ATS-DDS: a general alternating training strategy for dual-domain sparse-view CT reconstruction

Sparse scanning methods are widely used to reduce the harmful effects of radiation doses on the human body. Recently, deep learning methods have been widely used for challenging reconstruction tasks such as sparse angles, while dual domain methods have natural advantages because they can handle information in both the sinusoidal and image domains. However, the existing methods in both fields do not pay enough attention to the allocation and strategy of training costs in both fields, but instead use the same training resources for unified training. To address this issue, this article designs a universal training strategy that can adapt to various network model structures: a general alternating training strategy for dual-domain sparse-view CT reconstruction (ATS-DDS) and conducts simulation experiments on various classic networks and the latest deep learning models. The results indicate that alternating training in dual-domain training scenarios can reduce the overall resources of network training and accelerate network training. At the same time, appropriately increasing information exchange between the two fields is beneficial for accelerating network training.