[Paper Review] DU-GAN: Generative Adversarial Networks with Dual-Domain U-Net Based Discriminators for Low-Dose CT Denoising

Zhizhong Huang, Junping Zhang+2 more|arXiv (Cornell University)|Aug 24, 2021

Medical Imaging Techniques and Applications45 references199 citations

TL;DR

This paper proposes DU-GAN, a generative adversarial network with dual-domain U-Net-based discriminators for low-dose CT denoising. It leverages U-Net discriminators in both image and gradient domains to capture global structure and local details, improving denoising quality and reducing streak artifacts; the method achieves state-of-the-art performance on simulated and real-world datasets with enhanced edge preservation and uncertainty visualization via CutMix-augmented confidence maps.

ABSTRACT

LDCT has drawn major attention in the medical imaging field due to the potential health risks of CT-associated X-ray radiation to patients. Reducing the radiation dose, however, decreases the quality of the reconstructed images, which consequently compromises the diagnostic performance. Various deep learning techniques have been introduced to improve the image quality of LDCT images through denoising. GANs-based denoising methods usually leverage an additional classification network, i.e. discriminator, to learn the most discriminate difference between the denoised and normal-dose images and, hence, regularize the denoising model accordingly; it often focuses either on the global structure or local details. To better regularize the LDCT denoising model, this paper proposes a novel method, termed DU-GAN, which leverages U-Net based discriminators in the GANs framework to learn both global and local difference between the denoised and normal-dose images in both image and gradient domains. The merit of such a U-Net based discriminator is that it can not only provide the per-pixel feedback to the denoising network through the outputs of the U-Net but also focus on the global structure in a semantic level through the middle layer of the U-Net. In addition to the adversarial training in the image domain, we also apply another U-Net based discriminator in the image gradient domain to alleviate the artifacts caused by photon starvation and enhance the edge of the denoised CT images. Furthermore, the CutMix technique enables the per-pixel outputs of the U-Net based discriminator to provide radiologists with a confidence map to visualize the uncertainty of the denoised results, facilitating the LDCT-based screening and diagnosis. Extensive experiments on the simulated and real-world datasets demonstrate superior performance over recently published methods both qualitatively and quantitatively.

Motivation & Objective

• To address the challenge of noise and streak artifacts in low-dose CT (LDCT) images due to reduced radiation exposure.
• To improve denoising performance beyond traditional GANs by enhancing both global structural and local detail learning.
• To provide radiologists with a per-pixel confidence map for uncertainty visualization in denoised LDCT images.
• To reduce artifacts caused by photon starvation through gradient-domain adversarial training.
• To enhance discriminator robustness and generalization using CutMix data augmentation.

Proposed method

• Proposes a dual-domain GAN framework using U-Net-based discriminators in both image and gradient domains to learn multi-scale differences between denoised and normal-dose CT images.
• Employs a U-Net discriminator that outputs per-pixel confidence maps for local detail feedback and a global representation via the bottleneck layer for structural consistency.
• Introduces a second U-Net discriminator in the gradient domain to enhance edge details and suppress streak artifacts from photon starvation.
• Applies CutMix data augmentation to mix real and generated patches, improving discriminator generalization and enabling uncertainty estimation.
• Uses a multi-loss training objective combining adversarial loss, perceptual loss, and L1 loss to balance fidelity, texture, and structural quality.
• Trains the model on 64×64 patches to reduce computational cost while maintaining performance, with inference performed at full resolution.

Experimental results

Research questions

• RQ1 Can a U-Net-based discriminator improve both local detail and global structure learning in LDCT denoising compared to standard GAN discriminators?
• RQ2 Does dual-domain (image and gradient) adversarial training enhance edge preservation and artifact reduction in LDCT images?
• RQ3 Can CutMix-augmented U-Net discriminators provide reliable per-pixel uncertainty maps for clinical interpretation?
• RQ4 How does the patch size affect the performance of the U-Net discriminator in LDCT denoising?
• RQ5 Does the proposed DU-GAN generalize across different dose levels and real-world datasets?

Key findings

• The U-Net-based discriminator achieved the highest SSIM score (0.7454) among all discriminator types, outperforming patch, pixel, and global discriminators.
• The method with dual-domain training achieved PSNR of 23.1102, RMSE of 0.0724, and SSIM of 0.7343 on the Mayo-10% dataset, surpassing all ablation variants.
• Using CutMix improved the discriminator’s focus on local details, increasing SSIM and slightly reducing PSNR and RMSE.
• Smaller patch sizes (64×64) yielded better performance than larger ones, with PSNR of 22.3075 and SSIM of 0.7489.
• The dual-domain training effectively reduced streak artifacts and enhanced edge sharpness, as confirmed by visual and quantitative analysis.
• The method generalized well across simulated datasets at different dose levels and a real-world dataset with consistent architecture and hyperparameters.