Segmentation-assisted vessel centerline extraction from cerebral CT Angiography

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Accurate extraction of brain vessel centerlines from CT angiography (CTA) is crucial for diagnosing and treating cerebrovascular diseases like stroke.
• Challenges include complex cerebrovascular structures and variable imaging quality in CTA scans.

Purpose of the Study:

• To develop and validate a segmentation-assisted framework for enhanced accuracy and efficiency in brain vessel centerline extraction from CTA images.
• To streamline lumen segmentation generation without additional physician annotation, thereby improving centerline extraction effectiveness.

Main Methods:

• A framework integrating pre-processing (image registration, patch division), lumen segmentation generation (graph cuts, kernel regression), a dual-branch topology-aware UNet (DTUNet) with topology-aware loss (TAL), and post-processing (skeletonization, refinement).
• The DTUNet optimizes the use of annotated centerlines and generated lumen segmentation through its dual-branch structure and TAL.

Main Results:

• The framework was evaluated on an in-house dataset of 10 intracranial and 40 sub-image CTA scans.
• Demonstrated superior performance over state-of-the-art methods in average symmetric centerline distance (ASCD) and overlap (OV).
• Achieved ASCD of 0.84 and OV of 0.885 for intracranial images, and ASCD of 1.26 and OV of 0.824 for sub-images, showing promise for clinical stroke applications.

Conclusions:

• The DTUNet framework automates lumen segmentation and optimizes network design for high-performance vessel centerline extraction.
• Achieves excellent results without additional annotation demands, offering significant benefits for cerebrovascular disease management.