RANSAC-based global 3DUS to CT/MR rigid registration using liver surface and vessels

Area of Science:

• Medical Imaging
• Image Registration
• Computational Anatomy

Background:

• Fusion imaging combines ultrasound (US) with CT/MR for enhanced visualization.
• Accurate initial registration is crucial but challenging due to variable US probe positions.
• Current methods lack a standardized approach for automatic registration across all US sweep angles.

Purpose of the Study:

• To develop an automatic, position-independent global rigid 3D-3D registration technique for liver fusion imaging.
• To overcome limitations of manual registration and improve workflow efficiency.
• To create a robust method adaptable to various ultrasound acquisition geometries.

Main Methods:

• Utilized deep learning for segmenting liver and vessels (portal, hepatic veins) in US and CT/MR images.
• Extracted point clouds of liver surface and vessel centerlines for landmark identification.
• Employed RANSAC-based point cloud registration with incorporated constraints for speed and robustness.

Main Results:

• Quantitative evaluation on 80 patients demonstrated registration accuracy across intercostal, subcostal, and epigastric US views.
• Achieved mean registration errors of 7.3±3.2 mm (intercostal), 9.3±3.7 mm (subcostal), and 8.4±3.9 mm (epigastric).
• The method proved effective regardless of the ultrasound sweep position.

Conclusions:

• The proposed global rigid registration technique automates complex manual registration for liver fusion imaging.
• Significantly enhanced workflow efficiency for physicians and sonographers.
• Provides a reliable and automated solution for pre-procedural planning and intra-procedural guidance.