Enhancing Diagnostic Performance of Screening Mammography Readers Using an Intelligent Bayesian-Driven Adaptive Training System.

RATIONALE AND OBJECTIVES: To evaluate the effectiveness of an intelligent, Bayesian-driven adaptive mammogram reader training system in improving diagnostic performance of breast image readers and determining whether personalized AI-generated training can elevate readers' performance toward a predefined performance benchmark. MATERIALS AND METHODS: This prospective study assessed diagnostic performance before and after AI-guided adaptive training, with each participant serving as their own control. Eight readers (radiologists, breast physicians, and radiology trainees) undertook personalized mammogram reading training sessions generated by a Bayesian case-difficulty model derived from large-scale historical BREAST reader data (618 readers with 1746 completions of 9 screening mammogram test sets (540 cases)). Training sets were dynamically tailored to individual's diagnostic performance profiles across breast density, lesion type, and lesion size, and iteratively updated following each session. Performance metrics included specificity, case sensitivity, lesion sensitivity, ROC AUC, and JAFROC figure of merit. Training sessions conducted until the predefined performance benchmark (ROC AUC ≥ 0.90) was achieved. Pre- and post-training performances were compared using the Wilcoxon signed-rank test. Subgroup analyses for were conducted according to breast density, lesion types, and lesion sizes RESULTS: All participants reached the predefined performance benchmark after two to five training sessions with each comprised between 35 and 66 cases which were obtained over 1-8 months. Compared with baseline performances, significant post-training improvements of readers were observed in case sensitivity (0.814 vs 0.956; P = 0.012), lesion sensitivity (0.756 vs 0.890; P = 0.017), ROC AUC (0.846 vs 0.943; P = 0.012), and JAFROC (0.815 vs 0.906; P = 0.012). The largest gains were observed in both low and high dense breasts, small lesions (≤15 mm), architectural distortion, and asymmetric density. CONCLUSION: A Bayesian-driven AI adaptive training system significantly improved screening mammogram readers' performance while reducing training burden. Personalized, data-driven training offers a scalable approach to efficiently elevate reader performance toward a predefined performance benchmark, particularly for subtle and high-risk lesion presentations.