517 - Using Photon-counting CT and AI-assisted Image Registration Techniques to Improve Embolization Guidance

Purpose: Prostate artery embolization (PAE) procedures remain challenging due to complex pelvic vasculature. Photon-counting-detector CT (PCD-CT) offers ultra-high resolution virtual monoenergetic images and material maps that can improve semi-automated vessel segmentation; however, pre-procedural CT data require accurate registration with intra-procedural imaging. The aim of this study is to develop an AI-assisted, non-rigid registration framework combining pre-procedural PCD-CT with intra-procedural cone-beam CT (CBCT) and digital subtraction angiography (DSA) to improve image registration and thus vessel guidance during PAE.

Materials and Methods: Patients scheduled for PAE underwent pre-procedural CT angiography with PCD-CT (NAEOTOM Alpha). During intervention, CBCT without contrast and multiple DSA sequences were acquired (AXIOM Artis Q). Intra-procedural guidance was assisted by semi-automatically identified vessel traces on high resolution 50 keV PCD-CT data, rigidly registered and overlayed onto the live images (syngo Embolization Guidance). Subsequently and retrospectively, an AI-assisted, non-rigid registration algorithm was used to align the preprocedural Virtual-Non-Contrast (VNC) PCD-CT data to the procedural unenhanced CBCT. The transformation was then applied to the vessel tracings used for embolization guidance. These registered tracings were then forward-projected onto the 2D DSA plane for comparison with actual vessel positions. The difference between the projected traces and the DSA was used to iteratively inform the registration algorithm in the form of an additional loss function to update the model weights. The percent overlap between the projected traces and the vessels in the DSA views was used at each iteration to determine vessel registration accuracy.

Results: AI-assisted registration of VNC PCD-CT image reduced the residual 3D registration error by over 50% compared to rigid registration. The addition of 2D information from the DSA plane combined with the projected traces from 50keV PCD-CT improved the accuracy of the vessel registration from the initial 23% to 75%.

Conclusion: Leveraging the multi-energy property of high-resolution PCD-CT data paired with AI-assisted, non-rigid registration techniques provides more accurate vessel guidance for PAE procedures, particularly after incorporating 2D DSA data to focus the model’s attention on the relevant vessels.