AI-Assisted Decision-Support Framework for Breast Density and Background Parenchymal Enhancement Assessment in Contrast-Enhanced Mammography.

Authors

Abstract

Interobserver variability in breast density and Background Parenchymal Enhancement (BPE) assessment remains a major limitation in Contrast- Enhanced Mammography (CEM) reporting consistency. Building on the BPE-CEM Standard Scale (BCSS) framework introduced in Part 1, this study aimed to evaluate whether a structured artificial intelligence-assisted decision-support model based on expert-derived variables could improve consistency in BCSS-related interpretation, particularly in disagreement-prone dense breast categories, rather than function as an autonomous image-based grading system. We retrospectively analyzed 213 consecutive CEM examinations with BI-RADS 4-5 lesions and histologically confirmed malignancy. A multilayer perceptron model operating on expert-derived patient-level variables (BPE grade, breast density category, and age) generated continuous BCSS-related estimates intended as structured decision-support outputs rather than autonomous image grading. Inter-reader agreement analyses were based on multi-reader expert assessment with consensus reference labeling, and external testing on the VinDr-Mammo dataset was limited to density-related robustness evaluation. Model performance was compared with conventional radiologist assessment. Primary endpoints were prediction error and inter-reader agreement (Fleiss' κ). Secondary analyses included Area Under the Receiver-Operating-Characteristic Curve (AUC), precision, and recall across predefined BCSS categorization thresholds. External testing using the VinDr-Mammo dataset evaluated robustness of density-related predictions across independent imaging populations. Relative to conventional assessment, the predictive framework was associated with a 26% reduction in categorization disagreement and with higher agreement estimates within the structured evaluation setting, particularly in dense breasts (BI-RADS C/D), where interpretative variability is greatest. Overall performance supported feasibility within a decision-support context (AUC 0.75; precision 0.72; recall 0.69). AI-assisted modeling of structured BCSS variables showed potential to improve reporting consistency while preserving a human-in-the-loop workflow. Findings emphasize interpretative support rather than diagnostic automation. The retrospective single-center design and enriched malignant cohort limit generalizability to screening populations and warrant prospective multicenter validation. A structured AI-assisted predictive framework showed preliminary potential to enhance consistency of BCSS interpretation in CEM, particularly in dense breasts, while requiring prospective multicenter validation before broader clinical adoption. These results support the role of transparent decision-support tools aimed at reducing observer variability while maintaining radiologist oversight.

Tags

Topics