A novel morphological assessment approach for abdominal aortic aneurysms: automated segmentation and measurement using a semi-supervised deep learning algorithm.

Background: Morphological assessment of abdominal aortic aneurysm (AAA) is crucial for preoperative strategy formulation, but traditional manual assessment is time-consuming and subjective. Objectives: This study aims to develop a novel semi-supervised deep learning (DL) framework for the automatic and accurate segmentation and measurement of AAA. Methods: In this study, a semi-supervised learning algorithm was developed based on computed tomography angiography scans collected retrospectively from 88 AAA patients. The algorithm employed 16 labeled cases for supervised training and incorporated 47 unlabeled cases through consistency regularization to achieve accurate segmentation of the AAA lumen, aneurysm neck, and iliac arteries. To evaluate the model's segmentation performance, the Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95) were utilized, and the agreement between automatically derived measurements and manual annotations was assessed. Results: Among the models evaluated, Swin-UNETR achieved the best segmentation performance with 97.3% accuracy. The DSC was 0.965 ± 0.025 and 0.931 ± 0.021, and HD95 was 2.01 ± 0.39 mm and 2.93 ± 0.28 mm in the internal and independent test sets, respectively. Automation and expert measurements showed generally good-to-excellent agreement based on intraclass correlation coefficient (ICC) analysis. In addition, average time per case decreased from 53.08 ± 8.17 to 4.18 ± 0.90 min, mouse movement distance from 91.62 ± 9.95 to 13.03 ± 3.12 m, and clicks from 672.62 ± 129.47 to 157.69 ± 0.90 times (all Conclusions: The proposed Swin-UNETR algorithm enables fast, reliable, and reproducible automated segmentation and measurement of key anatomical details of AAA.