Reducing unnecessary biopsies of BI-RADS 4 lesions based on a deep learning model for mammography.

Objective: In this study, we aimed to explore the diagnostic value of a deep learning (DL) model based on mammography for Breast Imaging Reporting and Data System (BI-RADS) 4 lesions and to reduce unnecessary breast biopsies. Methods: We retrospectively collected clinical and imaging data of 557 BI-RADS 4 lesions (304 benign lesions, 195 malignant lesions, and 58 high-risk lesions which have risk of developing malignancy) obtained by mammography at Shenzhen People's Hospital and Luohu People's Hospital from January 2020 to June 2022. The DL model was constructed to predict the pathological classifications of these lesions, calculated its sensitivity, specificity, and accuracy, and evaluated its diagnostic performance using receiver operating characteristic curve and area under the curve (AUC). Results: This study included 557 patients with BI-RADS 4 lesions, including 381 patients (68.40%) with BI-RADS 4A, 106 patients (19.03%) with BI-RADS 4B, and 70 patients (12.57%) with BI-RADS 4C. For BI-RADS categories 4A, 4B, and 4C lesions, 70.9%, 27.4%, and 7.1% were respectively confirmed as benign through biopsy, surgical pathology, or follow-up. The DL model demonstrated high diagnostic performance in identifying BI-RADS 4 lesions, achieving a sensitivity of 81.0%, specificity of 76.9%, accuracy of 78.8%, and an AUC of 0.790. We found that our DL model could avoid unnecessary biopsies for BI-RADS 4 lesions by 40.6% in our included patients, reducing unnecessary biopsies for BI-RADS 4A, 4B, and 4C lesions by 55.1%, 18.9%, and 4.29%, respectively. Conclusion: Our DL model for classifying BI-RADS 4 lesions can accurately identify benign and high-risk lesions that do not necessitate biopsy, further enhancing the safety and convenience for patients.