Automated abdominal aortic calcification scoring via deep learning: a multi-center validation of LVLCRNet

TL;DR

This study developed a deep learning model called LVLCRNet to automatically score abdominal aortic calcification with high accuracy and reliability across multiple centers.

Contribution

The novel contribution is the development of LVLCRNet, a deep learning model integrating anatomic localization and contrastive rank-aware learning for automated calcification scoring.

Findings

LVLCRNet showed strong agreement with expert annotations (R² of 0.858 internally and 0.842/0.837 externally).

The model demonstrated high classification accuracy (82.94% internally and 79.62%/81.59% externally).

LVLCRNet outperformed baseline models with minimal systematic bias and strong generalizability.

Abstract

To develop and validate a deep learning model for automated quantification of abdominal aortic calcification scores (AACS) adhering to the Kauppila protocol, with multicenter clinical validation.

This retrospective multicenter study analyzed 2,660 lateral lumbar/thoracoabdominal radiographs from four centers, partitioned into development (training: n = 1,478; validation: n = 423) and test cohorts (internal: n = 211; external: n = 157 from Center C and n = 391 from Center D). We proposed the Lumbar Vertebrae Localization-Contrastive Rank-Aware Network (LVLCRNet), incorporating automatic lumbar vertebrae localization, aortic region segmentation, and contrastive rank-aware network for ordinal classification. Comparative analyses against baseline network and Lumbar Vertebrae Localization Network were conducted using expert-annotated AACS as ground truth (GT), evaluated through Wilcoxon matched-paired signed-rank test, intraclass correlation coefficient (ICC), mean absolute error (MAE), coefficient of determination (R²), Bland-Altman analysis, and multiclass accuracy.

No significant difference was found between LVLCRNet and GT, whereas the baseline network showed significant deviations from GT across all cohorts (p < 0.017, Bonferroni-corrected). LVLCRNet achieved superior agreement with GT, demonstrating R² of 0.858 (internal) and 0.842/0.837 (external), ICC of 0.916 (internal) and 0.904/0.899 (external), and MAE of 1.547 (internal) and 1.189/1.972 (external). Bland-Altman analysis showed minimal systemic bias. Classification accuracy reached 82.94% (internal) and 79.62%/81.59% (external), outperforming comparators by 5.10–9.98%.

LVLCRNet provides reliable automated AACS through integrated anatomic localization and contrastive rank-aware learning. Its strong generalizability and precision in severity grading support clinical utility for cardiovascular risk stratification.

The online version contains supplementary material available at 10.1186/s12880-025-02072-7.

Figures