Semi-supervised graph embedding-based feature extraction and adaptive kernel-based classification for computer-aided detection in CT colonography

Abstract

Feature extraction and classification are two important parts in computer-aided detection (CAD) of abnormalities. For feature selection, typically we are interested in determining which, of a large number of potentially redundant or noisy features, are most discriminative for classification. Graph embedding is a relatively new category of feature selection methods. It uses a graph structure to model the geometric and discriminant structure of the data manifold. It has obvious advantages over other conventional feature selection methods like Principal Component Analysis (PCA). However, it does not properly incorporate prior labeling information like Partial Least Square (PLS) does. In this paper, we deal with feature selection for CAD of colonic polyps with graph embedding method in a semi-supervised style (SemiGE), which considers the label information as additional constraints. Specifically, we constrain the space of solutions that we explore only to contain embedding results which are consistent with the labels. Experiments were conducted on patients' datasets to compare its performance with other conventional and popular feature selection methods. The results illustrated that this presented SemiGE strategy outperforms all other exciting feature selection methods, such as PCA, PLS, etc. On the other hand, the wellknown support vector machine (SVM) classifier has been widely used in CAD of colonic polyps in computed tomography conlonography (CTC). Several kernel functions have been implemented for the SVM classifier. Each kernel function affects noticeably to the classifying performance and variable performance has been observed among the kernel functions. A traditional implementation strategy and all the previous applications of SVM only preselect a kernel function (e.g., RBF, Sigmoid, .etc). This kind of pre-selection approach fails to provide a kernel function which is adaptive to the experimental data. In this paper, we introduce an improved SVM-based classifier with adaptive kernel implementation, AK-SVM, to conduct the polyp detection. To the best of our knowledge, this is the first one to do colonic polyp detection using SVM with adaptive kernel. The kernel operator is constructed by learning from the training data. We use the boosting paradigm to perform the kernel construction process. To do so, we modify the booster to accommodate the kernel operators. We demonstrated the effectiveness of our approach on patient datasets with comparison to the conventional SVM-based classifier with preselected kernel. We measured the performance of classification by the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC). On the dataset we experimented, the performance of the AK-SVM is systematically better than the conventional SVM with a fixed RBF kernel.