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

Low-dose computed tomography (LDCT) imaging is widely regarded as an effective approach for reducing patients' exposure to X-ray radiation. However, the quality of the CT images may be compromised by the presence of noise and artifacts. Although generative adversarial networks (GANs) have demonstrated promising results in LDCT denoising, the performance of conventional models may be limited in capturing fine anatomical details and contextual dependencies. Additionally, the self-similarity-based techniques may not be robust enough to handle the noise components.

To resolve the issues related to LDCT, we propose a novel architecture with a generator consisting of a Deformable-Asymmetric Convolutional Residual Block (DACRB) and a Learnable Non-Local Attention Block (LNLAB). This allows the method to adapt to noise with varying spatial patterns and effectively preserve boundaries and textures. In addition, a self-similarity loss function based on KL divergence is employed to ensure structural consistency. A Swin Transformer-based perceptual loss is employed to achieve better visual quality.

The proposed method has been tested and validated using two public datasets. The proposed method has achieved a Peak Signal-to-Noise Ratio (PSNR) of 38.1055 dB and a Structural Similarity Index Measure (SSIM) of 0.9780 on the Mayo 2016 dataset, and a PSNR of 38.2210 dB and an SSIM of 0.9785 on the LDCT projection dataset.

The numerical results demonstrate the efficacy of the proposed framework in LDCT image reconstruction with an optimal trade-off between noise reduction and structural detail preservation.