Deep learning for anatomical structure identification and endoleak detection in contrast-enhanced computed tomographies in endovascular aortic repair patients

Abdominal aortic aneurysms are frequently treated using endovascular aortic repair (EVAR). A common complication is the development of endoleaks (EL), which can lead to aortic rupture if they are not properly identified.

To develop and validate deep learning (DL) models that annotate anatomical structures in pre- and post-surgical computed tomography angiographies (PreCTA and PostCTA) and detect EL in PostCTA.

CTA from 192 patients were retrospectively extracted from two centres (97 patients from internal centre (IC), 95 from an external centre (EC)). This resulted in 446 CTA (127 PreCTA and 319 PostCTA).

Using a 5-fold cross-validation strategy, four convolutional neural networks (nnU-Net) were trained on IC data to perform multiple tasks: (i) segmentation of the aortic lumen and thrombus in PreCTA, (ii) segmentation of the aortic lumen, excluded aorta, and EL in PostCTA, and (iii) identification of 4 anatomical landmarks (celiac trunk, superior mesenteric artery, renal bifurcation, and aortic bifurcation) in Pre-CTA, along with 3 graft-related landmarks (graft proximal extent, proximal edge of fabric component, and distal end) in PostCTA. For external validation, 30 PreCTA and 40 PostCTA from the EC cohort were used to assess segmentation and landmark identification performance, while an additional 211 PostCTA were employed for independent evaluation of EL detection. All image annotations were done by one of three trained experts, with a subset of 11 Pre-CTA and 11 Post-CTA scans annotated by two observers to assess inter-observer variability.

Dice Score (DS) and 95% Hausdorff Distance (HD) were used to assess the difference between DL and manual segmentations, while DL and manual landmark locations were compared using Euclidean distance.

Median age was 75 [69-79] years in IC and 74 [70-80] in EC and the majority were male (92/97 in IC, 91/95 in EC). EL were present in 34/97 patients (45/109 PostCTA) and 46/95 patients (100/211 PostCTA) from IC and EC, respectively. The most common EL was type II (30/34 patients in IC and 41/46 in EC), followed by type Ia (2/34 and 4/46), type Ib (2/34 and 3/46) and type III (0/34 and 3/46).

DL segmentation and landmark identification in IC data was similar or better than inter-observer variability (Figure). In EC data, segmentations were excellent but landmark identification showed significant decrease in quality, but probably with no clinical impact (Table).

DL model detected EL barely better in IC data (sensitivity=0.93, specificity=0.8, accuracy=0.87) than in EC (sensitivity=0.76, specificity=0.91, accuracy=0.84). Missed EL were all type II in IC, whereas there were 2 type Ia and 22 type II in EC.

DL facilitates the identification of anatomical structures and detection of EL in EVAR patients, highlighting its potential as a supportive tool for monitoring and risk assessment.Table.

DL annotation metrics