Background: Non-mass breast lesions (NMBLs) pose significant diagnostic challenges in breast imaging. This study aimed to evaluate the diagnostic performance of integrating an artificial intelligence-powered ultrasound (AI-US) system with digital mammography (DM) for distinguishing benign from malignant NMBLs. Additionally, the study assessed the model’s short-term risk stratification capability and its temporal stability over clinically relevant decision intervals. Methods: In this retrospective, single-center study, 118 patients with 120 NMBLs were enrolled. Imaging assessments were performed using a triple-blinded design. A Cox proportional hazards model was employed to identify predictors of malignancy. To assess the robustness of risk stratification across clinically relevant follow-up intervals, time-dependent receiver operating characteristic (ROC) analysis was performed at 6, 12, and 24 months. Results: Cox model analysis identified lesion size (hazard ratio HR = 1.92 per cm increase), suspicious malignant calcifications on DM (HR = 12.7), and the AI-US malignant risk score (HR = 1.32 per 0.1-unit increase) as independent predictors of malignancy (all p < 0.001). The combined diagnostic model demonstrated strong performance in time-dependent ROC analysis. The area under the curve (AUC) was 0.93 (95% confidence interval CI: 0.88–0.97) at 6 months and 0.88 (95% CI: 0.81–0.93) at 24 months, significantly outperforming either modality alone. The combined model achieved a negative predictive value of 93.8%, potentially reducing unnecessary biopsies by 46%. Conclusions: The integration of AI-US with DM improves diagnostic accuracy for NMBLs and provides a robust risk stratification that remains stable over follow-up intervals of 6 to 24 months. This multimodal approach enables precise risk stratification, potentially reducing unnecessary biopsies and supporting personalized follow-up strategies.
Han et al. (Wed,) studied this question.