A content-based image retrieval scheme for identifying lung nodule malignancy levels

Abstract

Discriminating the lung nodules benign or malignant is an important task for computer aided diagnosis of lung cancer. The malignancy of the nodules is divided into five levels in Lung Image Database Consortium (LIDC) database. In this study, a new content-based image retrieval (CBIR) scheme is proposed for classification of the lung nodules with different ratings. A lung nodule dataset is assembled from LIDC lung CT database. Two nodule density dependent features are calculated to depict each nodule. For each queried nodule, a CBIR scheme is used to search for ten most similar reference nodules. A malignancy probability is computed to predict the malignancy of the lung nodules. The results show that accuracy and the area under the curve (AUC) are 0.6655 and 0.6901 when classifying nodules moderately suspicious for cancer (rating 4) and highly suspicious (rating 5). The ACC and AUC are 0.9231 and 0.8659 when testing our scheme on differentiating benign and malignant cases.