Dual-branch guided multi-scale half-instance normalization network for low-dose CT image denoising.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but introduces noise, compromising diagnostic accuracy.
• Effective LDCT image denoising is crucial for reliable medical diagnoses.

Purpose of the Study:

• To develop an efficient LDCT denoising model (DGMINet) that leverages adjacent frame information.
• To focus on preserving both local image details and global structural information.
• To ensure a competitive inference time for clinical applications.

Main Methods:

• Proposed DGMINet, a dual-branch guided multi-scale half-instance normalization network.
• Utilized adjacent frame CT images via an assistance module and dual-branch structure for feature fusion.
• Employed a multi-scale half-instance normalization module and Charbonnier loss function for enhanced feature extraction and detail preservation.

Main Results:

• DGMINet significantly outperformed state-of-the-art methods on AAPM and Piglet datasets.
• Achieved substantial improvements in PSNR, SSIM, and FSIM, with a decrease in RMSE.
• Demonstrated superior visual quality in detail preservation and noise removal, with competitive inference times.

Conclusions:

• DGMINet effectively denoises LDCT images while preserving critical details.
• The model shows significant potential for real-world clinical applications due to its performance and efficiency.