Endovascular Aneurysm Repair (EVAR) has seen a transformation from the use of the mobile fluoroscopic C-arms to modern high-tech hybrid operating rooms as a result of technical innovations. With increased technical possibilities in the contemporary hybrid operating rooms, intraoperative clinical decision-making remains predominantly based on visual inspection of images by the operating team. Artificial intelligence (AI) techniques, and particularly deep learning, can be used to analyse and fully exploit information in the acquired intraoperative images at no additional burden to the operating team. In terms of clinical implementation, AI has the potential to function as a supportive tool in improving the interobserver agreement and enhancing clinical decision-making in the hybrid operating room. Completion DSAs performed at the end of the EVAR procedure contain information of the stent graft's position and patency, potential endoleaks and blood flow dynamics. We propose to use deep learning to develop objective evaluation methods for stent graft deployment, position, and endoleak recognition on the completion angiogram to improve procedural outcomes. In this study, we present fully automatic stent graft segmentation on completion angiographies as a key step in objective intraoperative image analysis using deep learning.

Completion digital subtraction angiographies (cDSAs) of 47 patients treated for an infrarenal abdominal aortic aneurysms using EVAR were selected from a prospectively maintained database at the Amsterdam University Medical Centers, location VUmc and AMC, Amsterdam, the Netherlands. Patients in the study cohort received a bifurcated stent graft with or without an iliac branch device or a single-limb aortoiliac stent graft. The cDSAs were acquired using a multiphase acquisition protocol on a Philips Azurion FlexMove 7 C20 Hybrid Operating Room. Single two-dimensional minimum intensity projections (MinIP) along the temporal dimension were obtained from the completion angiography series. Pixel-wise manual segmentation masks of the stent graft were created by an expert for all cDSA series and served as the ground truth during training and evaluation of the deep learning network. A two-dimensional convolutional neural network with a U-Net architecture was subsequently trained based on the MinIP for fully automatic segmentation of the implanted stent graft on the completion angiographies. Data augmentation was applied to obtain additional synthetic data to increase the generalizability of the network and consisted of random horizontal flipping and rotation of the MinIP. The performance of the trained U-Net was evaluated based on a cross-validation of the full data set using the Dice similarity coefficient, which measures the overlap of the predicted segmentation of the U-Net with the ground truth segmentation. Furthermore, the average surface distance between the predicted segmentation and ground truth was calculated, which is the average distance from all points on the boundary of the predicted segmentation to the closest point on the boundary of the ground truth segmentation.

Cross-validation of the trained U-Net for automatic stent graft segmentation resulted in an average Dice similarity coefficient of 0.957 ± 0.041 and a median Dice similarity coefficient of 0.968 (IQR: 0.950 - 0.976). The mean and median of the average surface distance across all completion angiographies were 1.266 ± 1.506 mm and 0.870 mm (IQR: 0.490 - 1.430), respectively. The U-Net was able to correctly and automatically segment bifurcated stent-grafts, aorto-uni-iliac stent grafts, and bifurcated stent grafts with iliac branch devices, irrespective of the device manufacturer.

A fully automatic method was developed for stent graft segmentation on the completion digital subtraction angiographies during EVAR, utilizing a deep learning network. Automatic stent graft segmentation is a key step towards intraoperative stent graft deployment evaluation. Furthermore, it provides the platform for further development of intraoperative analytical applications in the endovascular hybrid operating room such as enhanced endoleak visualization and image fusion correction.