CGP-Uformer: A low-dose CT image denoising Uformer based on channel graph perception.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Low-dose computed tomography (CT) reduces radiation exposure but introduces significant image noise.
• This noise degrades image quality, hindering accurate analysis and diagnosis in medical applications.

Purpose of the Study:

• To introduce a novel deep learning network, the Channel Graph Perception based U-shaped Transformer (CGP-Uformer), for high-performance denoising of low-dose CT images.
• To address the challenge of noise in low-dose CT imaging, aiming to improve diagnostic accuracy.

Main Methods:

• The CGP-Uformer network integrates convolutional feed-forward Transformer (ConvF-Transformer) blocks for enhanced feature representation.
• It incorporates a channel graph perception block (CGPB) utilizing Graph Convolutional Networks (GCN) for inter-channel feature extraction.
• Spatial cross-attention (SC-Attention) blocks are employed to minimize semantic differences during feature fusion.

Main Results:

• Application of CGP-Uformer to the 2016 NIH AAPM-Mayo LDCT challenge dataset yielded a peak signal-to-noise ratio (PSNR) of 35.56.
• The network achieved a structural similarity index measure (SSIM) of 0.9221, indicating high image fidelity.
• Experimental results demonstrate superior denoising capabilities compared to four other leading networks.

Conclusions:

• The proposed CGP-Uformer network offers state-of-the-art performance in low-dose CT image denoising.
• It effectively preserves crucial image details while significantly reducing noise.
• This advancement holds promise for improving the diagnostic utility of low-dose CT scans.