Background: Calcification of abdominal arteries is an important risk marker in vascular disease. Automated, objective quantification methods could improve reproducibility and reduce observer dependency in clinical practice.

Purpose: To develop and evaluate a deep learning method for quantifying abdominal arterial calcification from contrast-enhanced CT angiography (CTA).

Material and methods: We retrospectively collected 223 CTA volumes, divided into 147 training and 76 test cases. Ground truth calcification segmentations were manually annotated, while vessel segmentations were generated by a previously trained neural network and manually refined. Two nnU-Net models were trained, one for artery segmentation and one for calcification segmentation. Renal, mesenteric, and common iliac arteries were shortened algorithmically. Performance of the models was evaluated using Dice score, volumetric similarity, sensitivity, precision, and Jaccard index. Calcification burden was defined as the ratio of calcified volume to artery volume. The amount and the average size of calcification clusters were investigated. The performance of the method was benchmarked against an idealized threshold-based approach and a more clinically realistic approach.

Results: The neural network achieved performance comparable to the optimized threshold-based method, with slight improvements across several segmentation metrics. Dice scores and volumetric similarity demonstrated reliable vessel and calcification detection. The predicted calcification burden score showed high correlation with the ground truth calcification burden score.

Conclusion: The proposed deep learning tool enables fast, reproducible, and observer-independent quantification of calcification in major abdominal vessels, offering a practical alternative to manual or threshold-based scoring methods.