EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Benchmarking deep-learning approaches for magnetic resonance imaging-based estimation of [18F]FP-CIT positron emission tomography

Chunfa Liu, Fan Fu, Yuxi Jin, Haizhou Liu, Jidong Han, Heng Jiang, Na Zhang, Hairong Zheng, Dong Liang, Biao Li, Zhanli Hu

Brain Network Disorders · 2026-11

DOI: 10.1016/j.bnd.2026.09.002 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Brain Network Disorders, 2026-11. Verfasst von Chunfa Liu und 10 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Shenzhen Science and Technology Innovation Commission, Guangdong Basic and Applied Basic Research Foundation, Shanghai Municipal Health Commission.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

Dopaminergic positron emission tomography (PET) provides critical molecular information for assessing Parkinson's disease (PD); however, it is limited by radiation exposure, tracer logistics, and scanner accessibility. This study investigates if [ 18 F] N-(3-fluoropropyl)-2β-carboxymethoxy-3β-(4-iodophenyl) nortropane (FP-CIT) PET, which is a tracer with highly localized striatal uptake, can be synthesized from structural magnetic resonance imaging (MRI) using representative deep-learning frameworks.

Methods

A hybrid PET/MRI cohort of 385 participants (210 PD and 175 non-PD [ND]) is used to benchmark five models, namely Pix2Pix, CycleGAN, 2.5D U-Net, denoising diffusion probabilistic model (DDPM), and ControlNet, using a unified preprocessing and evaluation pipeline. Image fidelity in the normalized intensity domain, agreement of regional values after standardized uptake value (SUV) range remapping, and exploratory PD/ND group separability were analyzed comprehensively.

Results

2.5D U-Net achieved the best global image similarity performance across the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean absolute error (MAE), and normalized root mean square error (NRMSE) in the normalized intensity domain. DDPM showed the strongest subject-level agreement of region of interest (ROI) values after the SUV range was remapped, whereas ControlNet provided competitive intermediate performance. The generative adversarial networks (GANs)-based methods underperformed with higher noise and artifacts. The synthesized images retained partial ratio-based PD/ND group separability within the study cohort, suggesting that MRI-derived models can capture dopamine transporter (DAT)-related patterns for methodological benchmarking and exploratory analyses.

Conclusion

This benchmark establishes reproducible baselines for MRI-to-PET synthesis and highlights the key architectural tradeoffs in dopaminergic imaging.

Förderung

  • Shenzhen Science and Technology Innovation Commission
  • Guangdong Basic and Applied Basic Research Foundation
  • Shanghai Municipal Health Commission
  • National Natural Science Foundation of China
  • Ministry of Science and Technology of the People's Republic of China

Zitieren

APA

Liu, C., Fu, F., Jin, Y., Liu, H., Han, J., Jiang, H., Zhang, N., Zheng, H., Liang, D., Li, B., & Hu, Z. (2026). Benchmarking deep-learning approaches for magnetic resonance imaging-based estimation of [18F]FP-CIT positron emission tomography. Brain Network Disorders. https://doi.org/10.1016/j.bnd.2026.09.002

Vancouver

Liu C, Fu F, Jin Y, Liu H, Han J, Jiang H, et al. Benchmarking deep-learning approaches for magnetic resonance imaging-based estimation of [18F]FP-CIT positron emission tomography. Brain Network Disorders. 2026. doi:10.1016/j.bnd.2026.09.002

BibTeX
@article{Liu2026,
  author = {Liu, Chunfa and Fu, Fan and Jin, Yuxi and Liu, Haizhou and Han, Jidong and Jiang, Heng and Zhang, Na and Zheng, Hairong and Liang, Dong and Li, Biao and Hu, Zhanli},
  title = {Benchmarking deep-learning approaches for magnetic resonance imaging-based estimation of [18F]FP-CIT positron emission tomography},
  journal = {Brain Network Disorders},
  year = {2026},
  doi = {10.1016/j.bnd.2026.09.002}
}
Kostenlos starten →