Dose reduction in dynamic synaptic vesicle glycoprotein 2A PET imaging using artificial neural networks

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Low-dose positron emission tomography (PET) imaging increases noise in dynamic frames, impacting kinetic parameter estimation.
• Existing denoising methods struggle to balance noise reduction with bias in low-dose PET scans.

Purpose of the Study:

• To develop and assess a spatiotemporal denoising technique for low-dose dynamic PET imaging.
• To improve the quality of parametric images derived from reduced-count dynamic PET frames.

Main Methods:

• Integration of a cascade artificial neural network (ANN) with the highly constrained back-projection (HYPR) scheme.
• Implementation using C-UCB-J PET imaging data of the synaptic vesicle glycoprotein 2A (SV2A) in the human brain.
• Patch-based ANN trained on subject-specific data and applied across subjects, followed by HYPR for temporal information utilization.

Main Results:

• The ANN+HYPR technique achieved 80% noise reduction with -2% bias in dynamic frames.
• Significant noise reduction was observed in tracer uptake (75%, -2% bias) and distribution volume (70%, -5% bias) images.
• The method demonstrated superior noise reduction with minimal bias compared to conventional techniques in healthy volunteers and Parkinson's disease patients.

Conclusions:

• The ANN+HYPR technique effectively denoises low-dose dynamic PET imaging, enhancing parametric quantification.
• This approach offers denoising capabilities equivalent to an 11-fold dose increase for C-UCB-J SV2A PET imaging.
• The developed technique holds promise for improving diagnostic accuracy and reducing radiation exposure in PET studies.