Performing an endovascular aneurysm repair (EVAR) is only possible with intraoperative fluoroscopic imaging. Technical innovations in the past decade have led to the replacement of the mobile C-arm with modern hybrid operating rooms. With the increased imaging capabilities of the current hybrid operating room, large numbers of fluoroscopy images and digital subtraction angiography (DSA) images are generated during endovascular aortic procedures. Completion DSA is typically performed at the end of the EVAR procedure after stent graft deployment to evaluate the stent graft position and to detect possible endoleaks. DSA images contain information on the stent graft position, possible endoleaks, patency of arteries and stent-graft limbs, and blood flow dynamics.(1) We propose to analyse these images with artificial intelligence methods using deep learning and perfusion DSA parameters.(2,3) In this paper, we present fully automatic visualisation of endoleaks in completion DSA, utilizing a deep learning network. We hypothesize that intraoperative automatic endoleak visualisation during EVAR is possible with objective endoleak analysis, and can aid the physician in clinical decision making.

We performed a single-centre, experimental study with retrospective collected data to create a deep-learning based intra-operative endoleak visualization. Recent EVAR procedural imaging to treat an infrarenal aortic aneurysm was collected from the hospital picture archiving and communication system. Images were acquired on the hybrid operating room with Azurion FlexMove 7 C20 of Philips (Philips Healthcare, Best, The Netherlands) in the Amsterdam University Medical Centers, location VUmc and AMC, between May 2017 and April 2019. 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 completion DSA images of 97 patients were collected, of which 49 with an endoleak. The majority of endoleaks were type 2 (35/49). The location of each endoleak was labelled with a rectangular bounding box using MeVisLab software (MeVis Medical Solutions AG, Bremen, Germany) on the original 3D DSA images, using the z-axis as timeframe. The centrepoint and vector of the bounding box were converted into an ellipsoid 3D binary mask and transformed into a 2D binary minimum intensity projection, which was used as reference labels for the deep learning model. A two-dimensional convolutional neural network (CNN) with a U-Net architecture was trained for automatic endoleak visualisation.(4) Perfusion DSA parameters were calculated per pixel; peak density (PD), time of arrival (TOA), time to peak (TTP) and area under the curve (AUC) (Figure 1). The PD represents the maximum arterial contrast attenuation, the TOA represents the time of contrast arrival in seconds, the TTP time to contrast peak (seconds), and AUC the total contrast volume passing through the pixel. The images were converted into a four-channel 2D image per patient. The patient cohort was divided into a training set (70%), validation set (10%) and test set (20%). The CNN was trained using an Adam optimizer. As data augmentation, random horizontal flipping and rotations up to 15° were used. The model was implemented in Pytorch and Medical Open Network for Artificial Intelligence (MONAI) and networks were trained and evaluated using an NVIDIA GeForce RTX 3090 GPU. The output of the network were 2D endoleak probability heatmaps, which were projected onto the initial PD perfusion DSA image. Statistical analysis was performed by sensitivity and specificity calculation and a receiver operating characteristic (ROC)-curve was created with corresponding AUC.

It was possible to transform all completion DSA into perfusion DSA. After training, the network was able to detect endoleaks on the test set which contains completion DSA never shown before. This resulted in endoleak prediction heatmaps on each completion DSA (Figure 2). From the ROC-curve, the calculated AUC of the sensitivity and specificity was 0.84.

We developed a fully automatic endoleak visualisation method based on the completion digital subtraction angiography during EVAR, utilizing a deep learning network. This objective analysis of intraoperative information can extract detailed knowledge and aid the physician in the endovascular hybrid operating room in clinical decision making with a deep learning algorithm. Future developments will focus on the classification of endoleak types and improvement of localisation.

1. Cho H, Lee JG, Kang SJ, et al. Angiography-Based Machine Learning for Predicting Fractional Flow Reserve in Intermediate Coronary Artery Lesions. J Am Heart Assoc. Feb 19 2019;8(4):e011685. doi:10.1161/JAHA.118.011685 2. Charalambous S, Kontopodis N, Papadakis AE, Ioannou CV, Tsetis D. Perfusion Digital Subtraction Angiography: Is it Time to Step Towards Functional Imaging of Endovascular Aneurysm Repair Patients? Eur J Vasc Endovasc Surg. Nov 2021;62(5):821-822. doi:10.1016/j.ejvs.2021.07.018 3. Raffort J, Adam C, Carrier M, et al. Artificial intelligence in abdominal aortic aneurysm. J Vasc Surg. Jul 2020;72(1):321-333 e1. doi:10.1016/j.jvs.2019.12.026 4. Ronneberger O, Fischer F, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015;abs/1505.04597. Accessed Mon, 13 Aug 2018 16:46:52 +0200. http://arxiv.org/abs/1505.04597