Abstract: Concerning the problem of a large amount of noise in Low-Dose Computed Tomography (LDCT) reconstructed images, a deep residual Convolutional Neural Network for Stationary Wavelet Transform (SWT-CNN) model was proposed to estimate Normal-Dose Computed Tomography (NDCT) image from LDCT image. In training phase, the high-frequency coefficients of LDCT images after Stationary Wavelet Transform (SWT) three-level decomposition were taken as inputs, the residual coefficients were obtained by subtracting the high-frequency coefficients of NDCT images from high-frequency coefficients of LDCT images were taken as labels, and the mapping relationship between inputs and labels could be learned by deep CNN. In testing phase, the high-frequency coefficients of NDCT image could be predicted from the high-frequency coefficients of LDCT image by using this mapping relationship. Finally, the predicted NDCT image could be reconstructed by Stationary Wavelet Inverse Transform (ISWT). With the size of 512 x 512, 50 pairs of normal-dose chest and abdominal scan sections of the same phantom and reconstructed images with noise added to the projection field were used as data sets, of which 45 pairs constituted a training set and the remaining 5 pairs constituted a test set. The SWT-CNN model was compared with the-state-of-the-art methods, such as Non-Local Means (NLM), K-Singular Value Decomposition (K-SVD) algorithm, Block-Matching and 3D filtering (BM3D), and Image domain CNN (Image-CNN). The experimental results show that, the Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) of NDCT image predicted by SWT-CNN model are higher, and its Root Mean Square Error (RMSE) is smaller than that of other algorithms. The proposed model is feasible and effective in improving the quality of low-dose CT images.

Key words: Low-Dose Computed Tomography (LDCT), Stationary Wavelet Transform (SWT), Convolutional Neural Network (CNN), residual learning, deep learning

摘要： 针对低剂量计算机断层扫描（LDCT）重建图像中存在大量噪声的问题，提出了一种平稳小波的深度残差卷积神经网络（SWT-CNN）模型，可以从LDCT图像估计标准剂量计算机断层扫描（NDCT）图像。该模型在训练阶段，将LDCT图像经平稳小波（SWT）三级分解后的高频系数作为输入，将LDCT图像高频系数与NDCT图像高频系数相减得到残差系数作为标签，通过深度卷积神经网络（CNN）学习输入和标签之间的映射关系；在测试阶段，利用此映射关系即可从LDCT图像的高频系数中预测NDCT高频系数，最后通过平稳小波反变换（ISWT）重构预测的NDCT图像。实验采用50对大小为512×512的同一体模的常规剂量胸腔及腹腔扫描切片和投影域添加噪声后的重建图像作为数据集，其中45对作为训练集，其余5对作为测试集。将所提模型与效果较好的非局部降噪算法、K-奇异值分解（K-SVD）算法、匹配三维滤波（BM3D）算法及图像域CNN（Image-CNN）模型对比，实验结果表明，SWT-CNN模型预测的NDCT图像信噪比（PSNR）和结构相似性（SSIM）高，且均方根误差（RMSE）小于其他算法处理结果。该模型对于提高低剂量CT图像质量是可行且有效的。

关键词: 低剂量计算机断层扫描, 平稳小波变换, 卷积神经网络, 残差学习, 深度学习