LM-LDA: Learned Descent Algorithm for List-Mode TOF-PET Image Reconstruction

Incorporating Time-of-Flight (TOF) information into Positron Emission Tomography (PET) imaging has been demonstrated to enhance reconstructed image quality and effectively suppress noise. List-mode reconstruction, known for its capacity to efficiently manage TOF data, provides significant benefits. Nevertheless, integrating TOF information poses substantial challenges due to the considerable memory requirements. To address this constraint, we introduce LM-LDA, an innovative model-driven deep learning framework designed to minimize network parameter requirements, specifically designed for 3D TOF-PET image reconstruction from list-mode data. In particular, we introduce a proximal gradient method using a learnable <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$l_{2.1} \text{-norm}$</tex> tailored to the reconstruction task. Experimental evaluations reveal that our approach surpasses existing mainstream list-mode reconstruction algorithms for TOF-PET while significantly reducing network parameter usage. This breakthrough paves the way for utilizing deep learning to process list-mode data more effectively, offering improved efficiency and accuracy in TOF-PET imaging.