LM-SPD-Net: list-mode TOF-PET image reconstruction using stochastic primal-dual network

Incorporating Time-of-Flight (TOF) information into Positron Emission Tomography (PET) imaging has demonstrated a notable enhancement in the quality of reconstructed PET images while effectively reducing noise. Nonetheless, achieving this enhancement poses challenges due to the high memory demands of model-based deep learning reconstruction techniques. To tackle this problem, we propose an innovative deep learning approach grounded in a model-based framework, LM-SPD-Net, specifically developed for list-mode TOF-PET reconstruction. LM-SPD-Net builds upon the iterative SPDHG algorithm by replacing traditional proximal operators with deep neural networks and incorporating the projection process into the training procedure. This approach enables the learned operators to achieve high robustness and generalization for reconstruction, while also enhancing the interpretability of the method. Experimental findings reveal that the suggested method not only surpasses the performance of current advanced TOF-PET list-mode reconstruction methods but also achieves further reductions in both spatial and temporal resource consumption during reconstruction. This advancement paves a new avenue for the application of deep learning technology in TOF list-mode data, highlighting its potential to significantly improve PET imaging efficiency and accuracy. Our findings suggest that LM-SPD-Net offers a novel approach to the challenges faced in TOF-PET imaging.