Weakly-Supervised Classification of Pulmonary Nodules Based on Shape Characters

摘要

Accurate classification and recognition of pulmonary nodules is an important and key process of Computer-Aided Diagnosis (CAD) system in lung cancer diagnose. Although it has become an increasingly popular research topic, it remains a lot of scientific and technical challenges. Not only do we lack the accurate and effective algorithm of recognition and classification, but also we have difficulties in shape features representation and samples labeling. So this paper presents a weakly-supervised method based on the Partial Label Error-Correcting Output Codes (PL-ECOC) algorithm for solving nodules' classification problem. During the training phase, we use a small amount of labeled pulmonary nodules from experts as weakly-supervised information, for generating a binary classifier. This classifier will be used to compare the Humming distance with the testing samples, in order to obtaining the final category labels. Experiments on Lung Imaging Database Consortium (LIDC) and real-world data sets have shown the efficient performance of our method.