Computer Aid for the Decision to Biopsy Breast Lesions

Abstract

Abstract : The goal of this project is to improve the accuracy of the diagnosis of breast cancer from mammograms by using a computer-based system to provide the mammographer with a second opinion on whether or not to perform a biopsy. An estimated 2% to 10% of true cancers are not biopsied but are instead followed, while between 60% and 90% of breast biopsies are performed on benign lesions. This report documents progress that has been made in improving the accuracy of diagnoses from mammograms using a Case-Based Reasoning (CBR) approach. The CBR approach predicts the outcome of a biopsy from the known biopsy outcomes for cases. The current version of the CBR performs with an accuracy of 61% on a retrospective set of consecutive cases for which the clinical diagnostic accuracy was 35%. The CBR algorithm has four fundamental tasks: (1) specify a reference set of cases, (2) define a metric for the distance between cases, (3) define a rule (based on the distance metric) for selecting similar cases from the reference set, and (4) specify a classification technique for predicting the outcome of biopsy from the known outcomes of reference cases. The reference database for this study contained about 1500 cases that were referred for biopsy at Duke University Medical Center between 1992 and 2000. Each case included the mammographer's description of the lesion using the BI-RADS (TM) lexicon, known risk factors, and outcomes in the form of benign or malignant status as determined by biopsy. At a sensitivity of 0.98 relative to all biopsied lesions, the specificity of CBR was found to be 0.4. Thus, through the use of CBR 40% of the benign biopsies could have been avoided at the cost of delaying diagnosis for 2% of the malignancies. The results demonstrate the feasibility of developing CBR as a decision aid for breast biopsy using the BI-RADS lexicon to index the cases. (5 tables, 6 figures, 11 refs.)