Dual-branch guided multi-scale half-instance normalization network for low-dose CT image denoising.

Abstract

Background: Low-dose computed tomography (LDCT) image denoising is a critical area of research in medical image processing. Compared to normal-dose CT, LDCT has gained significant attention due to its lower radiation dose, which reduces harm to the human body. However, the reduction in radiation dose introduces noise, which compromises the accuracy of medical diagnoses.

Purpose: The main goal of this study is to develop an efficient LDCT denoising model that can effectively extract adjacent frame image information, while focusing on both local image details and global structural information, and ensuring good inference time for clinical application.

Methods: This study proposes a dual-branch guided multi-scale half-instance normalization network (DGMINet) for LDCT image denoising. We introduce the efficient utilization of adjacent frame CT images to assist in denoising. By designing an adjacent frame image assistance module and constructing a dual-branch guided structure, the features of adjacent LDCT frame images are fused with those of the current frame image. This process enhances the representation of important regions, restores missing structural edges, and effectively retains local details. The fused features are then processed through a multi-scale half-instance normalization module, which captures multi-scale features using convolution kernels of varying sizes and innovatively adjusts the statistical properties of features at different scales through instance normalization. Additionally, the network employs the Charbonnier loss function to effectively preserving structural edges and texture features. These innovations enable DGMINet to effectively distinguish between noise and clean images, significantly improving denoising performance.

Results: Our experimental results show that the DGMINet method outperforms existing state-of-the-art denoising methods, demonstrating superior denoising performance. For example, on the AAPM dataset, compared to LDCT, the PSNR, SSIM, and FSIM metrics improved by 4.61 dB, 0.0544, and 0.0171, respectively, and the RMSE metric decreased by 5.95. On the real-world Piglet dataset, DGMINet also exhibited excellent denoising performance compared to LDCT at four different dose levels. Visually, DGMINet outperforms other denoising methods in terms of detail preservation and noise removal. Additionally, DGMINet maintains competitive inference times, proving its strong feasibility for practical applications.

Conclusions: The proposed DGMINet model achieves significant improvements in LDCT image denoising, offering an effective solution that removes noise while preserving crucial image details. Its outstanding performance and relatively efficient inference time highlight the model's potential for real-world clinical applications.