GPU-accelerated 3D Bayesian image reconstruction from Compton scattered data.

Area of Science:

• Medical Imaging
• Computational Physics
• Computer Science

Background:

• Compton cameras are crucial for imaging, but reconstruction is computationally intensive.
• Traditional reconstruction methods struggle with large system matrices and slow computations.
• Accelerating projection and backprojection is key for efficient iterative reconstruction.

Purpose of the Study:

• To develop fast Bayesian reconstruction methods for Compton cameras.
• To leverage commodity graphics hardware (GPUs) for accelerated computations.
• To overcome the limitations of conventional caching schemes for large system matrices.

Main Methods:

• Proposed GPU-accelerated methods for on-the-fly conical projection and backprojection.
• Developed voxel-based conical backprojection methods using two approximation schemes.
• Approximated intersecting chord lengths and treated voxels as dimensionless points for efficiency.

Main Results:

• GPU-based methods dramatically improve computational speed.
• Reconstruction accuracy shows only a minor loss compared to CPU-based methods.
• Simulation studies validate the effectiveness of the proposed GPU approach.

Conclusions:

• GPU-accelerated Bayesian reconstruction offers a significant speedup for Compton cameras.
• The developed methods provide a practical solution for handling large system matrices.
• Future high-resolution detectors will further minimize accuracy differences between GPU and CPU methods.