Intraoperative 2D/3D Image Registration via Differentiable X-ray Rendering.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Navigation

Background:

• Accurate 2D/3D registration is crucial for surgical decisions, aligning intraoperative 2D images with preoperative 3D scans.
• Existing methods like conventional optimization and supervised neural networks have limitations in speed, accuracy, and data requirements.

Purpose of the Study:

• To develop a self-supervised 2D/3D registration method that overcomes the limitations of current approaches.
• To achieve accurate and rapid registration without manual landmark supervision, improving intraoperative guidance.

Main Methods:

• Introduced DiffPose, a self-supervised approach utilizing patient-specific simulation and differentiable physics-based rendering.
• Employed a CNN for initial pose regression from synthetic X-rays and refined it with differentiable rendering.
• Incorporated novel methods for camera pose sampling in SE(3), sparse differentiable rendering, and tangent space optimization in se(3) using specialized losses.

Main Results:

• DiffPose achieved sub-millimeter accuracy on surgical datasets at intraoperative speeds.
• Demonstrated significant improvement over existing unsupervised methods, outperforming them by an order of magnitude.
• Outperformed even supervised baseline methods, highlighting its effectiveness.

Conclusions:

• DiffPose offers a robust and efficient solution for 2D/3D registration in surgical navigation.
• The self-supervised, simulation-based approach eliminates the need for manual annotations, making it more practical.
• This method has the potential to enhance surgical accuracy and safety through improved intraoperative guidance.