{
  "abstract": "Introduction Patients with cerebrovascular conditions undergo CT-angiography (CTA) as part of their initial workup. Despite major advances in imaging technology, automated identification and anatomical labeling of individual brain vessels in the Circle of Willis (CoW) from clinical CTA remains a difficult problem. One-dimensional centerline networks offer a promising approach, preserving the essential geometry and connectivity of the vasculature while being far simpler to work with than traditional 3D models. These centerline networks can be directly integrated into modern machine learning and simulation tools while avoiding the computational demand of 3D models.Methods Patients who had suffered aneurysmal subarachnoid hemorrhage and undergone CTA and serial transcranial Doppler monitoring were selected for inclusion in this work. 3D models of the CoW were generated via manual segmentation of CTAs. One-dimensional centerline networks were then extracted from each segmented vessel volume using a custom Python pipeline leveraging VMTK algorithm. Each vessel segment was manually labeled using a standardized naming system, and key measurements, including vessel length, radius, curvature, and how each segment connected to its neighbors, were extracted from the segmentations to build a structured one-dimensional network for each patient. A two-stage machine learning model was then developed, in which a graphical neural network first learned the structural relationships within each vascular network, and a gradient boosting decision tree used those learned features to predict the identity of each vessel segment.Results Twenty patient-specific one-dimensional vascular networks were successfully generated. The cohort included anatomical variants, including unilateral A1 and fetal posterior communicating artery, all of which were represented within the network structure. Vessel measurements including length, radius, branching order, and connectivity were encoded across all networks. The trained model correctly labeled CoW vessel segments in every patient within the cross validated training dataset.Conclusions Patient-specific one-dimensional centerline networks of the cerebral vasculature are a useful dataset for training graph-based machine learning models. This pipeline, from CT-angiogram to structured network to automated vessel labeling, shows feasibility in a clinically complex patient population with significant anatomical variability. Validation on independent patient data remains the essential next step toward building tools that can meaningfully support clinical care and cerebrovascular research.Disclosures L. Basner: None. Z. Abecassis: None.",
  "authors": [
    {
      "affiliations": [
        "School of Medicine, University of Washington, Seattle, WA"
      ],
      "name": "L Basner"
    },
    {
      "affiliations": [
        "Neurological Surgery, University of Washington, Seattle, WA"
      ],
      "name": "Z Abecassis"
    },
    {
      "affiliations": [
        "University of Pennsylvania, Philadelphia, PA"
      ],
      "name": "M Fung"
    },
    {
      "affiliations": [
        "Department of Neurological Surgery, Georgetown University School of Medicine, Washington, DC"
      ],
      "name": "JK Lim"
    },
    {
      "affiliations": [
        "University of Washington, Seattle, WA"
      ],
      "name": "M Levitt"
    },
    {
      "affiliations": [
        "Neurological Surgery, University of Washington, Seattle, WA"
      ],
      "name": "P Fillingham"
    }
  ],
  "title": "E-056 Automated vessel labeling of the circle of willis using one-dimensional centerline networks",
  "uid": "d9be4244-8ee3-51d1-8f5d-10002c3e8e22"
}