Generative Adversarial Networks for Noise Reduction in Low-Dose CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose CT (LDCT) scans are crucial for reducing radiation exposure but suffer from inherent noise.
• Image noise in LDCT can obscure diagnostic details and hinder accurate quantitative analysis.

Purpose of the Study:

• To develop and evaluate a deep learning approach for denoising LDCT images.
• To improve the image quality of LDCT scans to mimic routine-dose CT (RCT) images.

Main Methods:

• A convolutional neural network (CNN) generator was trained to denoise LDCT images.
• An adversarial CNN discriminator was used to improve the realism of denoised images.
• Three training strategies were compared: voxelwise loss, adversarial loss, and a combination of both.

Main Results:

• While voxelwise loss yielded the highest peak signal-to-noise ratio, adversarial training better captured routine-dose CT image statistics.
• Denoising improved the quantification of calcified inserts in phantom studies.
• Coronary calcium scoring was feasible in noisy LDCT patient images.

Conclusions:

• CNN-based image domain noise reduction for LDCT is feasible and effective.
• Adversarial training enhances the ability of CNNs to generate realistic LDCT images resembling RCT quality.
• This approach holds promise for improving diagnostic accuracy in low-dose CT imaging.