Generation of Synthetic brain PET images of synaptic density from MRI and FDG-PET using a Multi-stage U-Net

Synaptic vesicle glycoprotein 2A (SV2A) in human brains is an important biomarker of synaptic loss associated with several disorders, including Alzheimer’s disease (AD). However, SV2A tracers, such as [$\left.{ }^{11} \mathrm{C}\right]$ UCB-J, are less available in practice due to constrains that include cost and the need for an onsite cyclotron. Therefore, we investigated the possibility of using a 3D Multi-stage U-net (MS U-net) to generate synthetic [$\left.{ }^{11} \mathrm{C}\right]$ UCB-J SV2A PET images based on T1-weighted Magnetic Resonance Imaging (MRI) and/or [${ }^{18} \mathrm{~F}$]FDG PET imaging. In our study, we used paired data from 54 participants including 22 healthy controls (HC) and 32 participants with AD. Our results demonstrate that the MS U-net was able to generate desirable SV2A PET images that are visually similar to true images in terms of contrast and tracer distribution, quantitatively with low bias and high similarity. Thus, synthetic SV2A PET images can provide additional pathological insights into AD and other neurological disorders without increasing radiation exposure and cost. Our experiments show that through progressive learning MS U-net achieved less bias and outperformed U-net. Adopting residual connection, cross-stage feature fusion and supervised attention strategies contributes to better gradient flow, mitigates information loss and enables selective focus, leading to better generation results. Additionally, combining MRI and FDG PET images as the network input provides the best prediction performance, whereas having FDG PET alone produces better synthetic [
¹¹
C]UCB-J SV2A PET image than using MRI alone. We also evaluated our model performance for 18 typical brain regions and MS U-net achieved less than 5% bias in almost all regions. Furthermore, the differences between AD and HC in certain regions (e.g., hippocampus) are still significant in generated images. We conclude it is feasible to produce robust SV2A PET images for HC and AD from MRI and FDG PET.