Toward opportunistic coronary artery calcium assessment from dual-layer dual-energy chest radiography using deep learning: a simulation study
Background: Coronary artery calcium (CAC) predicts cardiovascular risk. Dual-energy (DE) chest radiography may enable accessible opportunistic CAC assessment by separating calcifications from overlapping bone. Purpose: To evaluate multi-task deep learning for CAC localization and CT-referenced score prediction and assess projection, image representation, and exposure effects. Methods: This retrospective simulation study used 550 calcium-scoring CT exams with reference CAC masks and Agatston scores. Dual-layer DE radiographs were simulated with a physics-based detector model. A modified nnU-Net jointly predicted CAC localization and scores from continuous probability maps. Comparisons included AP, 45°, and lateral views; lateral single-energy (SE), DE bone, and high-pass filtered DE bone images; and dose-matched models at 70, 50, and 24 mAs. Patient-level splitting reserved 10% for independent testing; five-fold models provided averaged inference. Metrics included Dice and AUC at CAC thresholds >0, >100, >400, and >1000, with bootstrap uncertainty and paired DeLong tests with Holm-Bonferroni correction. Results: Lateral views achieved Dice 0.34 versus 0.23 (AP) and 0.22 (45°), with AUCs for CAC >100 of 0.86, 0.81, and 0.77, respectively; AUC differences were not significant after correction. Lateral DE bone outperformed SE (Dice: 0.33 vs. 0.19; AUC for CAC >100: 0.86 vs. 0.64), with significant AUC advantages at >100, >400, and >1000. High-pass filtering added little classification benefit. Dice decreased with exposure, while AUC differences across doses were not significant. Conclusions: Simulation results support lateral dual-layer DE bone imaging for opportunistic CAC assessment, with significant advantages over SE. Physical and clinical validation is needed.