PolyPose: Deformable 2D/3D Registration via Polyrigid Transformations

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Intraoperative 3D patient pose determination from 2D X-rays is crucial for interventional procedures.
• Preoperative volumetric imaging (CT/MRI) offers precise 3D data but is unavailable during surgery.
• Existing 2D/3D registration methods struggle with sparse or limited-view X-ray data.

Purpose of the Study:

• To develop a robust and simple method for deformable 2D/3D registration using limited X-ray views.
• To enable integration of preoperative volumetric data into intraoperative guidance.
• To overcome limitations of current registration techniques in sparse-view settings.

Main Methods:

• Introduced PolyPose, a deformable 2D/3D registration method.
• Parameterized 3D deformation fields as compositions of rigid transforms (polyrigid formulation).
• Leveraged the biological constraint that bones are piecewise-rigid during movement, enforcing anatomically plausible priors.

Main Results:

• PolyPose successfully aligned preoperative volumes to as few as two X-rays.
• Demonstrated robust performance on diverse orthopedic surgery and radiotherapy datasets.
• Outperformed existing methods in challenging sparse-view and limited-angle scenarios.

Conclusions:

• The polyrigid formulation provides a strong inductive bias for accurate 3D pose estimation from limited 2D data.
• PolyPose eliminates the need for computationally expensive, patient-specific deformation regularizers.
• This method offers crucial 3D guidance for intraoperative interventions where traditional methods fail.