Asymmetric Convolution-based GAN Framework for Low-Dose CT Image Denoising

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose CT (LDCT) image denoising is crucial for enhancing diagnostic accuracy.
• Existing generative adversarial network (GAN) methods may lose high-frequency details and introduce structural distortions.
• Preserving local and global pixel correlations is vital for high-quality LDCT imaging.

Purpose of the Study:

• To develop a novel deep learning model for effective LDCT image denoising.
• To address limitations of current methods in preserving anatomical details and preventing shape distortion.
• To improve the diagnostic quality of LDCT images through advanced noise reduction.

Main Methods:

• A novel asymmetric convolution-based generator network (ACGNet) was developed.
• ACGNet utilizes 1D asymmetric convolutions (1x3 & 3x1) and a dynamic attention module (DAM).
• A neural structure preserving loss (NSPL) and differential content loss were incorporated for enhanced reconstruction.

Main Results:

• ACGNet demonstrated superior performance in preserving local and global pixel relations in denoised LDCT images.
• The method successfully prevented structural (shape) distortion, maintaining image integrity.
• ACGNet achieved state-of-the-art results on public datasets, with PSNR of 35.2015 dB and SSIM of 0.9560 on the Mayo 2016 dataset.

Conclusions:

• ACGNet efficiently denoises LDCT images while preserving critical anatomical details and structural integrity.
• The proposed NSPL and differential content loss contribute to human-perceived image quality and lesion boundary restoration.
• ACGNet represents a significant advancement in deep learning-based denoising for low-dose CT imaging.