Similarity in Mammography CAD Using CBIR Approach: A Validation Study

To provide assistance for radiologists in mammographic screening, many computer-aided detection and diagnosis systems (CAD) have been developed. However, there are a lot of problems which should be addressed in conventional mammographic CAD system, such as the relatively lower performance in detecting malignant masses, especially those subtle masses. The reasons which caused those errors may be the black-box type approach, which only cuing those suspicious masses but it is different to explain the reasoning of the CAD decision-making. Mammographic CAD using content-based image retrieval is another new type of CAD which can provide visual assistance instead of the type of black box method in conventional CAD for radiologists. Unlike those conventional CAD, in content-based image retrieval (CBIR) CAD, several most similar regions of interest (ROIs) are provided to radiologists as well as the decision index (DI) of one ROI which being a positive region. It has been proved that this visual aid tool could improve radiologists' performance. At present, there are two common types of CBIR CAD based on the calculation of similarity between testing ROI and reference ROI, one is the multi-feature based methods, and the other one is pixel-value-based template matching methods. The typical techniques used in these two types of CBIR CAD are multi-feature-based K-nearest neighbor (KNN) and template matching based system using mutual information (MI). The objective of this paper is to evaluate the performance of those methods commonly used in CBIR and discuss the approaches to improve CAD performance.