AI-Supported Modeling for Standardizing Breast Density and BPE in CEM - Part II

Building upon the BCSS framework validated in Part 1, this study addresses the persistent interobserver variability in background parenchymal enhancement (BPE) assessment through computational automation. Utilizing the same dataset of 213 contrast-enhanced mammography (CEM) cases, we developed a deep learning model (artificial neural networks) to automate BCSS categorization. The model reduced prediction error by 26% compared to traditional methods and improved inter-reader agreement (κ = 0.82, a 40% increase from baseline), with the strongest performance observed in high breast density cases (BI-RADS C/D), where interpretative variability most significantly affects clinical decisions. Key performance metrics including AUC (0.75), precision (0.72), and recall (0.69) demonstrate the system’s clinical viability. This AI-driven implementation of the BCSS significantly enhances standardization while preserving the diagnostic role of radiologists. Combined with the observational foundations established in Part 1, it provides an end-to-end solution for consistent BPE assessment in CEM.