Perfusion Digital Subtraction Angiography for Artificial Intelligence-Based Endoleak Visualisation During Endovascular Aneurysm Repair

Objective: Contemporary endovascular hybrid operating rooms produce an extensive amount of medical images during procedures. The advanced imaging capabilities of modern hybrid operating rooms lead to the generation of numerous fluoroscopy and digital subtraction angiography (DSA) images during endovascular aneurysm repair (EVAR). Nevertheless, intra-operative clinical decision making largely relies on the visual inspection of images by the operating team. Despite the acquisition of hundreds of images during a typical intervention, a significant portion remains unused, even though they may offer valuable insights to enhance procedural outcomes. The completion DSA is performed after stent graft deployment and reveals details about the stent graft position, potential endoleaks, arterial patency, stent graft limb status, and blood flow dynamics.1 To harness the potentially concealed and valuable haemodynamic information within these images for improved procedural outcomes, a detailed analysis by converting the regular DSA to perfusion DSA is suggested and thorough analysis with artificial intelligence (AI) based deep learning techniques.2, 3, 4 In this study, a fully automated endoleak visualisation process, employing a deep learning network based on perfusion DSA imaging obtained from real world EVAR procedures, is introduced.

Methods: A multicentre, experimental diagnostic accuracy study with retrospective collected data was performed. The procedural X-ray images of the hybrid operating rooms of Amsterdam UMC location AMC and Amsterdam UMC location VUmc were collected of 220 patients in total. The completion DSA was extracted from the procedural images and reviewed by an expert panel consisting of two vascular surgeons and two interventional radiologists on endoleak presence and type (1, 2, 3, 4, or unknown). The location of each endoleak was labelled with a rectangular bounding box in each DSA frame. Perfusion DSA parameters were calculated based on time–density curves; peak density (PD), time to peak (TTP), and area under the curve (AUC) (Fig. 1). A two dimensional convolutional neural network (CNN) with modified U-Net architecture was trained with multitask learning (regression and classification) and data augmentation. After fivefold cross-validation, the AUC)and f1 score were calculated. Statistical analysis was performed by sensitivity and specificity calculation.

Results: The calculated AUC for endoleak detection was 0.85 (Fig. 2) and the f1 score for endoleak localisation was 0.62. After training, the algorithm was able to detect and localise endoleaks on the test dataset, which contains images never shown before. This resulted in endoleak prediction heatmaps for each completion DSA, with corresponding bounding boxes for endoleak detection and localisation.

Conclusion: A fully automated AI based endoleak visualisation method was developed, based on completion perfusion DSA imaging of real world EVAR procedures. The method demonstrates a high detection rate and moderate localisation rate, providing objective, and detailed imaging information that can be of significant assistance in clinical decision making during intra-operative procedures.