基于VGG网络与深层字典的低剂量CT图像去噪算法

低剂量计算机断层扫描（LDCT）能够有效降低X射线辐射对人体健康造成的危害，已广泛应用于医学临床诊断。针对LDCT图像中存在大量的斑点噪声和条形伪影的问题，提出一种结合改进的VGG网络和深层字典的图像去噪算法，以弥补深层字典去噪能力的不足。在深层字典学习到第一层字典原子和稀疏矩阵后，通过改进的VGG网络将字典原子区分为信息原子和噪声原子，同时将稀疏矩阵中噪声原子所对应的元素设置为零，降低噪声原子对图像去噪效果的影响。实验结果表明，与K-SVD算法、正则化K-SVD算法和深层字典学习算法相比，该算法的峰值信噪比和结构相似性指数平均提高了1.4 dB和0.03，能够有效抑制LDCT图像噪声和伪影，且保留较多的边缘和细节信息。

Low-Dose CT Image Denoising Algorithm Based on VGG Network and Deep Dictionary

Low-Dose Computed Tomography (LDCT) can effectively reduce the harm caused by X-ray radiation to human health, and has been widely used in medical clinical diagnosis.Focusing on the problem of extensive speckle noise and strip artifacts in LDCT images, an image denoising algorithm combined with improved VGG network and deep dictionary is proposed to make up for the deficiency of deep dictionary denoising ability.After learning the first layer dictionary atom and sparse matrix in the deep dictionary, the dictionary atom is divided into an information atom and noise atom through the improved VGG network.The elements corresponding to the noise atom in the sparse matrix are then set to zero, to reduce the influence of the noise atom on image denoising.The experimental results show that compared with the K-SVD algorithm, the regularized K-SVD and deep dictionary learning algorithms, improve the peak signal-to-noise ratio and structural similarity index by 1.4 dB and 0.03 on average, which can effectively suppress the noise and artifacts of LDCT images while retaining more edge and detail information.