Dynamic positron emission tomography image reconstruction using spatiotemporal kernel method with deep image prior.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Dynamic Positron Emission Tomography (PET) imaging faces challenges in reconstructing high-quality images due to inherent ill-posedness and low photon counts.
• Deep learning, specifically the Deep Image Prior (DIP) framework, has shown promise in improving PET reconstruction without external training data.

Purpose of the Study:

• To develop and validate a novel DIP-based dynamic PET reconstruction algorithm integrated with a spatiotemporal kernel.
• To enhance the quality of reconstructed dynamic PET images, particularly under low-count conditions.

Main Methods:

• A novel dynamic PET reconstruction algorithm was developed, integrating a spatiotemporal kernel with the DIP framework.
• The reconstruction objective function was formulated as a constrained optimization problem solved using the alternating direction method of multipliers (ADMM).
• The algorithm supports list-mode reconstruction for full 3D imaging and scalability.

Main Results:

• The proposed method significantly improved image quality compared to conventional methods (MLEM, KEM, STKEM) and other deep learning methods (DIPRecon, NeuralKEM).
• It achieved higher signal-to-noise ratio (SNR) and structural similarity index (SSIM), with stable iterative reconstruction and comparable contrast recovery coefficient (CRC).
• The method demonstrated robust performance under low-count conditions, preserving image quality where conventional methods degraded.

Conclusions:

• The novel DIP-based spatiotemporal kernel method enhances dynamic PET reconstruction accuracy without external priors.
• Its modular design allows integration into existing workflows and adaptability to various PET acquisition protocols.
• The method shows potential for clinical and preclinical dynamic PET imaging, especially in low-dose or high-resolution scenarios.