ERA-WGAT: Edge-enhanced residual autoencoder with a window-based graph attention convolutional network for low-dose...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) is vital for medical diagnosis but poses radiation risks.
• Low-dose CT (LDCT) methods are needed to minimize patient exposure.
• Existing deep learning methods struggle with non-local information extraction.

Purpose of the Study:

• To develop an advanced deep learning model for effective LDCT image denoising.
• To improve the extraction of both local and non-local features in low-dose CT scans.
• To enhance the perceived image quality of LDCT images while reducing noise.

Main Methods:

• Proposed ERA-WGAT, a residual autoencoder with an edge enhancement module.
• Incorporated window-based graph attention convolutional network for non-local feature extraction.
• Utilized a compound loss function (MSE and multi-scale perceptual loss) to address over-smoothing.

Main Results:

• ERA-WGAT demonstrated superior noise suppression compared to existing LDCT denoising methods.
• The model achieved enhanced perceived image quality in low-dose CT scans.
• Effective extraction of rich edge information and non-local self-similarity was confirmed.

Conclusions:

• ERA-WGAT offers a promising solution for high-quality LDCT imaging.
• The method balances noise reduction with preservation of essential image details.
• This approach contributes to safer and more effective medical imaging practices.