深度学习应用于低剂量CT图像的伪影和噪声抑制已取得一定成果，成为该领域新的研究热点。然而，由于成对训练数据不足造成网络训练不充分，制约了伪影和噪声的抑制效果。本项目提出采用半监督学习机制训练生成对抗网络，用真实成对数据约束生成数据的质量，两种数据在判决器中形成循环对抗训练模式，通过交互式学习提高网络对伪影和噪声的抑制效果。配合训练模式的改变，提出对图像不同语义信息在变换域分通道训练的网络结构，并通过设计多描述损失函数约束训练过程，从而提高网络对不同语义信息的处理能力。此外，通过引入注意力子网和噪声水平估计子网，分通道有针对性地学习不同语义信息的增强特征，来克服网络对噪声和伪影特征表达能力不足的问题。本项目致力于探索在成对CT图像有限的约束条件下生成对抗网络的训练机制和模型结构设计，该问题具有重要的学术研究价值，研究成果可用于提高低剂量CT的成像质量，具有潜在的临床应用价值。

The application of deep learning to low-dose CT images’ artifacts and noise suppression has achieved certain results and become a new research hotspot in this field. However, network training is insufficient due to the lack of paired training data, which greatly restricts the effectiveness of artifacts and noise suppression. This project proposes a semi-supervised learning mechanism to train generative adversarial network, in which the real paired data is used to supervise the quality of the synthesized data. The two kinds of data form a circular adversarial training mode in the discriminator, thus improving the network’s effectiveness of artifacts and noise suppression through interactive learning. Cooperated with the transformation of the training model, this study proposes a new network structure to train different semantic information in the different sub-channels in transform domain, and designs multiple description loss functions to constrain the training process, thereby improving the sensitivity of the network to different semantic information. In addition, by introducing attention subnets and noise level estimation subnets, learning the enhanced features of different semantic information in different channels in a targeted way to improve the network's insufficient ability to express noise and artifact features. This project is devoted to exploring the training mechanism and model structure design of generative adversarial networks under the constraint that there are few pairs of CT images, which has important academic research value. The research results can be used to improve the imaging quality of low-dose CT, which have potential clinical application value.