Automated Opportunistic Detection of Abdominal Aortic Aneurysm on Noncontrast Lumbar Spine CT

Background: Commercial artificial intelligence tools can detect or quantify abdominal aortic aneurysm (AAA) on CT angiography and routine abdominal CT. However, their applicability to noncontrast lumbar spine CT, a high-volume examination with limited anatomic coverage and protocol-specific reconstruction, remains uncertain.

Purpose: To develop and validate a fully automated pipeline for aortic segmentation, maximal diameter measurement, and AAA detection specifically on noncontrast lumbar spine CT.

Methods: This retrospective study included 280 noncontrast lumbar spine CT examinations obtained for nonvascular indications. In 120 examinations (60 with AAA and 60 controls), the aorta was manually segmented voxel-wise and used to fine-tune an openly available nnU-Net-based segmentation model using stratified six-fold cross-validation. The six resulting models were combined in a probability-averaging ensemble and evaluated in an independent, nonoverlapping cohort of 160 examinations with radiologist-measured maximal aortic diameters (range, 15.6-64.8 mm; 85/160 at least 30 mm). Automated diameters were derived using surface extraction, centerline computation, orthogonal cross-sections, and ellipse fitting.

Results: Mean Dice similarity coefficient increased from 0.929 with the original model to 0.958 ± 0.030 after fine-tuning (117/120 examinations improved; p < 0.001); median 95th-percentile Hausdorff distance was 1.67 mm. In the independent cohort, automated measurements had a mean absolute error of 0.68 mm compared with radiologist measurements (bias, +0.04 mm; 95% limits of agreement, -2.20 to +2.28 mm; r = 0.995). At the 30-mm threshold, accuracy was 97.5%, sensitivity 97.6%, specificity 97.3%, and κ = 0.95; area under the receiver operating characteristic curve was 0.995.

Conclusion: This pipeline accurately detected AAA on noncontrast lumbar spine CT, extending validation of automated opportunistic aortic assessment to a dedicated imaging context for which published commercial-tool evidence remains limited. Its open-framework-based segmentation and explicit geometric measurement process may facilitate reproducible local implementation and further clinical validation.