A Content-based Image Retrieval Scheme for Lung Nodule Classification

Objective: To develop a new content-based image retrieval (CBIR) based computer-aided diagnosis (CAD) scheme to discriminate the lung nodules benign or malignant and to perform a preliminary evaluation of this CAD scheme and its robustness. Methods: Two lung nodule datasets from LIDC-IDRI lung CT database were assembled. Two nodule density related features were computed to represent each nodule. For each queried nodule, a twostep CBIR scheme was applied to retrieve the top ten most similar reference nodules. A classification likelihood value was calculated to predict the malignancy of the lung nodule. To assess the robustness of the CBIR scheme, we first tested this CAD scheme on the second dataset, and then used the second dataset to retrieve the first dataset. To verify the feasibility of the CBIR scheme, classification performance of our scheme was comparied with that of classical classifiers. Results: Through applying a leave-one-out validation method on the first dataset, an area under the ROC curve (AUC) of 0.915 was obtained, and the total classification accuracy was 83.0%. For robustness on the second dataset, the AUC was 0.727, and the total classification accuracy was 66.1%. When we used the second dataset to retrieve the first dataset, the AUC value and the total classification accuracy were 0.751 and 71.3%, respectively. The classification performance of the proposed scheme outperforms that of the classical classifiers. Conclusion: This study demonstrated that (1) a simple and efficient CBIR based CAD scheme applying two nodule density related features achieved high performance for classification of lung nodules and (2) this CAD scheme using CBIR approach also had high robustness performance in the future clinical application.