Time-frequency domain prior constrained deep unfolding network for low-dose CT reconstruction

Background: Low-dose computed tomography (LDCT) effectively reduces the risk of malignant disease; however, reducing the radiation dose introduces additional noise and stripe artifacts in the CT imaging process. While Convolutional Neural Networks (CNN) have demonstrated performance advantages in LDCT imaging tasks, their end-to-end network architecture limits adaptability to CT reconstruction tasks, leaving room for further performance improvement.

Objective: To propose a low-dose CT reconstruction network based on the iterative algorithms, incorporating an interpretable network architecture to achieve superior reconstruction performance.

Methods: To better adapt to CT reconstruction tasks, we proposed an interpretable deep unfolding network leveraging time-frequency and image domain priors to fully exploit the features extracted in the transform domain. The iterative optimization process of the proposed algorithm is mapped into a deep unfolding network, and a Stage Information Memory Network (SIMN) is designed to address information loss between adjacent stages and within each stage.

Results: Experimental results on Mayo and Piglet datasets show that the proposed model outperforms state-of-the-art techniques in both quantitative metrics and visual quality.

Conclusions: The proposed network effectively removes artifacts and noise from low-dose CT images, achieving excellent reconstruction performance.