Deep learning for segmentation and endoleak detection in contrast-enhanced computed tomography in endovascular aortic repair patients

Abstract Background Endovascular aortic repair (EVAR) is the most common technique for repairing abdominal aortic aneurysms (AAA), achieving low procedure-related complication and excellent patients’ recovery. However, lifelong imaging surveillance is needed to monitor possible complications following EVAR, such as endoleaks 1. Purpose To develop and evaluate deep learning (DL) models for (i) aortic segmentation in pre- (PreI) and (ii) post-intervention (PostI) contrast-enhanced computed tomography angiographies (CT) and for (iii) endoleak (EL) detection in PostI CT images. Methods A total of 341 multi-vendor CT images (152 PreI and 189 PostI CT) from 142 EVAR patients were retrospectively extracted from two clinical centres. The aortic lumen and thrombus were manually segmented in all PreI CT images while the aortic lumen, the part of the aorta excluded by the prosthesis, and the EL (if present) were annotated in all PostI CT images by one of three experts. Moreover, two experts annotated 22 CT images to assess inter-observer variability. Three convolutional neural networks called nnU-Net 2 were trained with data from one centre for (i) PreI CT scans segmentation, (ii) aortic lumen and excluded aorta segmentation in PostI CT images, and (iii) EL detection. The training set consisted of CT images from 97 patients (97 PreI and 109 PostI CT scans). CT images from 55 patients of the two centres (55 PreI CT, 80 PostI CT, 22 with EL) were used for testing. DL-based segmentations were compared with manual segmentations using Dice Score (DS) and 95th percentile Hausdorff distance (HD). The EL detection algorithm was evaluated in terms of accuracy, specificity and sensitivity compared to manual annotations. Results Mean time for DL annotation was 2.7 and 3.1 minutes for PreI CT and PostI CT scans, respectively. Segmentation performances were excellent for each segmentation task in both preI and postI CT images of the two centres, including the one that was not used for training (Table 1, Figure 1). Moreover, the error of DL predictions was similar or lower than inter-observer variability in all tasks and centres. The EL detection algorithm achieved an accuracy of 0.875, a specificity of 0.966, and a sensitivity of 0.636 on the multi-centre testing set. Conclusion Deep learning can reduce time and variability in aortic segmentation of CT scans and EL detection, leading to faster and more reliable clinical assessments to improve risk stratification of EVAR patients.Table 1.Segmentation metrics Figure 1.Example of segmentation