Related Experiment Video
Updated: Aug 9, 2025
05:32
Retrospective Cardiac Gating with A Prototype Small-Animal X-ray Computed Tomograph
Published on: February 21, 2025
357
Fast myocardial perfusion SPECT denoising using an attention-guided generative adversarial network
Jingzhang Sun1, Bang-Hung Yang2,3, Chien-Ying Li2,3
1Biomedical Imaging Laboratory (BIG), Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Taipa, Macao SAR, China.
Frontiers in Medicine
|February 23, 2023
Summary
A novel 3D attention-guided generative adversarial network (AttGAN) effectively denoises myocardial perfusion (MP) SPECT images. This deep learning approach outperforms conventional methods, improving image quality for faster scans.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Nuclear Medicine
Background:
- Deep learning shows promise for denoising myocardial perfusion (MP) SPECT.
- Conventional convolutional neural networks (CNNs) struggle with learning feature dependencies across large regions due to fixed kernel sizes.
- Attention mechanisms can learn relationships between local receptive fields and other image voxels.
Purpose of the Study:
- To propose a 3D attention-guided generative adversarial network (AttGAN) for denoising fast MP-SPECT images.
- To evaluate the effectiveness of AttGAN in learning feature dependencies across large regions for improved MP-SPECT denoising.
Main Methods:
- Fifty patients underwent 99mTc-sestamibi stress SPECT/CT scans.
- Fast MP-SPECT projection images (1-7s) were generated from full-time data.
- AttGAN, AttGAN with defect information (AttGAN-def), cGAN, and Unet were implemented and trained.
- Quantitative indices, including perfusion defect size (PDS), were analyzed using cross-validation on all 50 patients.
Main Results:
- AttGAN-based networks demonstrated superior quantitative indices compared to cGAN and Unet across all acquisition times.
- AttGAN-def further enhanced performance, achieving a mean absolute error of 1.60 for PDS at 1s/prj, compared to 2.36 (AttGAN), 2.76 (cGAN), and 3.02 (Unet).
Conclusions:
- AttGAN-based denoising significantly outperforms conventional CNN-based networks for MP-SPECT.
- The attention mechanism in AttGAN is effective for learning long-range dependencies, leading to improved denoising of fast MP-SPECT images.
