Related Experiment Video
Updated: Mar 1, 2026
10:44
Three-Dimensional Phase Resolved Functional Lung Magnetic Resonance Imaging
Published on: June 21, 2024
1.3K
Image quality restoration in 15-s breath-hold PET using a diffusion-based neural network
Jingxi Hu1,2,3, Lianghua Li4, Qiyang Zhang1,2
1Research Center for Medical AI, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China.
Medical Physics
|February 28, 2026
Summary
This study introduces TAM-DiffPET, a novel deep learning framework that significantly reduces noise in short, breath-hold PET scans. The method enhances lesion visibility and maintains quantitative accuracy for improved diagnostic potential.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiochemistry
Background:
- Breath-hold PET imaging reduces motion artifacts but suffers from noise and poor lesion detection due to short acquisition times.
- Improving image quality in ultrashort PET scans is a significant technical challenge for clinical applications.
Purpose of the Study:
- To develop a deep learning framework using a diffusion probabilistic model to enhance image quality in 15-second breath-hold PET scans.
- The goal is to suppress noise and improve lesion visibility while preserving quantitative accuracy.
Main Methods:
- A denoising diffusion probabilistic model (DDPM) with Temporal Attention Modulation (TAM) was developed, named TAM-DiffPET.
- The model was trained on paired 15-s breath-hold and 5-min free-breathing PET scans from 230 patients.
- Performance was evaluated using PSNR, SSIM, and SUV error analysis, compared against U-Net, CycleGAN, and vanilla DDPM.
Main Results:
- TAM-DiffPET achieved superior performance with the highest PSNR (40.2 dB) and SSIM (0.995) compared to other deep learning methods.
- Quantitative analysis showed lower SUV error within lesion regions, and visual assessment revealed enhanced lesion contrast and reduced noise.
- Difference maps and SUV distribution analysis confirmed minimal deviation from reference scans and preserved tracer uptake consistency.
Conclusions:
- The proposed diffusion-based framework effectively denoises ultrashort breath-hold PET images, improving visual clarity and quantitative accuracy.
- This technology holds potential for clinical use in motion-prone imaging, enhancing diagnostic accuracy while reducing scan time and radiation exposure.
