Low-Count PET Image Reconstruction With Generalized Sparsity Priors via Unrolled Deep Networks.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Deep learning shows promise for low-count Positron Emission Tomography (PET) image reconstruction.
• Existing methods often overlook PET's physical characteristics, limiting performance and interpretability.

Purpose of the Study:

• Introduce GS-Net, an unrolled deep network for enhanced PET image reconstruction.
• Improve fidelity and prior regularization by leveraging PET physics.

Main Methods:

• Utilize maximum likelihood estimation for Poisson distribution and Generalized domain transformation for Sparsity learning (GS-Net).
• Employ Alternating Direction Method of Multipliers (ADMM) with Expectation Maximization (EM) and L1 norm optimization.
• Implement end-to-end adaptive hyperparameter learning.

Main Results:

• GS-Net demonstrated superior performance over traditional and existing deep learning methods.
• Evaluations on simulated and real clinical PET datasets showed significant improvements.
• Qualitative and quantitative analyses confirmed advanced reconstruction capabilities.

Conclusions:

• GS-Net offers a novel approach to PET image reconstruction by integrating physical characteristics.
• The method achieves state-of-the-art results, enhancing diagnostic accuracy in low-count scenarios.
• Adaptive hyperparameter tuning simplifies the process and improves generalizability.