Fast and memory-efficient Monte Carlo-based image reconstruction for whole-body PET

Area of Science:

• Medical Imaging
• Computational Physics

Background:

• Accurate system modeling in Positron Emission Tomography (PET) is crucial for image quality.
• Monte Carlo (MC) techniques offer high accuracy but pose significant computational and memory challenges for system matrix calculation and storage in whole-body PET scanners.

Purpose of the Study:

• To develop and evaluate a simulation-reconstruction framework to address the computational cost and memory requirements of MC-based system matrix modeling for the Philips Gemini GS PET scanner.
• To improve image reconstruction quality by optimizing the system matrix calculation and storage.

Main Methods:

• Developed a fast, realistic system matrix simulation module using egs_pet (based on EGSnrc).
• Implemented a rotator-based ordered subset expectation maximization (OS-EM) algorithm leveraging the cylindrical symmetry of the PET scanner.
• Utilized sparse storage techniques to compress the system matrix.

Main Results:

• System matrix simulation completed in five days on 50 cores, yielding a 2.01 GB matrix storable in standard PC memory.
• Demonstrated considerable improvement in image quality, particularly in contrast-noise trade-offs, compared to standard OS-EM.
• Significantly enhanced contrast recovery coefficients for small hot and cold spheres.

Conclusions:

• The proposed framework effectively improves image quality for the Philips Gemini GS PET scanner.
• The methodology is adaptable to other whole-body and preclinical PET scanners with similar geometries.