Joint Reconstruction of Dynamic PET Activity and Kinetic Parametric Images Using Learnable Descent Framework

Positron Emission Tomography(PET) is a sophisticated imaging technique used in nuclear medicine. Dynamic PET images provide more quantitative information compared to single-frame static PET images, playing an indispensable role in scientific research and clinical diagnosis. However, the limited photon counts in each frame measurement data continuously challenge the quality of dynamic PET imaging. At this point, due to the insufficient photon counts in a single-frame dataset that fails to effectively reflect statistical characteristics, methods that individually reconstruct each frame using statistical-based reconstruction algorithms often result in inaccurate outcomes. To address this issue, this paper proposes a system framework that simultaneously achieves PET image reconstruction and dynamic parameter estimation. The algorithm, based on the foundation of combining deep learning to construct a learnable network regularization term in the image domain, incorporates prior information from the temporal axis derived from the tracer's dynamic model. This enhances the quality of PET reconstruction images and directly estimates dynamic parameters. Moreover, the proposed algorithm's iterative process combines the advantages of the physical principles in PET imaging and the rapid reconstruction speed of model-based approaches, demonstrating certain advantages with respect to generalization.