Related Experiment Video
Updated: May 23, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
Simultaneous partial volume correction and denoising of brain PET images, using transformers and transfer learning
Sanaz Kaviani1,2, Amirhossein Sanaat3, Mersede Mokri1,2
1Faculté de médecine, Université de Montréal, Montreal, Canada.
EJNMMI Research
|May 22, 2026
Summary
A novel deep learning (DL) method effectively corrects for partial volume effects (PVE) and noise in low-count Positron Emission Tomography (PET) brain scans. This DL-PVE approach enhances image quality and quantitative accuracy without requiring MRI data.
Area of Science:
- Medical Imaging
- Neuroscience
- Artificial Intelligence
Background:
- Positron Emission Tomography (PET) is crucial for quantitative brain imaging but is limited by image noise and partial volume effects (PVE), especially with reduced radiotracer doses or acquisition times.
- Conventional partial volume correction (PVC) methods struggle with noise and often require MRI, which is not always available or practical in clinical settings.
- Existing deep learning (DL) methods primarily focus on denoising and do not simultaneously address PVE, highlighting the need for integrated solutions.
Purpose of the Study:
- To evaluate a DL approach for generating high-quality, PVC brain PET images from low-count scans.
- To assess the DL method's ability to simultaneously mitigate noise and PVE, improving quantitative accuracy.
Main Methods:
- A DL-PVC model based on Transformer and Unet architectures was developed and trained using both phantom and clinical brain PET datasets (18F-FDG and 18F-Florbetapir).
- Transfer learning was employed to combine phantom ground truth data with real clinical PET complexities.
- The DL-PVC method was compared against low-count non-PVC PET and Iterative Yang PVC using metrics like SSIM, PSNR, and rRMSE.
Main Results:
- The DL-PVC method significantly outperformed low-count non-PVC PET and Iterative Yang PVC in both phantom and clinical datasets.
- Improvements included higher SSIM (approx. 2%), higher PSNR (approx. 11-22%), and lower rRMSE (up to 50%) compared to baseline low-count PET.
- The DL-PVC model demonstrated substantial quantitative accuracy gains for both 18F-FDG and 18F-Florbetapir tracers.
Conclusions:
- A novel DL-based PVC method was successfully developed, leveraging Transformer and Unet architectures.
- The method effectively improves PET image quality and quantitative accuracy by jointly addressing noise and PVE.
- Crucially, the DL-PVC model requires only PET input at inference, enabling its use without co-registered MRI data.