Artificial intelligence decision support for breast ultrasound: potential to reduce benign biopsies.

PURPOSE: The purpose of this study is to evaluate the potential impact of a Food and Drug Administration (FDA)-cleared artificial intelligence (AI)-based decision support tool for breast ultrasound in reducing benign biopsies. METHODS: This study was conducted at two academic medical centers from March to May 2025 following the clinical implementation of an FDA-cleared AI-based decision support tool for breast ultrasound interpretation (Koios DS; New York, NY). Fourteen breast imaging radiologists documented BI-RADS 4 (suspicious) cases, along with the corresponding AI output. AI assessments were compared with biopsy pathology results to characterize AI classification within this selected cohort. RESULTS: Overall, 271 lesions in 253 women (mean age, 54 years; range, 24-92 years) were categorized as BI-RADS 4 and were biopsied, aspirated, and/or surgically excised. Of these, 23.2% (63/271) yielded malignant pathology, while 76.8% (208/271) were benign. The AI tool classified 90.5% (57/63) of malignant lesions as BI-RADS 4 and 9.5% (6/63) as BI-RADS 2 or 3. Among benign cases, the AI tool assigned 41.8% (87/208) as BI-RADS 2 (benign), 18.8% (39/208) as BI-RADS 3 (probably benign), and 39.4% (82/208) as BI-RADS 4 (suspicious). Retrospective application of the AI output to biopsy triage decisions suggests the potential for reduction in benign biopsy recommendations, with 60.6% (126/208) of benign lesions classified by the AI tool as BI-RADS 2 or 3. CONCLUSION: AI may help reduce benign biopsy recommendations among lesions assessed as BI-RADS 4 by radiologists; however, careful integration with radiologist judgment and clinical history is essential to ensure timely cancer diagnosis.