DARL-Net: Deformable-asymmetric residual learning and learnable non-local attention-based low-dose CT image denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) reduces X-ray radiation exposure but can suffer from noise and artifacts.
• Conventional generative adversarial networks (GANs) and self-similarity methods have limitations in preserving anatomical details and handling noise patterns.

Purpose of the Study:

• To develop an advanced LDCT denoising method that overcomes limitations of existing techniques.
• To improve the preservation of fine anatomical details and contextual dependencies in noisy LDCT images.

Main Methods:

• A novel architecture featuring a Deformable-Asymmetric Convolutional Residual Block (DACRB) and Learnable Non-Local Attention Block (LNLAB) for adaptive noise handling.
• Implementation of a self-similarity loss function using KL divergence for structural consistency.
• Integration of a Swin Transformer-based perceptual loss for enhanced visual quality.

Main Results:

• The proposed method achieved a PSNR of 38.1055 dB and SSIM of 0.9780 on the Mayo 2016 dataset.
• On the LDCT projection dataset, the method yielded a PSNR of 38.2210 dB and SSIM of 0.9785.
• Demonstrated superior performance in noise reduction while preserving structural integrity.

Conclusions:

• The proposed framework effectively reconstructs LDCT images, balancing noise reduction with structural detail preservation.
• The novel architecture shows significant potential for improving diagnostic accuracy in low-dose CT imaging.