Related Experiment Video
Updated: Oct 19, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
Anatomical-guided attention enhances unsupervised PET image denoising performance
Yuya Onishi1, Fumio Hashimoto1, Kibo Ote1
1Central Research Laboratory, Hamamatsu Photonics K. K., 5000 Hirakuchi, Hamakita-ku, Hamamatsu 434-8601, Japan.
Medical Image Analysis
|September 26, 2021
Summary
This study introduces an unsupervised deep learning method for denoising Positron Emission Tomography (PET) images using anatomical guidance from MRI scans. The magnetic resonance-guided deep decoder (MR-GDD) effectively reduces noise, enabling shorter scan times and lower tracer doses without compromising image quality.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiochemistry
Background:
- Supervised deep learning methods for Positron Emission Tomography (PET) image denoising require extensive paired datasets of low- and high-quality images.
- Unsupervised methods offer an alternative but often struggle to achieve comparable denoising performance without compromising image quality.
Purpose of the Study:
- To develop and evaluate an unsupervised 3D PET image denoising method that leverages anatomical information from magnetic resonance imaging (MRI).
- To improve denoising performance and preserve spatial resolution and quantitative accuracy in PET images, enabling reduced scan times and tracer doses.
Main Methods:
- Proposed a magnetic resonance-guided deep decoder (MR-GDD) network incorporating an encoder-decoder architecture and an attention mechanism.
- Utilized anatomical MR images to guide the denoising process, enhancing the extraction of spatial and semantic features from PET images.
- Validated the method using Monte Carlo simulations of [18F]fluoro-2-deoxy-D-glucose (FDG) PET data and experimental studies with preclinical ([18F]FDG, [11C]raclopride) and clinical ([18F]florbetapir) PET data.
Main Results:
- The MR-GDD method achieved superior denoising performance, evidenced by the highest peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) in simulations compared to conventional and deep learning-based methods.
- Demonstrated state-of-the-art denoising in preclinical and clinical studies, maintaining spatial resolution and quantitative accuracy even with significantly reduced PET counts (1/10th).
- Visualized the optimization process, offering insights into the unsupervised learning behavior.
Conclusions:
- The proposed unsupervised MR-GDD method offers effective PET image denoising by integrating anatomical guidance from MRI.
- This approach facilitates substantial reductions in PET scan times and tracer doses, making PET imaging more accessible and efficient.
- The MR-GDD method shows promise for widespread clinical application, improving patient comfort and reducing healthcare costs.
