Investigation of Three-Group Classifiers to Fully Automate Detection and Classification of Breast Lesions in an Intelligent CAD Mammography Workstation

Abstract

Abstract : Our goal is to develop a fully automated classification scheme for computer-aided diagnosis in mammography. Our proposed scheme would classify computer detections into three groups: malignant lesions, benign lesions, and false-positive computer detections. We proved that the area under the ROC curve (AUC) is not useful in classification tasks with three or more groups, and showed that the three decision boundary lines used by the three-group ideal observer are intricately related to one another. We analyzed several recently proposed three-group classification methods in terms of the ideal observer. We collected a database of 270 mammographic images with clustered microcalcification lesions. We have developed a novel performance metric that may generalize better than AUC to tasks with more than two groups. A three-group classifier could potentially allow radiologists to detect more malignant breast lesions without increasing their false-positive biopsy rates.