Multi-Cycle-Consistent Adversarial Networks for CT Image Denoising

Jinglan Liu, Yukun Ding, Jinjun Xiong, Qianjun Jia, Meiping Huang, Jian Zhuang, Bike Xie, Chun-Chen Liu, Yiyu Shi

Abstract

Low-dose Computed Tomography (CT) imaging is essential for reducing radiation risk but introduces significant noise that can impair clinical diagnosis. Traditional supervised denoising models require paired noisy-clean data, which is difficult to obtain in practice. This paper introduces the Multi-Cycle-Consistent Adversarial Network (MCCAN), an unsupervised framework that decomposes the complex denoising task into smaller steps using an intermediate noise domain. By enforcing both local and global cycle consistency, MCCAN achieves superior noise reduction and anatomical preservation compared to standard CycleGAN approaches.

Introduction

Computed Tomography (CT) is a cornerstone of modern medical diagnostics. However, the ionizing radiation associated with CT scans poses potential health risks. While reducing the radiation dose is desirable, it leads to increased image noise and artifacts, potentially obscuring critical clinical details.

Deep learning has shown great promise in image denoising, but most state-of-the-art models rely on paired training data—identical scans at both low and normal doses. Obtaining such data is clinically impractical. Unsupervised methods, such as CycleGAN (referred to as CCADN in this study), address this by learning mappings between unpaired domains. However, when the noise gap between domains is large, these models often struggle, leading to "hallucinations" or the loss of fine anatomical structures because the transformation is too complex to learn in a single leap.

Methodology

The core innovation of MCCAN is the introduction of an intermediate domain ($Z$) that acts as a "stepping stone" between the noisy source domain ($X$) and the clean target domain ($Y$). Instead of a direct $X \to Y$ mapping, the model learns a sequential transformation: $X \to Z \to Y$. Domain $Z$ consists of images with moderate noise levels, simplifying the learning process for each individual generator.

The MCCAN architecture (a) and its multi-cycle consistency mapping logic (b).

To ensure the integrity of the transformed images, MCCAN employs two types of cycle consistency:

  • Global Cycles: These ensure that an image passing through the entire chain ($X \to Z \to Y \to Z \to X$) remains consistent with its original state. This couples all four generators, preventing the accumulation of errors across steps.
  • Local Cycles: These enforce consistency between adjacent domains (e.g., $X \to Z \to X$). Local cycles provide "shallow" supervision, which is easier to optimize and crucial for anchoring the generators to the specific anatomical content of the input, preventing "content swapping" where the model might scramble background and foreground textures.

Experimental Setup

The model was trained on a dataset of 3D CT images obtained from a 256-slice MDCT scanner. The training set included 200 normal-dose (clean) and 200 low-dose (noisy) scans. The intermediate domain $Z$ was constructed by identifying scans within the dataset that naturally exhibited intermediate noise characteristics. Performance was evaluated on a separate test set of 11 images, comparing MCCAN against standard CCADN and various ablation configurations.

Results and Discussion

MCCAN demonstrated significant improvements in both qualitative and quantitative metrics. Visually, the model produced the cleanest images with the sharpest edges, effectively removing noise while maintaining the clarity of small vessels and bone structures.

Qualitative comparison of CT denoising results, highlighting the preservation of anatomical details in the red-boxed regions.

Quantitatively, noise reduction was measured using the Standard Deviation (SD) in homogeneous regions. While the standard CCADN reduced noise by 15–22%, MCCAN achieved a reduction of 24–32%. Furthermore, MCCAN maintained mean pixel values closest to the original images, confirming its ability to preserve the physical "substance" of the tissue. Ablation studies highlighted the necessity of local cycles; without them, the GANs often produced "meaningless output" where anatomical structures were incorrectly swapped or distorted.

Conclusion

MCCAN effectively addresses the challenges of unsupervised CT denoising by leveraging intermediate domains and multi-cycle consistency. By breaking down the transformation into manageable steps, the model achieves state-of-the-art results without paired data. Future research will focus on automating the selection of intermediate domains and exploring the potential for multiple intermediate stages in even more complex imaging tasks.

Summary updated

BibTeX
@inproceedings{liu2020multi,
  author = {Jinglan Liu and Yukun Ding and Jinjun Xiong and Qianjun Jia and Meiping Huang and Jian Zhuang and Bike Xie and Chun-Chen Liu and Yiyu Shi},
  title = {{Multi-Cycle-Consistent Adversarial Networks for CT Image Denoising}},
  booktitle = {ISBI},
  year = {2020},
  doi = {10.1109/isbi45749.2020.9098683},
}
Figure