GPU-Accelerated Forward and Back-Projections with Spatially Varying Kernels for 3D DIRECT TOF PET Reconstruction.

Area of Science:

• Medical Imaging
• Computer Science
• High-Performance Computing

Background:

• Iterative reconstruction methods are crucial for medical imaging, particularly in Time-of-Flight (TOF) Positron Emission Tomography (PET).
• Modeling spatially variant system response kernels presents computational challenges, impacting reconstruction speed and accuracy.
• Existing methods, like Fast Fourier Transform (FFT)-based approaches, struggle with generic, shift-variant kernels.

Purpose of the Study:

• To develop a GPU-accelerated framework for efficient modeling and projection operations using spatially variant system response kernels.
• To optimize the framework for DIRECT (Direct Image Reconstruction for TOF) iterative reconstruction.
• To address memory cache performance issues in GPU computations for non-axis aligned TOF data.

Main Methods:

• Implemented a GPU-accelerated framework leveraging specific GPU memory types and instruction-level parallelism.
• Focused on optimizing GPU memory access patterns to maximize cache performance.
• Developed forward- and back-projection operators capable of handling generic, shift-variant system response kernels.

Main Results:

• The GPU implementation achieved comparable or faster performance than state-of-the-art FFT-based methods.
• The framework successfully handled spatially symmetric and asymmetric, shift-variant kernels.
• Demonstrated efficient modeling of spatially variant system response kernels for TOF imaging.

Conclusions:

• The developed GPU framework offers a powerful and efficient solution for iterative image reconstruction in TOF systems.
• This approach overcomes limitations of FFT-based methods by accommodating complex, generic system response kernels.
• The optimization strategies enhance performance and enable advanced modeling capabilities in medical imaging reconstruction.