., support vector machine (SVM), random forest (RF), linear discriminant analysis (LDA), artificial neural network (ANN), logistic regression (LR), quadratic discriminant analysis (QDA), and classification and regression tree (CART). The performances of these classifiers were evaluated based on 786 initially detected patches, including 64 true polyps. Our results show that SVM, RF and LDA perform the best for the detection task and also most robustly in dealing with datasets with different unbalanced level. It can be concluded that these three classifiers are strong, good classifiers. While ANN delivers less favorable result, it provides good complementary information and can be labeled as weak, good classifier. By this comparative study, we conjecture that the combination of these classifiers can be a stronger classifier and worth for further investigation, because they are complementary to each other. From this comparative study, we further conclude that integrating a strong classifier for texture analysis would be a logical choice for CAD in CTC.