Mathematical and numerical methods for accurate aorta segmentation from non-enhanced CT Data yielding reliable identification and evaluation of large vessel vasculitis

TL;DR

This work addresses the challenge of accurately segmenting the entire aorta from non-enhanced CT data to support large-vessel vasculitis assessment. It introduces a three-step pipeline: (i) a minimal-path path extraction inside the aorta using a novel potential that blends image intensity with edge information, (ii) a 3D Lagrangian curve evolution to center the path along the aorta, and (iii) the GSUBSURF model initialized from the centerline to obtain a precise aorta surface. The methods demonstrate robust performance across data with weak edges, achieving a worst-case mean Hausdorff distance of $2.175 \pm 0.605$ mm at a voxel spacing of $0.977$ mm, and enable ROI-based LVV evaluation when aligned with FDG-PET/CT imaging. The framework provides reproducible metrics (e.g., SUV ratios) for vasculitis diagnosis and treatment monitoring, with results that align with expert clinicians and offer a path toward semi-automatic clinical workflow improvements.

Abstract

Segmentation of the aorta is crucial for various medical analyses, such as the diagnosis and treatment of cardiovascular diseases. This work presents mathematical models and methods yielding a semi-automatic segmentation of the aorta from non-enhanced CT data. Our framework consists of three steps. First, using the minimal path approach, we extract a path within the aorta from two user-supplied points. Then, using 3D Lagrangian curve evolution, we move the initial path to the approximate centerline of the aorta. The centered path is used in the last step to construct the initial condition for the generalized subjective surface method (GSUBSURF). Applying the GSUBSURF method with this initial condition yields an accurate segmentation of the aorta. The segmentation results and the manual segmentations overlap, with a worst-case mean Hausdorff distance of $2.175 \pm 0.605$ mm for a voxel spacing of $0.977$ mm. Using the aorta centerline and segmentation, we define precise regions of interest along the aorta to assess large-vessel vasculitis from patient FDG-PET/CT image data. The application shows promising results, as we demonstrated widespread inflammation throughout the aorta in a patient before treatment. After treatment, we observed a significant reduction in inflammation while accurately identifying the aorta regions where inflammation persisted. These findings also align with those of experienced medical doctors who have worked on the same cases.

Figures (13)

• Figure 1: Aorta location in the body Kovacs_2010
• Figure 2: Paths extraction within complex 3D shape. The yellow dots are the endpoints. The red path is a wrong path due to a shortcut. The blue path is located within the structure, but it is not centered, and the green path is centered.
• Figure 3: Extracted path with missing edges near endpoints (red) and far away from the endpoints (green).
• Figure 4: The first row shows three different axial views of the CT images where the aorta is visible (red circles). The second row shows the edge image computed from the previous input slices. The third row shows the potential function corresponding to each slice. The slices are extracted from the 3D volume of Patient 1.
• Figure 5: Extracted paths for patients 4 and 5. The aorta (orange) is segmented manually for visualisation purposes. The yellow dots are the starting and final points. The green path is the expected path for patient 4 using the potential $\Tilde{P}$, and the red path is an incorrect path for patient 5 due to a shortcut.