Defining the Clinical Indication for Real-Time Artificial Intelligence-Assisted Ultrasound in Breast Lesions with Low Suspicion of Malignancy.

Abstract

Objectives: To evaluate whether real-time artificial intelligence (AI) breast ultrasound (US) can identify a subgroup of breast imaging reporting and data system (BI-RADS) category 4A lesions without suspicious mammographic (MMG) findings in whom unnecessary biopsy may be safely avoided.

Methods: This retrospective analysis of single-center biopsy registry included 641 consecutive US BI-RADS 4A lesions that underwent vacuum-assisted breast biopsy. Before conducting biopsy, the lesions were reassessed using a real-time AI US system (CadAI-B®), and MMG findings were classified as suspicious or non-suspicious based on formal reports. Seven prespecified biopsy-decision strategies were evaluated, including 3 AI-only strategies, 3 mammography-gated AI strategies, and the current standard strategy of biopsy all BI-RADS 4A lesions. The primary outcomes were biopsy reduction rate, missed cancers, sensitivity, specificity, positive predictive value, negative predictive value, and accuracy, with final pathology as the reference standard.

Results: Among 641 cases of US BI-RADS 4A lesions, 26 cases of malignancies (4.1%) were identified, including 14 cases of ductal carcinoma in situ (DCIS), 11 cases of invasive ductal carcinomas, and 1 case of apocrine carcinoma. AI reassigned 289 cases of lesions (45.1%) to BI-RADS categories 1-3. AI-only biopsy-deferral strategies missed 30.8% and 38.5% of malignancies when biopsy was deferred for AI BI-RADS ≤2 and ≤3 lesions, respectively, with DCIS accounting for most missed cancers. In contrast, among lesions without suspicious MMG findings, no malignancies were observed in AI BI-RADS 1-3 lesions. The sequential strategy of deferring biopsy only for lesions with AI BI-RADS ≤3 and non-suspicious MMG findings reduced biopsies by 43.5% (279/641) while maintaining 100.0% sensitivity (95% CI, 87.1-100.0%) and increasing specificity to 45.4% (95% CI, 41.4-49.4%).

Conclusion: Real-time AI breast US alone could not safely support biopsy deferral in US BI-RADS 4A lesions. However, after suspicious MMG findings had been excluded, AI identified a low-risk subgroup in whom unnecessary biopsy could be substantially reduced without compromising cancer detection. These findings suggest that the clinical utility of real-time AI breast US lies in patients without suspicious MMG findings. Prospective multicenter validation is warranted before clinical implementation.