Potential role of AI decision support in reducing unnecessary ultrasound-guided breast biopsies

Abstract

Background Ultrasound (US)-guided breast biopsy is widely used for evaluating suspicious breast lesions; yet, many biopsied lesions prove to be benign, resulting in potentially unnecessary invasive procedures. Improving diagnostic specificity without compromising sensitivity remains a key challenge in breast imaging, and artificial intelligence-based decision support systems have emerged as promising tools to assist radiologists in biopsy decision-making. This study aimed to compare the diagnostic effectiveness of an AI-DS system with that of experienced breast radiologists in evaluating lesions that underwent US-guided biopsy, and to assess the potential of AI-DS in reducing unnecessary biopsies. Methods BI-RADS scores assigned by radiologists and the AI-DS were retrospectively analyzed in 453 breast lesions biopsied between October 2021 and March 2024. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and area under the curve (AUC) were calculated for reader A, reader B, and the AI-DS. Diagnostic performance was compared using ROC analysis, McNemar tests, and generalized estimating equations. Results The AI-DS demonstrated significantly higher specificity, PPV, and accuracy than both radiologists (all p < 0.001). The AUC was 0.681 (95% CI = 0.633-0.729) for the AI-DS, 0.556 (95% CI = 0.502-0.610) for reader A, and 0.574 (95% CI = 0.521-0.627) for reader B. Subgroup analyses showed that the AI-DS yielded more accurate results than radiologists across most subcategories. In a hypothetical scenario, if 84 BI-RADS 4A lesions downgraded by AI-DS had not been biopsied, the biopsy rate could have been reduced by 18.6% (84/453) while missing only one malignant lesion. Conclusion The AI-DS demonstrated sensitivity comparable to radiologists while providing higher specificity and accuracy. In cases where radiologists classify lesions as BI-RADS 4A but the AI-DS predicts BI-RADS 2 or 3, the AI output could prompt reconsideration of biopsy necessity and support short-term imaging follow-up instead.