Beyond Cancer Detection: An AI Framework for Multidimensional Risk Profiling on Contrast-Enhanced Mammography

This is the authors' abstract. We don't add key points for this paper.

Purpose: The purpose of this study is to assess whether AI-based models improve reproducibility of breast density (BD) and background parenchymal enhancement (BPE) classification and to explore whether contrast-enhanced mammography (CEM) can serve as a proof-of-concept platform for systemic risk surrogates. Materials and Methods: In this retrospective single-center study, 213 women (mean age 58.3 years; range 28–80) underwent CEM in 2022–2023. Histology was obtained when lesions were present (BI-RADS 4/5). Five radiologists independently graded BD and BPE; consensus served as the ground truth. Linear regression and a deep neural network (DNN) were compared with a simple linear baseline. Inter-reader agreement was measured with Fleiss’ κ. External validation was performed on 500 BI-RADS C/D cases from VinDr-Mammo targeted density endpoints. A secondary exploratory analysis tested a multi-output DNN to predict BD/BPE together with bone mineral density and systolic blood pressure surrogates. Results: Baseline inter-reader agreement was κ = 0.68 (BD) and κ = 0.54 (BPE). With AI support, agreement improved to κ = 0.82. Linear regression reduced the prediction error by 26% versus the baseline (MSE 0.641 vs. 0.864), while DNN achieved similar performance (MSE 0.638). AI assistance decreased false positives in C/D by 22% and shortened the reading time by 35% (6.3→4.1 min). Validation confirmed stability (MSE ~0.65; AUC 0.74–0.75). In exploratory analysis, surrogates correlated with DXA (r = 0.82) and sphygmomanometry (r = 0.76). Conclusions: AI significantly improves reproducibility and efficiency of BD/BPE assessments in CEM and supports feasibility of systemic risk profiling.