Related Experiment Video
Updated: Jan 18, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Relevancy aware cascaded generative adversarial network for LSO-transmission image denoising in CT-less PET
Chetana Krishnan1, Mohammadreza Teimoorisichani2
1The University of Alabama at Birmingham, AL, United States of America.
Biomedical Physics & Engineering Express
|September 10, 2025
Summary
This study introduces a novel relevancy-aware Generative Adversarial Network (reGAN) for denoising medical images in low-dose PET scans without CT. The reGAN significantly enhances image quality and diagnostic reliability, offering a viable solution for quantitative accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Image Reconstruction
Background:
- High-quality PET imaging is challenged by low radiation doses and scan times, especially without CT-based attenuation maps.
- Joint reconstruction algorithms (MLAA, MLACF) improve PET imaging but can produce noisy attenuation maps (μ-maps).
- Denoising these μ-maps is crucial for accurate PET image interpretation and diagnosis.
Purpose of the Study:
- To develop a novel cascaded relevancy-aware Generative Adversarial Network (reGAN) for denoising μ-maps.
- To enhance the diagnostic reliability and quality of PET imaging, particularly in low-dose scenarios.
- To improve quantitative accuracy in PET scans lacking CT-based attenuation correction.
Main Methods:
- A cascaded reGAN architecture was designed, integrating UPlus GAN modules, relevancy mapping, and contextual attention.
- The model was trained on PET/CT data from 16 patients, using MLAA/MLACF-derived μ-maps as input and CT-derived μ-maps as ground truth.
- Performance was evaluated using SSIM, PSNR, VIF, and MSE, with comparisons against other 2D and 3D GANs.
Main Results:
- reGAN achieved superior performance with the highest SSIM (0.91-0.93), PSNR (34.7-36.2 dB), and VIF (0.89-0.91), and lowest MSE (0.018-0.021).
- Qualitative analysis confirmed reGAN's ability to preserve fine details (e.g., bony structures) and effectively reduce artifacts.
- Relevancy maps provided pixel-wise confidence indicators, enhancing interpretability and diagnostic reliability.
Conclusions:
- The proposed reGAN offers a robust method for medical image denoising by combining generative modeling with diagnostic confidence metrics.
- This approach is a viable solution for achieving quantitative accuracy in low-dose PET imaging without CT-based attenuation maps.
- reGAN improves the reliability and diagnostic value of PET scans, addressing key challenges in modern medical imaging.
