Evaluation of a Content-Based Image Retrieval Computer-Aided Diagnosis System for Breast Ultrasound Images Through Distance Similarity Measures

초록

Purpose A content-based image retrieval (CBIR) computer-aided diagnosis (CADx) system using breast masses in ultrasound images has been developed and evaluated to assist radiologists with characterization processes. The purpose of this study is to improve the accuracy of breast cancer diagnoses by analyzing images and providing quantitative information to radiologists through the CADx system. Methods Two morphological features and six texture features of breast masses were extracted to design how the CADx system retrieves a mass similar to a query mass in a reference library. Based on extracted features from breast masses, the CADx system retrieves masses which are similar to the query mass from the reference library using a k-nearest neighbor (k-NN) method. To evaluate the CBIR CADx system, 39 similarity measures (nine similarity families, F-0-F-8) based on the distance similarity were used. A receiver operating characteristic (ROC) analysis was conducted to evaluate the performance of the distance similarity measures. Conclusions The F-0 family (Mahalanobis distance) measure used with the k-NN classifier provided slightly higher performance for the classification of malignant and benign masses as compared to those with the F-1-F-8 family measures.

키워드