Deep Image Prior-Incorporated Direct PET Parametric Image Reconstruction.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Computational Imaging

Background:

• Direct parametric reconstruction algorithms aim to enhance statistical reliability in dynamic PET imaging.
• Estimates from these algorithms are often degraded by noise, impacting accuracy.
• Noise arises from measurement errors and propagation during the reconstruction process.

Purpose of the Study:

• To develop a novel deep image prior (DIP) regularized direct reconstruction method for dynamic PET imaging.
• To leverage DIP networks for improved representation and noise reduction in parametric image estimation.
• To evaluate the method's performance in reconstructing K₁ from myocardial perfusion imaging data.

Main Methods:

• A deep image prior (DIP) network was employed to represent parametric images during direct reconstruction.
• The DIP network was initialized with pre-trained weights and updated during reconstruction to learn intermediate information.
• The method was applied to simulated and patient ⁸²Rb dynamic PET myocardial perfusion imaging data.

Main Results:

• The DIP-regularized method demonstrated superior noise versus bias/mean performance compared to traditional methods.
• Performance improvements were observed against indirect, direct reconstruction, and other regularization techniques (quadratic smoothness, dictionary learning, FCNN).
• The nonlinear representation capability of the DIP network was key to achieving enhanced reconstruction quality.

Conclusions:

• The proposed DIP-regularized direct reconstruction method shows significant potential for improving dynamic PET imaging precision.
• Enhanced precision in PET measurements can lead to improved diagnostic accuracy.
• The method offers a promising tool for more reliable disease monitoring using dynamic PET data.