Disentangled generative adversarial network for low-dose CT

Yi Zhang, Wenchao Du, Hu Chen, Hongyu Yang, Yi Zhang 0018

Introduction

Computed Tomography (CT) is an indispensable tool in modern medicine, yet the ionizing radiation associated with standard-dose CT (SDCT) poses potential health risks to patients. While low-dose CT (LDCT) significantly reduces radiation exposure, it introduces substantial noise and artifacts that can compromise diagnostic accuracy. Traditional denoising methods often struggle with a trade-off: they either over-smooth the image, leading to a "waxy" appearance that obscures fine details (the "Blur Trap"), or they introduce artificial textures that can be mistaken for pathology (the "Artifact Trap").

The DNSGAN Framework: A Disentangled Approach

To address these challenges, researchers have proposed DNSGAN, a novel "divide and conquer" strategy. The core innovation lies in disentangling the denoising process into two distinct, sequential tasks: noise suppression and structural enhancement.

Architecture of the DNSGAN, featuring the Disentangled Filter Network (DFN) and the Residual Dense Network (RDN).

The generator in DNSGAN consists of two primary modules:

  1. Noise Removal Module (DFN): Utilizing a Dynamic Filter Network combined with an LSTM, this module adaptively generates custom filters for each input slice. It iteratively refines the image over multiple time steps, focusing on suppressing stochastic noise while maintaining the base image structure.
  2. Structure Enhancement Module (RDN): A Residual Dense Network then takes the cleaned but potentially blurry output and restores fine anatomical textures. This module is trained using a combination of content, feature, and adversarial losses to ensure high fidelity to real tissue patterns.

Furthermore, the system employs a Relativistic Multi-scale PatchGAN as a discriminator. Unlike standard discriminators, it evaluates the relative realism of generated images compared to real ones at multiple scales, forcing the generator to produce more natural local textures.

Experimental Evaluation

The DNSGAN model was evaluated using two datasets:

  • Mayo Clinic CT Dataset: A benchmark dataset with simulated "quarter-dose" noise, allowing for quantitative comparison against ground-truth normal-dose scans.
  • Dongpu General Hospital (DGH) Dataset: A real-world dataset of head CT scans at "one-sixth" dose, testing the model's robustness against clinical noise patterns without perfectly registered ground truth.

Results and Clinical Implications

DNSGAN demonstrated superior performance both qualitatively and quantitatively. While models optimized solely for Peak Signal-to-Noise Ratio (PSNR) often produce visually unappealing results, DNSGAN prioritizes the Fréchet Inception Distance (FID), which better aligns with human visual perception.

Visual comparison of denoising results across different methods, highlighting DNSGAN's superior texture preservation and FID scores.

As shown in the comparative results, DNSGAN achieves significantly better FID scores (3.05) compared to other state-of-the-art methods like WGAN-VGG (5.75). Visually, the DNSGAN outputs retain sharp edges and realistic textures, avoiding the artificial "mottle" or over-smoothing seen in competing approaches. This high-fidelity restoration is crucial for radiologists to identify small lesions and subtle pathological changes accurately.

Conclusion

By disentangling noise removal from structural recovery, DNSGAN provides a robust solution for high-quality LDCT imaging. The model's ability to generalize from simulated training data to real-world clinical scans suggests significant potential for reducing patient radiation dose without sacrificing diagnostic confidence. Future research will focus on unsupervised learning techniques to further enhance performance on diverse clinical datasets where paired ground-truth data is unavailable.

Summary updated

Abstract

Generative adversarial network (GAN) has been applied for low-dose CT images to predict normal-dose CT images. However, the undesired artifacts and details bring uncertainty to the clinical diagnosis. In order to improve the visual quality while suppressing the noise, in this paper, we mainly studied the two key components of deep learning based low-dose CT (LDCT) restoration models—network architecture and adversarial loss, and proposed a disentangled noise suppression method based on GAN (DNSGAN) for LDCT. Specifically, a generator network, which contains the noise suppression and structure recovery modules, is proposed. Furthermore, a multi-scaled relativistic adversarial loss is introduced to preserve the finer structures of generated images. Experiments on simulated and real LDCT datasets show that the proposed method can effectively remove noise while recovering finer details and provide better visual perception than other state-of-the-art methods.
BibTeX
@article{zhang2021disentangled,
  author = {Yi Zhang and Wenchao Du and Hu Chen and Hongyu Yang and Yi Zhang},
  title = {{Disentangled generative adversarial network for low-dose CT}},
  journal = {Eurasip J. Adv. Signal Process},
  year = {2021},
  doi = {10.1186/s13634-021-00749-z},
}
Figure