Breast biopsy prediction using a case-based reasoning classifier for masses versus calcifications

Abstract

We investigated how the subdivision of breast biopsy cases into masses and calcifications influences breast cancer prediction for a case-based reasoning (CBR) classifier system. Mammographers' BI-RADS (TM) descriptions of mammographic lesions were used as input to predict breast biopsy outcome. The CBR classifier compared the case to be examined to a reference collection of cases and identified similar cases. The decision variable for each case was formed as the ratio of malignant similar cases to all similar cases. The reference data collection consisted of 1433 biopsy-proven mammography cases, and was divided into 3 categories: mass cases, calcification cases, and other. Performance was evaluated using ROC analysis and Round Robin sampling, and variance was estimated using a bootstrap analysis. The best ROC area for masses was 0.92+/- 0.01. At 98% sensitivity, about 209 (51%) patients with benign mass lesions might have been spared biopsy, while missing 5 (2%) malignancies. The best ROC area for calcifications was only 0.64+/- 0.02. At 98% sensitivity, 50 (12%) benign calcification cases could have been spared, while missing 5 (2%) malignancies. The CBR system performed substantially better on the masses than on the calcifications.