A Deep Learning Decision Support Tool to Improve Risk Stratification and Reduce Unnecessary Biopsies in BI-RADS 4 Mammograms

Purpose: To evaluate the performance of a biopsy decision support algorithmic model, the intelligent-augmented breast cancer risk calculator (iBRISK), on a multicenter patient dataset. Materials and Methods: iBRISK was previously developed by applying deep learning to clinical risk factors and mammographic descriptors from 9700 patient records at the primary institution and validated using another 1078 patients. All patients were seen from March 2006 to December 2016. In this multicenter study, iBRISK was further assessed on an independent, retrospective dataset (January 2015–June 2019) from three major health care institutions in Texas, with Breast Imaging Reporting and Data System (BI-RADS) category 4 lesions. Data were dichotomized and trichotomized to measure precision in risk stratification and probability of malignancy (POM) estimation. iBRISK score was also evaluated as a continuous predictor of malignancy, and cost savings analysis was performed. Results: The iBRISK model’s accuracy was 89.5%, area under the receiver operating characteristic curve (AUC) was 0.93 (95% CI: 0.92, 0.95), sensitivity was 100%, and specificity was 81%. A total of 4209 women (median age, 56 years [IQR, 45–65 years]) were included in the multicenter dataset. Only two of 1228 patients (0.16%) in the “low” POM group had malignant lesions, while in the “high” POM group, the malignancy rate was 85.9%. iBRISK score as a continuous predictor of malignancy yielded an AUC of 0.97 (95% CI: 0.97, 0.98). Estimated potential cost savings were more than $420 million. Conclusion: iBRISK demonstrated high sensitivity in the malignancy prediction of BI-RADS 4 lesions. iBRISK may safely obviate biopsies in up to 50% of patients in low or moderate POM groups and reduce biopsy-associated costs.