Poster T18 - Unlocking Aneurysm Growth Dynamics: Deep Learning and Volumetric Vasculature Modeling for 3D Growth Mapping Across Longitudinal Medical Imaging

Introduction: Aortic aneurysms, characterized by the weakening and ballooning of the aortic wall, pose a significant health risk due to their potential to rupture, leading to life-threatening internal bleeding. By closely monitoring the growth in aortic aneurysms, clinicians can make informed decisions about the timing of interventions, such as surgical repair, to mitigate the risks associated with this potentially catastrophic condition. Current clinical routines to assess aortic dilatations from medical images are typically manual or semi-manual, and focused only on the maximum aneurysm diameter. This laborious procedure, entailing measurements by a trained radiologist at predetermined locations across multiple scans, is susceptible to inherent human errors, operator bias, and the common challenge of miscommunication between physicians/surgeons and radiologists concerning the precise location for assessing the growth. We introduce a novel, automated, and 3D tool designed to extract digital twins from computed tomography angiographies (CTAs) and reproducibly compute the relative surface growth rate (in mm/year) for the entire aortic wall, providing reliable quantification of dilatation with the ability to predict the most probable locations of rupture.

Materials and

Methods: We used a clinical cohort of 100+ CTAs for 50+ patients, with at least 2 longitudinal imaging available, including both full body and chest CTA. An optimized 3D U-Net was developed to effectively segment full aorta, and the results were meticulously cleaned and leveraged to construct a single closed VTK (The Visualization Toolkit) mesh representing the digital twin of aorta. To ensure accurate alignment across CTAs obtained at different timestamps for each patient, automated co-registration of CTAs was performed using ITK (The Insight Toolkit) Elastix B-Spline non-linear registration. For each aorta mesh, the centerline was then extracted and the distance to the centerline at every surface point was calculated. Extracting point-to-point correspondence of surface points from newer and older aorta models includes (Figure 1A): (1) Identifying for each surface point pnew in newer aorta, its corresponding centerline point cnew, (2) Transforming cnew to older aorta as cnew_transformed and finding the corresponding centerline point cold, (3) Transforming pnew to older aorta as pnew_transformed, (4) Using cold as start point and pnew_transformed as direction point of a ray, performing parallel surface ray casting using OBB (oriented bounding boxes) Tree, and finding the intersection point with older aorta surface pold, and (5) Comparing the previously calculated distance to the centerline of pnew and pold, calculating surface growth rate.

Results, Conclusions, and Discussions: Results and

Discussion: We successfully developed and validated a reliable pipeline for 3D mapping of aortic growth rate in a clinical cohort (Figure 1B). The color code applied to the surface of aorta indicates the level of dilatations, with darker shades of red representing severer dilatations. Due to the nature of the tube-like model, the data closer to the aortic model end points are less reliable (negative growth rate), which can be simply filtered out as they are not clinically relevant. We applied multiple optimization and parallel processing techniques to reduce the computation speed to < 6 mins. The pipeline has been tested on the entire cohort, and being successfully applied to a larger population with no limitation. Our results were not only consistent with the visual assessments of doctors from CTAs in terms of the locations of dilatation, but also provide precise quantification and high accuracy.

Conclusions: We've leveraged recent AI and modeling advancements to generate accurate digital twins of the aorta from CTAs with varying ROIs. This includes automatic meshing, registration, and point-to-point matching for precise 3D map extraction of growth rates over time. Our tool has potential for seamless integration into imaging packages for real-time aorta morphological mapping, offering invaluable assistance to clinicians. Reduced processing time allows for large-scale studies on aorta morphological evolution, aiding research on Aortopathy, aging, and sex dimorphism. It also empowers clinicians to predict rupture risk, enhancing treatment decision-making.

Acknowledgements (Optional):