High performance volume-of-intersection projectors for 3D-PET image reconstruction based on polar symmetries and SIMD...

Area of Science:

• Medical Imaging
• Computational Science
• Image Reconstruction

Background:

• Iterative 3D PET image reconstruction demands efficient forward-backward projectors for speed.
• The system response matrix (SRM) in PET involves billions of elements, causing computational challenges for algorithms like MLEM.
• Exploiting symmetries in the SRM is crucial for matrix compression and improved computational performance.

Purpose of the Study:

• To develop an efficient, symmetry-driven evaluation of forward-backward projectors for 3D PET image reconstruction.
• To overcome memory access inefficiencies and enhance numerical throughput in PET reconstruction algorithms.
• To accelerate the calculation time of iterative PET reconstruction using optimized data organization and parallel processing.

Main Methods:

• Implemented a novel symmetry-driven evaluation of forward-backward projectors, leveraging polar symmetries from the PRESTO toolkit.
• Reorganized image and line-of-response (LOR) data in memory to enable efficient single instruction multiple data (SIMD) vectorization.
• Utilized simultaneous multi-threading (SMT) to further reduce calculation time.

Main Results:

• Achieved a global speedup factor of 11 without SMT and over 100 with SMT for the CPU-based implementation.
• Demonstrated equivalent numerical results compared to previous methods.
• Overcame limitations of ray-driven implementations by improving memory access patterns and cache utilization.

Conclusions:

• The proposed symmetry-driven projector evaluation significantly accelerates 3D PET image reconstruction.
• Optimized data organization and parallel processing (SIMD, SMT) are key to achieving substantial speedups.
• This approach enhances the computational efficiency of iterative PET reconstruction, enabling faster and more accurate imaging.