Automated quantification of abdominal aortic calcification using 3D nnU-Net: a novel approach to assess AAA rupture risk.

Abstract

Background: Abdominal aortic aneurysms (AAA) pose a serious rupture risk, heightened by aortic calcification. Traditional calcification scoring methods are slow and require expertise. This study aims to construct a convolutional neural network (nnU-Net) model for automatic quantification and segmentation of abdominal aortic calcification from a single CTA scan.

Methods: This retrospective study included 100 patients who underwent abdominal aortic CTA between January 2018 and October 2023, meeting specific inclusion criteria. Vessel and calcification segmentation were manually scored by two physicians, and an nnU-Net deep learning model was developed to automate calcification measurement. Model performance was assessed using Dice scores. Agreement between manual and model-based scoring was assessed using Spearman rank correlation and Bland-Altman analysis.

Results: The nnU-Net model achieved median Dice scores of 93.60% for blood vessels and 81.06% for calcification. Average Dice scores were 92.37 ± 4.87% for blood vessel segmentation and 81.03 ± 5.11% for calcified plaque. The model's Agatston scores correlated closely with manual scores (Spearman's ρ = 0.969), with a mean difference of -229.51 (95% limits of agreement: -6003.92 to 5544.90). The model's evaluation time was also shorter than manual scoring (112 ± 4.4 s vs. 3796 ± 6.6 s, p < 0.001).

Conclusion: The nnU-Net-based model shows potential as an automated tool for accurately segmenting and quantifying abdominal aortic calcification, offering comparable results to manual scoring with significantly reduced evaluation time. This approach may assist in more efficient assessment of AAA rupture risk, supporting clinical decision-making in patient management.