Angiogram-Free Intracranial Vessel Localization for Plaque Assessment
W Fan1*, Y Sang1, H Zhou1, Z Hu2, J Xiao2, Z Fan2,3, D Ruan1,4, (1) UCLA, Los Angeles, CA, (2) Cedars-Sinai Medical Center, Los Angeles, CA, (3) Keck School of Medicine of USC, Los Angeles, CA, (4) UCLA School of Medicine, Los Angeles, CA
Presentations
SU-IePD-TRACK 3-2 (Sunday, 7/25/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]
Purpose: Inspection of vessel morphology is important in assessing cardiovascular risks. Magnetic resonance (MR) vessel wall imaging (VWI) is routinely used with a companion MR angiogram (MRA) that localizes the vessel segments of interest. However, the acquisition of MRA prolongs scanning time, extra cost in clinical workflow, and is also subject to slight displacement with respect to the VWI. This study aims to investigate the feasibility to infer the vessel location directly from VWI.
Methods: The (VWI, MRA) pair of images from 30 patients were included; and were randomly split into 25 cases for training doubled as atlas, and 5 cases for testing. To regulate vessel topology while preserving the target’s specific appearance, such as plaque presentation, a novel atlas-based pipeline with deep correction was proposed. For each target VWI, all atlas VWI were first rigidly registered and the one with the highest mutual information was chosen to proceed to be deformable, generating an initial deformation vector field (DVF). An unsupervised error correction network was further introduced to improve the DVF, particularly in alignment and conformality to vessel defined via training MRA. The end-to-end inference applied the refined DVF to the reference MRA for final vessel extraction and localization.
Results: The 90% Hausdorff distance on the vessel segment of interest is around 10mm, which is satisfactory for localization purposes. Close inspection shows that major vessels such as the internal carotid artery (ICA) agree well with the benchmark MRA, while vertebral arteries and basilar arteries showed inferior local match.
Conclusion: Our method showed significant improvement, particularly in vessel structure integrity compared to the commonly used deep segmentation or deep synthesis. We are working on regulating the error correction network to better account for DVF heterogeneity and focal vessel performance.
ePosters
Keywords
Cerebral Vasculature, MRI, Blood Vessels
Taxonomy
IM- MRI : Machine learning, computer vision
Contact Email