低剂量CT图像全变分深度展开去噪网络

Deep Total Variation Denoising Network for Low-Dose CT Images

摘要:

对低剂量CT图像去噪进行了研究，分析了神经网络去噪在伪影抑制中计算性能低、泛化性不足的问题。采用各向异性全变分深度展开去噪网络，新方法结合图像相邻体素的边缘特性，引入各向异性TV正则项保留图像结构信息，避免各向同性TV导致的边缘模糊，并通过Chambolle-Pock算法求解数学模型，适配深度展开到卷积神经网络。此外，结合像素注意力机制进行网络优化，捕捉图像中的重要细节。经实验测试，基于Mayo 2016数据集，该方法在图像去噪效果上优于传统方法及其他先进网络模型，在PSNR、SSIM和VIF等指标上表现更优，满足低剂量CT图像高质量重建的需求。

Abstract:

A study was conducted on denoising low-dose CT images, analyzing the issues of low computational performance and insufficient generalization in neural network denoising for artifact suppression. An Anisotropic Total Variation Deep Unfolding Denoising Network was adopted, with the new method incorporating the edge characteristics of adjacent voxels by introducing an anisotropic TV regularization term to preserve the structural information of images and avoid edge blurring caused by isotropic TV. The Chambolle-Pock algorithm was employed to solve the mathematical model, adapting it for deep unfolding into convolutional neural networks. Additionally, a pixel attention mechanism was integrated for network optimization to capture important image details. Experimental tests based on the Mayo 2016 dataset demonstrated that this method outperforms traditional methods and other advanced network models in image denoising, showing superior performance in PSNR, SSIM, and VIF metrics. This method meets the requirements for high-quality reconstruction of low-dose CT images.