An iterative reconstruction algorithm for unsupervised PET image.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Imaging

Background:

• Convolutional Neural Networks (CNNs) show promise for Positron Emission Tomography (PET) image reconstruction but require extensive training data.
• Existing methods often involve complex optimization or rely on large datasets, limiting clinical applicability.

Purpose of the Study:

• To develop a novel iterative PET image reconstruction algorithm integrating the Deep Image Prior (DIP) framework.
• To improve reconstruction quality and speed using patient-specific prior information and sinogram data.
• To overcome limitations of data-hungry CNNs in clinical settings.

Main Methods:

• Developed a constrained optimization problem using the DIP framework and patient-specific prior information (e.g., MRI).
• Integrated Nesterov's acceleration and a restart mechanism into the iterative algorithm for enhanced performance.
• Utilized existing PET image reconstruction packages to streamline calculations.

Main Results:

• The proposed algorithm demonstrated superior performance compared to MLEM-GF, KEM, and DIPRecon methods in 2D PET data experiments.
• Experiments included both computer simulations and real patient data, validating the algorithm's effectiveness.
• The method achieved improved reconstruction quality and speed.

Conclusions:

• The novel iterative algorithm leverages inter-patient information, eliminating the need for large training datasets typical of CNNs.
• Optimized iterative processes enhance reconstruction performance and efficiency.
• The algorithm's design allows for easy integration into standard PET reconstruction implementations with minimal modifications.