A novel window-based hybrid attention network for low-dose CT image denoising

CT is a widely used medical imaging modality in clinical medicine. However, during the CT scanning process, X-ray radiation may cause radiation damage to the human body and even increase the risk of cancer. A common solution is to obtain low-dose CT (LDCT) images by reducing the working current of the X-ray tube, thereby effectively reducing the radiation dose received by patients. However, the quality of the resulting LDCT images also declines, exhibiting more noise and artifacts compared to normal-dose CT (NDCT) images. This paper proposes an LDCT image denoising method based on a window hybrid attention network. The overall structure of the network is based on the U-Net framework. In the encoder part, the proposed window hybrid attention block is used to extract image information. In the decoder part, the image features are efficiently fused to significantly restore the overall structure and texture details of the image. Additionally, we have also improved the feedforward network in the window hybrid attention block by proposing a multi-scale fusion feedforward network. This network utilizes three parallel branches to extract multi-scale local features and fuse multi-scale information, effectively enhancing the denoising capability of the network. Our experimental results demonstrate that the proposed denoising method achieves excellent objective scores of 29.0726 in PSNR and 0.8569 in SSIM on the AAPM real dataset. Our method reduces noise in LDCT images while preserving structural details more effectively than existing approaches, both on the QIN-LUNG-CT simulated dataset and on the AAPM real dataset, and will help enhance diagnostic accuracy for radiologists.