Fast and memory-efficient reconstruction of sparse Poisson data in listmode with non-smooth priors with application...

Area of Science:

• Medical Imaging
• Computational Science
• Physics

Background:

• Time-of-flight positron emission tomography (TOF PET) systems generate large sinogram datasets, posing memory and computational challenges for iterative reconstruction algorithms.
• Existing algorithms like stochastic primal-dual hybrid gradient (SPDHG) require substantial memory, limiting their application on state-of-the-art TOF PET systems.
• The increasing size of TOF PET data, due to improved TOF resolution and larger fields of view, exacerbates these computational limitations.

Purpose of the Study:

• To develop and analyze a novel listmode (LM) extension of the SPDHG algorithm tailored for sparse TOF PET data.
• To evaluate the performance of the proposed LM-SPDHG algorithm in terms of memory reduction and reconstruction speed.
• To enable the application of advanced iterative reconstruction techniques on high-performance computing hardware for clinical TOF PET imaging.

Main Methods:

• A new listmode (LM) extension of the stochastic primal-dual hybrid gradient (SPDHG) algorithm was developed for Poisson distributed sparse data.
• The algorithm was evaluated using realistic 2D and 3D simulations and a real data set from a state-of-the-art TOF PET/CT system.
• Performance was compared against conventional sinogram SPDHG and listmode EM-TV algorithms.

Main Results:

• The proposed LM-SPDHG algorithm demonstrated convergence speed equivalent to the conventional sinogram SPDHG.
• Significant memory reduction was observed: from ~56 GB to 0.7 GB for dynamic frames and 12.4 GB for static acquisitions.
• These memory savings facilitate pure GPU implementation, eliminating host-GPU data transfers.

Conclusions:

• The LM-SPDHG algorithm effectively addresses the memory limitations of reconstructing large TOF PET datasets.
• Reduced memory footprint enables efficient, pure GPU-based reconstruction, leading to substantially accelerated processing times.
• This advancement is crucial for the routine clinical application of advanced iterative reconstruction methods in TOF PET imaging.