Abstract: In order to solve the noise/artifact problem in Low-Dose Computed Tomography（LDCT） images, a Convolution Neural Network（CNN） based on Non-Sample Shearlet Transformation（NSST）, NSST-CNN model, is proposed in this paper. During training, NSST decomposition is performed on Normal-Dose Computed Tomography（NDCT） and LDCT images in the data set. The high-frequency sub-images of LDCT are used as the input, and the residual images of the high-frequency sub-images of NDCT and LDCT are used as the label. The mapping relationship between sub-images of LDCT and high-frequency sub-images of the residual images is then learned through CNN training. When testing, the high-frequency sub-images of LDCT are subtracted from the trained noise/artifact and the inverse NSST transform is performed to obtain a high-quality LDCT image. Experimental results show that NSST-CNN achieves a better balance between suppressing artifacts/noise and protecting structural details than KSVD, BM3D, and image space CNN method.

Key words: low-dose Computed Tomography（CT）, image restoration, non-sample Shearlet transformation, convolutional neural network, residual learning

摘要： 为解决低剂量CT（Low-Dose Computed Tomography，LDCT）图像中的噪声/伪影问题，提出一种基于非下采样Shearlet变换（Non-Sample Shearlet Transformation，NSST）的卷积神经网络（Convolution Neural Network，CNN）的NSST-CNN模型。训练时，对数据集中的常规剂量CT（Normal-Dose Computed Tomography，NDCT）和LDCT图像做NSST分解，将LDCT图像的高频子图作为输入，LDCT和NDCT图像的高频子图的残差图像作为标签，通过CNN训练，学习LDCT高频子图和高频残差子图的映射关系；测试时，将LDCT图像的高频子图减去利用映射关系预测的主要包括噪声/伪影的高频子图，然后做NSST反变换得到高质量的LDCT图像。实验结果表明，与KSVD、BM3D以及图像域CNN方法相比，NSST-CNN模型得到的结果具有更高的峰值信噪比和结构相似度，更接近NDCT图像。

关键词: 低剂量CT, 图像恢复, 非下采Shearlet变换, 卷积神经网络, 残差学习