Compound feature attention network with edge enhancement for low-dose CT denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose CT (LDCT) imaging is crucial for reducing radiation exposure.
• LDCT images often suffer from significant noise and artifacts, compromising image quality and diagnostic accuracy.
• Existing denoising methods struggle to balance noise reduction with preservation of essential image details.

Purpose of the Study:

• To introduce a novel deep learning model, the Compound Feature Attention Network with Edge Enhancement (CFAN-Net), for effective LDCT denoising.
• To address the limitations of current methods in preserving image texture and anatomical details during noise reduction.
• To improve the overall readability and diagnostic utility of low-dose CT scans.

Main Methods:

• CFAN-Net integrates an edge-enhanced module utilizing trainable Sobel convolution for detail extraction.
• A Compound Feature Attention Block (CFAB) is proposed, comprising interactive feature learning, multi-scale feature fusion, and joint attention modules.
• Noise is reduced in a coarse-to-fine manner through cascaded CFAB applications, enhancing feature learning efficiency.

Main Results:

• CFAN-Net achieved a peak signal-to-noise ratio (PSNR) of 33.9692 on the 2016 NIH AAPM-Mayo LDCT challenge dataset.
• The model demonstrated a structural similarity index (SSIM) of 0.9198, indicating high fidelity to the original image structure.
• Experimental results confirm superior performance compared to existing LDCT denoising algorithms.

Conclusions:

• CFAN-Net effectively removes noise and artifacts from low-dose CT images.
• The proposed network excels at preserving the crucial texture and structural details of CT images.
• CFAN-Net offers a promising solution for enhancing the quality and diagnostic value of low-dose CT scans.