Deep unrolled primal dual network for TOF-PET list-mode image reconstruction.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Time-of-flight (TOF) information enhances positron emission tomography (PET) image quality and noise reduction.
• List-mode reconstruction effectively utilizes TOF data, but advanced algorithms struggle with low-count datasets.
• Deep learning shows promise for PET reconstruction, yet TOF integration presents storage challenges for deep unrolled methods.

Purpose of the Study:

• To introduce LMPDnet, a novel deep unrolled primal dual network for TOF-PET list-mode reconstruction.
• To address the limitations of current TOF-PET reconstruction algorithms, particularly with low-count data and deep learning storage demands.

Main Methods:

• Proposed LMPDnet, a deep unrolled primal dual network with multiple phases.
• Each phase incorporates a dual network for list-mode updates and a primal network for image updates.
• Utilized CUDA for parallel acceleration and efficient system matrix computation for TOF list-mode data.

Main Results:

• LMPDnet demonstrated superior noise suppression and image quality compared to conventional methods (OSEM, TV-LREM, SPDHG, TV-SPDHG) and HistoCNN-2D.
• Performance improvements were observed across various TOF resolutions and data count levels in both visual and quantitative analyses.
• The proposed method significantly enhances reconstructed image quality for TOF-PET list-mode data.

Conclusions:

• Deep unrolled methods show significant potential for TOF-PET list-mode data reconstruction.
• LMPDnet outperforms current mainstream TOF-PET list-mode reconstruction algorithms.
• This work offers new insights into applying deep learning for TOF list-mode data.