3D path planning for robot-assisted vertebroplasty from arbitrary Bi-plane X-ray via differentiable rendering.

Area of Science:

• Medical Robotics
• Image-Guided Interventions
• Computer-Aided Surgery

Background:

• Robotic systems enhance accuracy and reduce radiation in image-guided interventions.
• Surgical path planning often requires 3D CT scans, which are not always available or cost-effective for procedures like vertebroplasty.
• Current methods for integrating 2D intraoperative images with 3D preoperative data are burdensome.

Purpose of the Study:

• To develop a CT-free 3D transpedicular path planning framework for robot-assisted vertebroplasty using bi-planar 2D X-rays.
• To integrate differentiable rendering with a Statistical Shape Model (SSM) for dynamic refinement of vertebral shape and pose.
• To enable versatile and accurate surgical planning independent of preoperative CT scans and fixed imaging geometries.

Main Methods:

• A differentiable rendering-based framework was developed for 3D path planning.
• A vertebral atlas was generated using a Statistical Shape Model (SSM).
• A learned similarity loss was employed to dynamically refine the SSM shape and pose using bi-planar 2D X-rays.

Main Results:

• The framework achieved superior reconstruction metrics compared to a normalized cross-correlation baseline (DICE: 0.75 vs. 0.65).
• Performance was comparable to the state-of-the-art ReVerteR model (DICE: 0.77) and generalized to arbitrary X-ray views.
• Bipedicular planning success rates reached 82% (synthetic) and 75% (cadaver), significantly outperforming a 2D-to-3D baseline (66% and 31%).

Conclusions:

• The proposed framework enables versatile, CT-free 3D path planning for robot-assisted vertebroplasty.
• The method accommodates diverse intraoperative imaging conditions without preoperative CT scans.
• This approach enhances the feasibility and efficiency of image-guided interventions in spinal procedures.