Automated Multi-View Planning for Endovascular Aneurysm Repair Procedures

During Endovascular aneurysm repair (EVAR) procedures, surgeons always require several views of vessel structures to accurately assess the size, shape, and location of the aneurysm, along with the surrounding vasculature. However, even expert surgeons often require multiple attempts to find a desired view, which leads to increased radiation exposure, high doses of contrast agents for patients, and time-consuming re-positioning of the C-arm. This paper introduces an automatic framework to provide optimal multi-view for the whole EVAR procedure. First, a 3D nnUNet is employed to extract geometric information and semantic information, providing accurate vascular and aneurysm segmentation as well as semantic bifurcation detection. Then, a semantic vessel tree model is built by integrating semantic information and geometric information. A local 3D plane at each critical bifurcation is fitted based on the centerlines surrounding this bifurcation, where we regard the estimated 3D local plane as a good view plane in patient physical space. Next, some 3D points are collected from these centerlines, projected onto the estimated local 3D plane, and transformed to the image domain to get the paired 2D points. Finally, based on the geometric information of the C-arm X-ray imaging device, the most informative view pose for C-arm positioning is solved via RANSAC Perspective-n-Point algorithm with the Levenberg-Marquardt optimization. Our work not only streamlines the surgical planning process, but also helps in customizing the patient-specific strategies to reduce risks and improve surgical outcomes. Our framework has been validated using an in-house dataset collected from 27 patients, which contains preoperative CTA data and intraoperative X-ray angiography images. The qualitative and quantitative results demonstrate the reliability and effectiveness of our approach. Meanwhile, our system achieved an average runtime of 6 min per patient.