A novel denoising method for CT images based on U-net and multi-attention

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Reducing radiation dose in computed tomography (CT) increases image noise, potentially impacting diagnostic accuracy.
• Existing denoising methods for low-dose CT (LDCT) images may degrade critical clinical lesion edge information.
• There is a need for advanced denoising techniques that preserve image details and mimic human visual attention.

Purpose of the Study:

• To propose a novel denoising method for medical CT images based on U-Net and a multi-attention mechanism.
• To enhance the retention of detailed features and clinical lesion information in denoised LDCT images.
• To develop an anthropomorphic denoising network that simulates human observational attention.

Main Methods:

• A novel denoising method integrating U-Net architecture with a multi-attention mechanism.
• Introduction of three attention modules: local attention, multi-feature channel attention, and hierarchical attention.
• Incorporation of an enhanced learning module with stacked convolutional, batch normalization, and activation layers to increase network depth and detail retention.

Main Results:

• The developed method effectively removes noise from CT images, significantly improving image quality.
• Quantitative evaluation using peak signal to noise ratio (PSNR) and structural similarity (SSIM) metrics demonstrated superior performance.
• Achieved high PSNR (34.7329) and SSIM (0.9293) on the QIN_LUNG_CT dataset (σ=10) and strong results on the Mayo Clinic LDCT dataset.

Conclusions:

• The proposed U-Net and multi-attention based denoising method is effective for LDCT image enhancement.
• The method successfully preserves important clinical lesion edge information, crucial for accurate diagnosis.
• This approach offers a promising solution for improving the diagnostic utility of low-dose CT imaging.