Related Experiment Video
Updated: Aug 31, 2025
07:53
Author Spotlight: Advancing 3D Modeling for Enhanced Diagnosis and Treatment of Pulmonary Nodules in Early-Stage Lung Cancer
Published on: October 13, 2023
1.6K
Improving image quality and lung nodule detection for low-dose chest CT by using generative adversarial network
1Department of Radiology, Ruijin Hospital affiliated to School of Medicine, Shanghai Jiao Tong University, Shanghai Jiao Tong, China.
The British Journal of Radiology
|August 22, 2022
Summary
Generative adversarial network (GAN) denoising models improve chest low-dose CT (LDCT) image quality and lung nodule detection. The Res model, targeting image residuals, outperformed the Dir model, showing potential clinical benefit.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Low-dose computed tomography (LDCT) is crucial for reducing radiation exposure in chest imaging.
- Image noise and reduced image quality in LDCT can hinder the accurate detection of pulmonary nodules.
- Generative adversarial networks (GANs) offer potential for image denoising and quality enhancement.
Purpose of the Study:
- To evaluate the effectiveness of two GAN-based denoising models (Dir and Res) in improving image quality and lung nodule detectability in chest LDCT.
- To compare the performance of GAN denoising models against standard LDCT and hybrid iterative reconstruction (IR) images.
- To investigate the impact of different GAN training targets (direct image vs. residual image) on denoising performance.
Main Methods:
- Two GAN denoising models, Dir and Res, were trained using LDCT images simulated from standard-dose CT (SDCT) data of 200 participants.
- The Dir model was trained to target SDCT images, while the Res model targeted the residual difference between SDCT and LDCT images.
- Model performance was evaluated using a phantom and 95 clinical chest LDCT scans, assessing image quality metrics and pulmonary nodule detection sensitivity by two radiologists.
Main Results:
- Both Res and Dir GAN models demonstrated improved structural similarity and peak signal-to-noise ratio compared to standard LDCT.
- The Res model exhibited the lowest standard deviation, indicating superior noise reduction.
- Radiologist diagnostic sensitivity for lung nodules increased with both GAN models, with the Res model showing a slight advantage (79/83%) over the Dir model (72/79%) and LDCT (72/77%).
Conclusions:
- GAN-based denoising models, particularly the Res model trained on image residuals, effectively reduce noise in chest LDCT.
- The Res model achieved superior image quality and enhanced lung nodule detectability compared to the Dir model and hybrid IR images.
- These findings highlight the clinical potential of GAN denoising for improving diagnostic accuracy in chest LDCT examinations.
