Iterative Label Propagation Based on Semi-Supervised Learning for Classifying Benign and Malignant Pulmonary Nodules

Classification of benign and malignant pulmonary nodules is a critical task for developing a computer-aided diagnosis (CAD) system of lung cancer. The graph-based semi-supervised learning (SSL) has received considerable attention in machine learning community. Most of existing graph-based SSL algorithms propagate label information from the labeled data to the unlabeled data that are connected to the labeled data on a fixed graph according to pairwise similarities between the data. However, such a fixed similarity graph in the feature domain is unreliable in presence of image noise or outliers. To address these problems, we propose a novel classification algorithm based on semi-supervised iterative label propagation (SSILP) for classifying benign and malignant pulmonary nodules. To construct a graph, we propose a simple yet effective graph construction method. Geodesics distance is employed to define a weighted matrix. To propagate the label information of labeled data, an iteration procedure is proposed to reduce the risk of erroneous propagation. Experimental results on the LIDC dataset demonstrate that the proposed SSILP algorithm achieves the satisfactory classification results in terms of accuracy, sensitivity and specificity.