Generating synthetic brain PET images of synaptic density based on MR T1 images using deep learning

Area of Science:

• Neuroimaging
• Medical Physics
• Radiochemistry

Background:

• Synaptic vesicle glycoprotein 2 A (SV2A) is a key biomarker for synaptic loss in neurological disorders.
• Current SV2A PET tracers like [11C]UCB-J face accessibility challenges due to cost and radiation exposure.

Purpose of the Study:

• To develop a method for generating synthetic [11C]UCB-J PET images from MRI data.
• To enable wider research into neurological disorders by improving SV2A PET accessibility.

Main Methods:

• A 3D encoder-decoder model was implemented for image synthesis.
• The model was trained on 160 participants' MRI and [11C]UCB-J PET data.
• Image quality was assessed using metrics like MSE, SSIM, bias, and correlation, with ROI analysis.

Main Results:

• Synthetic SV2A PET images closely matched ground truth in visual appearance and quantitative metrics (bias < 2%, similarity > 0.9).
• The model achieved less than 5% average bias across various diagnostic groups and brain regions.
• The method also improved image quality by reducing noise compared to low-dose scans.

Conclusions:

• Generating accurate SV2A PET images from MRI is feasible using a data-driven approach.
• This technique offers a promising solution for overcoming limitations of conventional SV2A PET imaging.