Artificial intelligence for endoleak detection and anatomical annotation of pre- and post-endovascular aortic repair computed tomography angiographies

Abstract Background Endovascular aortic repair (EVAR) is widely used for abdominal aortic repair. Endoleaks (EL) are prevalent and should be detected as they may lead to adverse events or need for re-intervention. Purpose To develop and validate artificial intelligence (AI) algorithms that segment the aorta and locate anatomical landmarks in pre- and post-intervention computed tomography angiographies (PreCTA and PostCTA) and detect EL in PostCTA. Methods 446 CTA (127 PreCTA and 319 PostCTA) from 192 patients were retrospectively extracted from two clinical centres (97 patients from internal centre (C1), 95 from an external centre (C2)). Four convolutional neural networks (nnU-Net) were trained with 5-fold cross-validation configuration on data from C1 to (i) segment aortic lumen and thrombus in PreCTA, (ii) segment aortic lumen, excluded aorta, and EL in PostCTA, and (iii) locate 4 aortic landmarks (celiac trunk, superior mesenteric artery, renal bifurcation, and aortic bifurcation) in PreCTA plus 3 EVAR landmarks (proximal extent, proximal edge of fabric component, end of the prosthesis) in PostCTA. Thirty PreCTA and 40 PostCTA from C2 were used for external validation of segmentations and landmark detection, while 211 PostCTA from C2 were used for the external validation of EL detection. All annotations were done by one of three experts, and 11 PreCTA and 11 PostCTA were annotated by two for inter-observer variability assessment. Manual segmentations were used to evaluate the AI segmentations through the Dice Score (DS) and 95% Hausdorff Distance (HD). AI and manual landmark locations were compared using Euclidean distance. Results Most of patients were male (92/97 in C1, 91/95 in C2) and age was 75 69-79 years in C1 and 74 70-80 in C2. Pre-EVAR aortic diameter was 61 55-69 mm in C1 and 58 54-68 mm in C2, while 34/97 and 46/95 of patients presented EL in a total of 45 and 100 PostCTA in C1 and C2 scans, respectively. Type II EL was the most frequent (30/34 patients in C1 and 41/46 in C2), followed by type Ia (2/34 and 4/46), type Ib (2/34 and 3/46), and type III (0/34 and 3/46). In CTA from C1, AI models had excellent segmentation (Figure) and landmark detection performance, generally in-line with both inter-observer variability and state-of-the-art AI performances. The generalizability test showed that performances in data from C2 were excellent for all segmentation tasks, but landmark detection performances showed a slight, unlikely clinically significant decline (Table). EL detection achieved slightly better results in C1 (accuracy=0.87, sensitivity=0.93, specificity=0.8) than in C2 data (accuracy=0.84, sensitivity=0.76, specificity=0.91). All 4/45 EL missed in C1 data were type II, while 2 type Ia and 22 type II EL were not detected in C2. Conclusion AI enables precise aortic segmentation and EL detection in EVAR patients, offering potential as a valuable tool for patient monitoring and risk stratification.Table.AI annotation metrics Figure.Manual and AI segmentations