Fast Auto-Differentiable Digitally Reconstructed Radiographs for Solving Inverse Problems in Intraoperative Imaging

Area of Science:

• Medical imaging
• Computer vision
• Scientific computing

Background:

• Digitally reconstructed radiographs (DRRs) are crucial for preoperative tasks like registration and 3D reconstruction.
• Intraoperative DRR generation faces challenges in real-time synthesis and gradient computation for optimization.
• Existing DRR methods are often slow for optimization due to difficulties in obtaining gradients.

Purpose of the Study:

• To create a fully differentiable DRR generator compatible with gradient-based optimization and deep learning.
• To accelerate DRR synthesis for intraoperative imaging and computer-assisted interventions.
• To improve the efficiency of slice-to-volume registration using gradient-based methods.

Main Methods:

• Reformulated Siddon's ray-tracing algorithm as vectorized tensor operations.
• Implemented the vectorized method in PyTorch, leveraging its automatic differentiation engine.
• Utilized GPU-accelerated tensor computation for high-speed rendering.

Main Results:

• Achieved rendering speeds comparable to state-of-the-art CUDA and C++ implementations.
• Demonstrated the method's utility in slice-to-volume registration.
• Simulations indicated convex loss landscapes for registration, favoring gradient-based optimization.

Conclusions:

• The proposed differentiable DRR generator enables fast, gradient-based optimization for intraoperative image guidance.
• This advancement supports the development of efficient computer vision algorithms for minimally invasive procedures.
• The implementation is publicly available for further research and application.