CNST: Cross-Network Semantic Transfer for low-dose CT denoising with attention encoding

Deep low-dose CT (LDCT) reconstruction methods commonly face the spectral bias problem. Where at the aspect of parameter updating, the gradients of low-frequency components tend to dominate the optimization process, thereby impeding the effective backpropagation of high-frequency component gradients. Regarding this problem, we propose a novel Cross-Network Semantic Transfer (CNST) framework. In the framework, to prevent the obstruction of low-frequency to high-frequency, the parameter updates of low-frequency and high-frequency components is decoupled into two separate backpropagation processes. Then, module-based frequency feature encoding is employed to construct a constraint loss, thereby enabling the integration of high and low frequency semantic information. In addition, for low-frequency encoding, to reduce the parameters involved in semantic a novel re-parameterization block is introduced. For high-frequency encoding, to compensate for the insufficient ability of data fidelity item in preserving high-frequency textures, based on attention encoding, an extra high-frequency loss are introduced. And both the high-frequency and low-frequency modules are trained within a generative adversarial framework. Experimental results demonstrate that CNST achieves competitive performance compared with some state-of-the-art methods. While effectively mitigating the spectral bias problem, CNST reconstructs textures closest to the real full-dose CT image.