A Hybrid Method for Rapid, Automated Landmarking of Lung Vasculature In Free-Breathing CT
Abstract
Purpose
Dense lung vasculature landmarking is vital for assessing thoracic registration accuracy in treatment planning. We demonstrate an automated method for identifying vascular bifurcations in free-breathing CT images.
Methods
Lungs were segmented using TotalSegmentator (500x500x395 helical CT), cropped, and eroded (r=10 voxels) to eliminate wall contamination. Images were denoised via Chambolle TV (weight=25 HU). We created a point cloud of initial vessel locations where the image intensity was > 525.0 HU within the eroded lung mask (N_Points ~50,000). We iteratively shifted each point within using a ray-traced mean-shift technique that sampled the surface of the vessel (500.0 HU) 1024 times about each point, limiting the iterations and max search distance on a per-point basis using the initial radius. We calculated the radius of the vessel at each point by finding the plane with the smallest area encompassing the point. We formed a graph with each point as a node, where an edge was formed if the points were closer to each other than either were to connected nodes. We identified nodes as endpoints (degree = 1); unfiltered bifurcations (degree > 2); and filtered bifurcations (identified during endpoint-endpoint pathfinding). Finally, we visually inspected each bifurcation location using a localized and rotated window centered on each bifurcation node.
Results
We found that the filtered bifurcations had an accuracy of 87.4% (132 out of 151) and the unfiltered bifurcations had an accuracy of 72.7% (173 our of 238). The computation time per image, including TotalSegmentator and review image generation, was < 5 minutes using an Nvidia A100 (Ampere, 40GB VRAM).
Conclusion
This method enables rapid identification of vascular bifurcation locations. Future work will focus on refining midline resampling to further improve filtered bifurcation accuracy.