A deep unrolling network based on cartoon texture decomposition for low-dose CT reconstruction

Deep unrolling network, as a promising deep learning approach for low-dose computed tomography (LDCT) reconstruction, can efficiently address the issues of severe noise and artifacts in LDCT imaging. However, most existing methods predominantly unrolled the fidelity term to convolutional neural networks (CNNs) while failing to unroll regularization terms, which inevitably limits the feature capture capability of the network. To further improve the performance of deep unrolling networks, this paper proposed the CTDNet, a cartoon texture decomposition-based deep unrolling network that unrolls not only the data fidelity term but also regularization terms into CNNs.

The cartoon texture decomposition model of Meyer was incorporated as regularization terms, together with the data fidelity term, to formulate the reconstruction optimization problem. This problem was solved by the Chambolle-Pock (CP) algorithm, yielding a single-loop iterative algorithm instance. This instance was then unrolled to the deep reconstruction network for LDCT, by replacing each update step of iterative process with a simple sub-network. To further enhance the quality of reconstruction results, a lightweight image refinement and fusion module was employed to perform detail enhancement and remove residual noise of the reconstructed images.

Extensive experiments were conducted on the "Low-Dose CT Image and Projection Data" dataset and the "Piglet Dataset". The experiment results demonstrated that the CTDNet effectively removes artifacts and noise from LDCT images while maximally preserving textural structures, which enables it to outperform in terms of visual effects and objective metrics.

This work further unrolls the regularization terms to CNNs on the basis of unrolling the data fidelity term, providing a novel unrolling strategy for the future design of deep unrolling networks. .