A Low-dose CT Reconstruction Network Based on TV-regularized OSEM Algorithm

TL;DR

The paper addresses noise-induced degradation in LDCT by integrating TV regularization directly into the M-step of the EM reconstruction and solving with the Chambolle–Pock algorithm, complemented by an OS view-by-view strategy (OSEM-CP). It further unrolls this iterative method into an end-to-end network (OSEM-CPNN) that learns proximal operators and hyperparameters, enabling high-quality reconstructions with a single full-view iteration. Across synthetic and public datasets, OSEM-CP outperforms traditional TV-regularized and OS-based methods, while OSEM-CPNN achieves state-of-the-art or competitive performance with practical inference times and strong fine-tuning adaptability. The work highlights a practical pathway for accurate LDCT reconstruction by marrying statistical priors with data-driven priors in an unrolled optimization framework.

Abstract

Low-dose computed tomography (LDCT) offers significant advantages in reducing the potential harm to human bodies. However, reducing the X-ray dose in CT scanning often leads to severe noise and artifacts in the reconstructed images, which might adversely affect diagnosis. By utilizing the expectation maximization (EM) algorithm, statistical priors could be combined with artificial priors to improve LDCT reconstruction quality. However, conventional EM-based regularization methods adopt an alternating solving strategy, i.e. full reconstruction followed by image-regularization, resulting in over-smoothing and slow convergence. In this paper, we propose to integrate TV regularization into the ``M''-step of the EM algorithm, thus achieving effective and efficient regularization. Besides, by employing the Chambolle-Pock (CP) algorithm and the ordered subset (OS) strategy, we propose the OSEM-CP algorithm for LDCT reconstruction, in which both reconstruction and regularization are conducted view-by-view. Furthermore, by unrolling OSEM-CP, we propose an end-to-end reconstruction neural network (NN), named OSEM-CPNN, with remarkable performance and efficiency that achieves high-quality reconstructions in just one full-view iteration. Experiments on different models and datasets demonstrate our methods' outstanding performance compared to traditional and state-of-the-art deep-learning methods.

Figures (8)

• Figure 1: The structure of the $n_{th}$ sub-module in the proposed OSEM-CPNN.
• Figure 2: The Shepp-Logan image reconstruction results of the OSEM algorithm and the proposed OSEM-CP at different dose levels from $I_0=1\times{10^3}$ to $I_0=1\times{10^5}$. The displayed windows of the gray values are $[0,0.5]$.
• Figure 3: Results of the experiments with the Shepp-Logan phantom at the low-dose conditions of $I_{0}=5\times{10^3}$. The displayed windows of the gray values are set to $[0,0.5]$.
• Figure 4: Results of the experiments with the Ellipses-010215 phantom at the low-dose conditions of $I_{0}=1\times{10^4}$. The displayed windows of the gray values are set to $[0.78, 1.45]$.
• Figure 5: Results of the experiments with the IDRI-0001-012 phantom at the low-dose conditions of $I_{0}=5\times{10^4}$. The display windows are set to $[-1024, 2048]$ HU globally, $[-1024, -410]$ HU in the upper zoomed-in areas and $[783, 1295]$ HU in the lower zoomed-in areas.