Two-step Content-based Retrieval for Pulmonary Nodule Diagnosis

Abstract

Similarity measurement of pulmonary nodules can be useful in content-based retrieval for pulmonary nodule diagnosis on computed tomography (CT). Unlike previous retrieval schemes, which concentrate on the feature extracting, we focus on the similarity measurement of pulmonary nodules. Similar to our previous studies, in this study, the pulmonary nodule dataset is from the LIDC-IDRI lung CT images, which includes 746 pulmonary nodules, 375 malignant nodules and 371 benign nodules. Each nodule is represented by a vector of 26 texture features. Two-step similarity measurement is proposed to construct a content-based image retrieval (CBIR) scheme to discriminant benign and malignant nodules. The similarities of pulmonary nodules are defined as semantic relevance and visual similarity. In the first step, semantic relevance is used to screen the nodules, which are semantic relevance to the query nodule. For the second step, visual similarity is applied to calculate the nodules, which look like the query nodules. Two Mahalanobis distances are learned to preserve semantic relevance and visual similarity of lung nodules, respectively. A retrieval scheme applies the learned Mahalanobis distances to calculate the similar nodules. Classification accuracy is used to evaluate the scheme performance, the area under the ROC curve (AUC) can reach 0.956±0.005.