Computer-aided Detection: The Impact of Machine Learning Classifier and Image Feature Selection on Scheme Performance

Computer-aided detection and diagnosis (CAD) schemes have been developed and applied to detect suspicious lesions depicted on biomedical images. After identifying initial candidates for the targeted suspicious lesions, most CAD schemes use a pre-trained multi-image-feature based machine learning classifier to classify these candidates into two groups of positive and negative detections. Although a large number of image features and machine learning classifiers have been developed and tested using different image databases, selecting the optimal image features and a machine learning classifier remains a challenged issue in CAD development. In this study, we assembled two independent image datasets for training and testing. We optimized four machine learning classifiers (namely, artificial neural network, support vector machine, Bayesian belief network, and k-nearest neighbor algorithm), which were trained and tested using the same dataset with two sets of image features. The results showed that using the first feature set, the case-based classification performance of four classifiers measured with the normalized areas under FROC-type performance curves (AUCs) ranged from 0.925 to 0.943 without statistically significant difference (p > 0.05). When using the second image feature set, AUC values of four classifies significantly reduced to the range from 0.886 to 0.903 (p < 0.01). This study suggested that although these four classifiers were built based on different machine learning concepts, their actual performance levels were likely to converge to the similar level when using the same image features and an independent testing dataset. Thus, selecting image features rather than a machine learning classifier plays a more important role in determining CAD performance.