Dynamic low-count PET image reconstruction using spatio-temporal primal dual network

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Dynamic positron emission tomography (PET) is crucial for monitoring physiological metabolism in clinical diagnosis and cancer treatment.
• Reconstructing dynamic PET images is challenging due to limited counts per frame, especially in ultra-short frames.
• Existing deep learning methods often overlook temporal correlations, focusing primarily on spatial aspects.

Purpose of the Study:

• To develop an advanced deep learning model for dynamic low-count PET image reconstruction.
• To address the limitations of current methods by incorporating both spatial and temporal information.
• To enhance the interpretability and physical constraints in PET image reconstruction.

Main Methods:

• Proposed the Spatio-Temporal Primal Dual Network (STPDnet), inspired by the learned primal dual (LPD) algorithm.
• Utilized 3D convolution operators to encode both spatial and temporal correlations.
• Integrated physical PET projection principles into the network's iterative learning process for enhanced interpretability and constraints.

Main Results:

• STPDnet demonstrated substantial noise reduction in both temporal and spatial domains.
• The proposed method outperformed traditional methods like Maximum Likelihood Expectation Maximization (MLEM), spatio-temporal kernel methods, LPD, and FBPnet.
• Achieved superior reconstruction performance in low-count scenarios.

Conclusions:

• STPDnet offers improved reconstruction performance for dynamic low-count PET imaging.
• The method is particularly suitable for whole-body dynamic and parametric PET imaging requiring ultra-short frames and handling high noise levels.
• This advancement holds significant potential for enhancing diagnostic capabilities in challenging PET applications.