Diagnostic Performance of Artificial Intelligence in Predicting Malignant Upgrade of B3 Breast Lesions: Systematic Review and Meta-Analysis

Background/Objectives: High-risk (B3) breast lesions are a heterogeneous group with uncertain malignant potential. Methods: We systematically reviewed and meta-analyzed the ability of artificial-intelligence (AI) models to predict malignant upgrades (a ductal carcinoma in situ or an invasive carcinoma) after biopsy. A comprehensive search of medical and engineering databases through 27 July 2025 identified retrospective studies that developed or validated AI models for upgrade prediction in cohorts with ≥20 B3 lesions and confirmed outcomes at surgical excision or after ≥24 months of follow-up. Results: Three single-center studies (557 lesions, 91 upgrades) met the eligibility criteria. Pooled analysis focused on clinically meaningful operating points rather than raw accuracy metrics. Models tuned for high sensitivity achieved high negative predictive values (pooled 0.95), suggesting reliable identification of lesions suitable for surveillance, but positive predictive values were modest and heterogenous (0.15–1.00), reflecting trade-offs between avoiding missed upgrades and reducing unnecessary excisions. Only two studies reported area-under-the-receiver-operating-characteristic curves, which pooled to 0.72, indicating moderate discrimination. Conclusions: Although limited by small sample sizes and single-center designs, these findings suggest that AI could aid decision-making for B3 lesion management. Prospective multicenter validation and standardized reporting are needed to evaluate clinical utility.