Registration of vascular structures using a hybrid mixture model

Area of Science:

• Medical imaging
• Computational anatomy
• Biomedical engineering

Background:

• Vascular changes due to surgery or pathology complicate anatomical image registration.
• Existing vessel registration methods are application-specific and struggle with size, shape, and missing branch variations.
• A generic approach is needed for diverse clinical applications and anatomical regions.

Purpose of the Study:

• To develop a generic vessel registration framework adaptable to various clinical applications and anatomical regions.
• To address challenges in vasculature registration, including differing vessel morphology and missing branches.
• To improve the accuracy and flexibility of image registration for anatomical structures.

Main Methods:

• A probabilistic registration framework utilizing a hybrid mixture model (HdMM+) with a refinement mechanism for missing branches.
• Vascular structures represented as 6D hybrid point sets: spatial positions (Student's t-distributions) and centerline orientations (Watson distributions).

Main Results:

• HdMM+ demonstrated significant error reduction (over [Formula: see text]) in intraoperative brain shift compensation and pulmonary vasculature monitoring.
• Outperformed state-of-the-art methods (Coherent Point Drift, Student's t-distribution mixture model) in accuracy metrics (mean surface distance, Hausdorff distance, Dice, Jaccard).
• Validated on both synthetic and patient data, showing robust performance across different applications.

Conclusions:

• The proposed HdMM+ framework effectively models complex vascular structures using hybrid representations.
• It accommodates intricate variations in vascular morphology and missing branches.
• The generic and flexible design enables broad applicability across diverse clinical scenarios.