Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss

Qingsong YangPingkun YanYanbo ZhangHengyong YuYongyi ShiXuanqin MouMannudeep K. KalraYi ZhangLing SunGe Wang

article2017IEEE Transactions on Medical Imaging1,454 citations

Proposes a Wasserstein generative adversarial framework with perceptual loss to suppress noise and artifacts in low-dose CT scans while preserving fine diagnostic structures.

Listen

Computed tomography (CT) imaging is a critical diagnostic tool in modern medicine, but concerns over patient radiation exposure have driven efforts to reduce radiation doses. However, low-dose CT scans inherently suffer from increased image noise and streak artifacts, which can obscure critical diagnostic details and undermine clinical decision-making. While standard machine learning denoising approaches improve mathematical pixel errors, they frequently produce over-smoothed, blurry images with artificial waxy textures that degrade subtle anatomical features. The article evaluates a deep learning framework designed to reduce noise in low-dose CT scans while preserving critical structural details and visual fidelity.

The evaluated method, named WGAN-VGG, integrates an advanced generative adversarial framework using Wasserstein distance with a perceptual loss function derived from a pre-trained feature extraction network. The generative adversarial framework aligns the statistical noise distribution of low-dose CT images with normal-dose images, while the perceptual loss compares high-level structural features rather than relying on conventional pixel-by-pixel comparisons. The authors trained and validated the system using a clinical dataset from the 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge, consisting of 10 anonymous patients with normal-dose and simulated quarter-dose abdominal CT scans, extracting over 100,000 image patch pairs for training. Performance was evaluated against traditional iterative reconstruction, standard convolutional neural networks, and alternative adversarial architectures using quantitative metrics, regional statistical properties, and a blinded qualitative study conducted by two independent radiologists.

The findings show that the proposed framework delivers superior diagnostic image quality compared to traditional and standard deep learning methods. In the blind reader evaluation, the WGAN-VGG method achieved the highest overall image quality score (3.70 out of 5) and the top artifact reduction score (3.45 out of 5), substantially outperforming standard pixel-loss networks and iterative reconstruction. While conventional networks and iterative reconstruction scored higher in raw noise suppression and traditional metrics such as peak signal-to-noise ratio, regional statistical analysis revealed that this was caused by severe over-smoothing that smeared out fine anatomical structures like small blood vessels. The proposed approach maintained standard deviation and tissue density values closely aligned with gold-standard normal-dose scans without generating synthetic distortions, successfully enhancing the visibility of subtle liver and kidney lesions.

These results demonstrate that standard pixel-based error metrics are insufficient for evaluating diagnostic medical imaging, as they reward blurriness over clinical accuracy. By preserving fine structural features and avoiding waxy artifacts, the framework enables diagnostic confidence at reduced radiation levels, directly supporting patient safety without requiring expensive scanner hardware replacements. Because the technique functions as an efficient post-processing step on reconstructed images, it offers a practical, computationally lightweight alternative to iterative reconstruction techniques that require access to proprietary raw scanner data.

To build on these results, decision-makers should support the integration and testing of perceptual loss frameworks in clinical imaging pipelines. Future technical development should focus on testing more advanced neural network generator architectures and extending deep learning directly to raw scanner projection data to recover information lost during initial image reconstruction. However, several operational limitations require caution: the framework operates strictly as a post-processing step on already-reconstructed images, meaning information fully lost during initial reconstruction cannot be recovered. Additionally, the network was evaluated on a single dataset with one simulated dose-reduction setting (quarter-dose). Model parameters will require retraining and fine-tuning across different scanner models, anatomical regions, and variable noise levels before broad clinical deployment.

Harvard University Huaxi MR Research Center Massachusetts General Hospital Rensselaer Polytechnic Institute Sichuan University University of Massachusetts Lowell West China Hospital Xi'an Jiaotong University ct image feature space gan low dose CT image denoising generative adversarial network Wasserstein distance

arXiv: 1708.00961

Paper: Wasserstein GAN, Martin Arjovsky et al. (2017). This foundational work introduces the Wasserstein GAN framework using the Earth Mover distance to stabilize adversarial optimization, which serves as the core training objective for the CT denoising model.
Paper: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, Christian Ledig et al. (2017). This paper establishes the concept of combining adversarial training with perceptual loss computed in deep feature spaces to preserve structural texture, directly inspiring the CT denoising network's objective.
Paper: Improved Training of Wasserstein GANs, Ishaan Gulrajani et al. (2017). This paper introduces the gradient penalty method to enforce Lipschitz continuity in Wasserstein GANs, resolving key stability limitations of weight clipping in adversarial generative training.
Paper: Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN), Hu Chen et al. (2017). This study establishes deep residual encoder-decoder architectures for low-dose CT denoising in the image domain, providing the baseline context and domain motivation.
Paper: Generative Adversarial Networks, Ian J. Goodfellow et al. (2014). This seminal text defines the generative adversarial network framework and minimax training formulation upon which all subsequent adversarial imaging models rely.
Paper: Image-to-Image Translation with Conditional Adversarial Networks, Phillip Isola et al. (2017). This work demonstrates paired image-to-image translation with conditional adversarial networks and combined loss functions, laying the groundwork for conditional image restoration.
Paper: Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising, Kai Zhang et al. (2016). This paper introduces deep residual learning specifically designed for image denoising to isolate and remove image noise components effectively.
Paper: Deep Convolutional Neural Network for Inverse Problems in Imaging, Kyong Hwan Jin et al. (2016). This paper introduces convolutional neural network solutions for medical imaging inverse problems, including sparse-view CT reconstruction and artifact reduction.

Paper: Generative Adversarial Network in Medical Imaging: A Review, Xin Yi et al. (2018). This comprehensive review categorizes and contextualizes GAN applications across medical imaging modalities, including low-dose CT reconstruction and denoising techniques.
Paper: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric, Richard Zhang et al. (2018). This study systematically validates the unreasonable effectiveness of deep neural network activations as perceptual metrics across diverse distortions and restoration tasks.
Paper: ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks, Xintao Wang et al. (2018). This work refines adversarial perceptual optimization by introducing relativistic discriminators and unactivated feature-space losses for superior texture preservation.
Paper: Noise2Noise: Learning Image Restoration without Clean Data, Jaakko Lehtinen et al. (2018). This work generalizes image denoising frameworks by demonstrating that deep networks can be trained directly on pairs of noisy images without requiring ground-truth clean scans.
Paper: Noise2Void - Learning Denoising From Single Noisy Images, Alexander Krull et al. (2018). This paper introduces a self-supervised blind-spot denoising formulation that removes the requirement for clean or paired target images in biomedical domains.
Paper: Spectral Normalization for Generative Adversarial Networks, Takeru Miyato et al. (2018). This paper advances Wasserstein and adversarial training stability by introducing spectral normalization on discriminator weights to strictly bound the Lipschitz constant.

Cover for Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss

Abstract

In this paper, we introduce a new CT image denoising method based on the generative adversarial network (GAN) with Wasserstein distance and perceptual similarity. The Wasserstein distance is a key concept of the optimal transform theory, and promises to improve the performance of the GAN. The perceptual loss compares the perceptual features of a denoised output against those of the ground truth in an established feature space, while the GAN helps migrate the data noise distribution from strong to weak. Therefore, our proposed method transfers our knowledge of visual perception to the image denoising task, is capable of not only reducing the image noise level but also keeping the critical information at the same time. Promising results have been obtained in our experiments with clinical CT images.

Topics

ct image feature space gan generative adversarial networks image noise image denoising methods non-local means wavelet thresholding human visual systems

Keywords

low dose CT image denoising generative adversarial network Wasserstein distance perceptual loss computed tomography image restoration artifact reduction WGAN-VGG

I Introduction
II Methods
II-A Noise Reduction Model
II-B WGAN
II-C Perceptual Loss
II-D Network Structures
II-E Other Networks
III Experiments
III-A Experimental Datasets
III-B Network Training
III-C Network Convergence
III-D Denoising Results
III-E Quantitative Analysis
III-F VGG Feature Extractor
IV Discussions and Conclusion
References

Knowls

Knowl 1 — WGAN-VGG Framework for Low-Dose CT Denoising
model/method
The WGAN-VGG framework performs low-dose CT (LDCT) image denoising by framing the task as a distribution mapping problem from low-dose images to normal-dose CT (NDCT) images while enforcing feature-level perceptual fidelity.

Traditional denoising networks that minimize mean squared error (MSE) compute per-pixel Euclidean distances, which tends to average high-resolution image patches and leads to blurred edges and waxy artifacts. WGAN-VGG addresses this through two complementary components:
1. A Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) that estimates the Earth Mover distance between the generated denoised image distribution and the true NDCT image distribution, driving the generator to statistically shift the noise distribution from strong to weak without pixel-wise blurring.
2. A perceptual loss computed using feature representations extracted from a pre-trained VGG-19 network, which measures differences between the denoised output and the ground truth in a high-dimensional feature space corresponding to structural features (edges, boundaries, and textures) rather than raw pixel intensities.
Knowl 2 — Joint Loss Function of WGAN-VGG
equation
The WGAN-VGG network is trained by solving the minimax optimization problem:

$\min_G \max_D \mathcal{L}_{\text{WGAN}}(D, G) + \lambda_1 \mathcal{L}_{\text{VGG}}(G)$

where $\lambda_1 > 0$ is a scalar hyperparameter that balances the adversarial distribution matching loss and the perceptual reconstruction loss.

The WGAN loss $\mathcal{L}_{\text{WGAN}}(D, G)$ with gradient penalty is given by:

$\mathcal{L}_{\text{WGAN}}(D, G) = -\mathbb{E}_{x \sim P_r}[D(x)] + \mathbb{E}_{z \sim P_L}[D(G(z))] + \lambda \mathbb{E}_{\hat{x} \sim P_{\hat{x}}}\left[\left(\|\nabla_{\hat{x}} D(\hat{x})\|_2 - 1\right)^2\right]$

where $x \in \mathbb{R}^{N \times N}$ denotes a normal-dose CT image sampled from the real NDCT distribution $P_r$ , $z \in \mathbb{R}^{N \times N}$ denotes a low-dose CT image sampled from the LDCT distribution $P_L$ , $G$ is the generator network, $D$ is the discriminator (critic) network without a final sigmoid activation, $\lambda$ is the gradient penalty coefficient, and $\hat{x} = \epsilon x + (1 - \epsilon) G(z)$ is uniformly sampled along straight lines connecting real and generated samples with $\epsilon \sim \text{Uniform}[0, 1]$ .
Knowl 3 — VGG Perceptual Loss for Grayscale CT Images
equation
The perceptual loss evaluates the structural discrepancy between the denoised image $G(z)$ and the reference normal-dose CT image $x$ in a deep convolutional feature space:

$\mathcal{L}_{\text{VGG}}(G) = \mathbb{E}_{(x, z)}\left[\frac{1}{w h d} \|\text{VGG}(G(z)) - \text{VGG}(x)\|_F^2\right]$

where $\text{VGG}(\cdot)$ denotes the feature map extracted from the 16th convolutional layer of a pre-trained VGG-19 network, $w$ , $h$ , and $d$ denote the width, height, and depth (channel count) of the extracted feature tensor respectively, and $\|\cdot\|_F$ is the Frobenius norm.

Because pre-trained VGG-19 expects three-channel RGB inputs while CT images are single-channel grayscale, the grayscale CT slices are duplicated across three identical color channels prior to being fed into the VGG feature extractor. The parameters of the VGG network remain fixed throughout training.
Knowl 4 — Generator and Discriminator Network Architectures in WGAN-VGG
model/method
The WGAN-VGG model consists of two trainable subnetworks:
- Generator ( $G$ ): An 8-layer fully convolutional network. Layers 1 through 7 each consist of 32 filters of size $3 \times 3$ with stride 1, followed by Rectified Linear Unit (ReLU) activations. Layer 8 uses a single $3 \times 3$ filter with stride 1 to produce the 1-channel denoised image output without pooling or strided downsampling.
- Discriminator ( $D$ ): A 6-layer convolutional critic followed by fully connected layers. Layers 1 and 2 each contain 64 filters ( $3 \times 3$ ); layers 3 and 4 each contain 128 filters ( $3 \times 3$ ); layers 5 and 6 each contain 256 filters ( $3 \times 3$ ). All convolutional layers use LeakyReLU activations. These are followed by two fully connected layers: the first has 1024 output units with LeakyReLU, and the second produces a single scalar output without a sigmoid activation.
Knowl 5 — WGAN-VGG Training Optimization Algorithm
algorithm
The WGAN-VGG model is trained by alternating updates between the discriminator $D$ and generator $G$ using the Adam optimizer with a gradient penalty on interpolated samples.
Input: Gradient penalty weight $\lambda = 10$ , VGG loss weight $\lambda_1 = 0.1$ , Adam learning rate $\alpha = 10^{-5}$ , momentum parameters $\beta_1 = 0.5, \beta_2 = 0.9$
Input: Total epochs $N_{\text{epoch}} = 100$ , discriminator iterations per step $N_D = 4$ , batch size $m = 128$ , patch size $80 \times 80$
Input: Initial discriminator parameters $w_0$ , initial generator parameters $\theta_0$ , fixed pre-trained VGG-19 parameters
for num_epoch = 1 to $N_{\text{epoch}}$ do
    for $t = 1$ to $N_D$ do
        Sample batch of NDCT patches $\{x^{(i)}\}_{i=1}^m$ , LDCT patches $\{z^{(i)}\}_{i=1}^m$ , and $\{\epsilon^{(i)}\}_{i=1}^m \sim \text{Uniform}[0, 1]$
        for $i = 1$ to $m$ do
             $\hat{x}^{(i)} \leftarrow \epsilon^{(i)} x^{(i)} + (1 - \epsilon^{(i)}) G(z^{(i)})$
             $L^{(i)}(D) \leftarrow D(G(z^{(i)})) - D(x^{(i)}) + \lambda (\|\nabla_{\hat{x}} D(\hat{x}^{(i)})\|_2 - 1)^2$
        end for
         $w \leftarrow \text{Adam}(\nabla_w \frac{1}{m} \sum_{i=1}^m L^{(i)}(D), w, \alpha, \beta_1, \beta_2)$
    end for
    Sample batch of LDCT patches $\{z^{(i)}\}_{i=1}^m$ and NDCT patches $\{x^{(i)}\}_{i=1}^m$
    for $i = 1$ to $m$ do
         $L^{(i)}(G) \leftarrow \lambda_1 L_{\text{VGG}}(z^{(i)}, x^{(i)}) - D(G(z^{(i)}))$
    end for
     $\theta \leftarrow \text{Adam}(\nabla_\theta \frac{1}{m} \sum_{i=1}^m L^{(i)}(G), \theta, \alpha, \beta_1, \beta_2)$
end for
Output: Trained generator parameters $\theta$
Knowl 6 — Experimental Setup and Baseline Network Configurations
experimental setup
The experiments utilized the clinical dataset from the 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge, containing 10 anonymous patients' normal-dose abdominal CT images and corresponding simulated quarter-dose CT images.
- Data extraction: 100,096 pairs of $64 \times 64$ patches extracted from 4,000 CT images for training (excluding background air patches), and 5,056 validation patch pairs extracted from 2,000 images.
- Evaluated configurations:
  - CNN-MSE: Generator trained with $\min_G L_{\text{MSE}}(G)$
  - CNN-VGG: Generator trained with $\min_G L_{\text{VGG}}(G)$
  - WGAN: Adversarial Wasserstein GAN with no content loss ( $\min_G \max_D L_{\text{WGAN}}(G, D)$ )
  - GAN: Standard minimax GAN ( $\min_G \max_D L_{\text{GAN}}(G, D)$ )
  - WGAN-MSE: WGAN regularized by MSE loss ( $\min_G \max_D L_{\text{WGAN}}(G, D) + \lambda_2 L_{\text{MSE}}(G)$ , $\lambda_2 = 0.1$ )
  - WGAN-VGG: Proposed network ( $\min_G \max_D L_{\text{WGAN}}(G, D) + \lambda_1 L_{\text{VGG}}(G)$ , $\lambda_1 = 0.1$ )
  - DictRecon: Iterative reconstruction based on 3D dictionary learning directly from raw projection sinogram data.

Knowl 7 — Quantitative Evaluation: PSNR and SSIM Discrepancy with Visual Quality

data/table

Peak Signal-to-Noise Ratio (PSNR, in dB) and Structural Similarity Index (SSIM) were computed on representative patient abdominal CT slices containing low-attenuation liver lesions:

Method	Fig. 5 Slice		Fig. 7 Slice
	PSNR (dB)	SSIM	PSNR (dB)	SSIM
LDCT	19.7904	0.7496	18.4519	0.6471
CNN-MSE	24.4894	0.7966	23.2649	0.7022
WGAN-MSE	24.0637	0.8090	22.7255	0.7122
CNN-VGG	23.2322	0.7926	22.0950	0.6972
WGAN-VGG	23.3942	0.7923	22.1620	0.6949
WGAN	22.0168	0.7745	20.9051	0.6759
GAN	21.8676	0.7581	21.0042	0.6632
DictRecon	24.2516	0.8148	24.0992	0.7631

CNN-MSE and DictRecon achieve the highest PSNR and SSIM metrics because they directly optimize pixel-level mean squared error or apply heavy regularization. However, these higher metrics correspond to over-smoothed textures and waxy artifacts, demonstrating that standard PSNR and SSIM alone do not adequately reflect diagnostic image quality and perceptual sharpness in medical CT denoising.

Knowl 8 — Hounsfield Unit Statistical Fidelity in Uniform CT Regions

data/table

Statistical analysis of CT numbers (mean and standard deviation in Hounsfield Units, HU) was conducted across flat regions-of-interest (ROIs) on two patient slices to evaluate noise texture preservation without over-smoothing, taking normal-dose CT (NDCT) as the gold standard:

Method	Fig. 5 ROI		Fig. 7 ROI
	Mean (HU)	SD (HU)	Mean (HU)	SD (HU)
NDCT (Gold Standard)	9	36	118	38
LDCT	11	74	118	66
CNN-MSE	12	18	120	15
WGAN-MSE	9	28	115	25
CNN-VGG	4	30	104	28
WGAN-VGG	9	31	111	29
WGAN	23	37	135	33
GAN	8	35	110	32
DictRecon	4	11	111	13

CNN-MSE and DictRecon produced standard deviations (SD = 18/15 HU and 11/13 HU) substantially lower than the gold standard NDCT (SD = 36/38 HU), confirming over-smoothing. WGAN-VGG maintained CT means matching NDCT (9 HU and 111 HU) and preserved a natural noise texture standard deviation (31 HU and 29 HU), closer to NDCT than MSE-based methods.

Knowl 9 — Radiologist Blind Reader Study on Denoising Quality

data/table

A blind reader study was conducted where two radiologists independently evaluated 10 groups of CT images across three metrics on a 5-point scale ( $1 = \text{unacceptable}$ , $5 = \text{excellent}$ ):

Method	Noise Suppression	Artifact Reduction	Overall Quality
NDCT	-	-	3.95 $\pm$ 0.20
LDCT	-	-	1.35 $\pm$ 0.16
CNN-MSE	4.35 $\pm$ 0.24	1.70 $\pm$ 0.28	2.15 $\pm$ 0.25
CNN-VGG	3.10 $\pm$ 0.23	2.85 $\pm$ 0.32	3.05 $\pm$ 0.20
WGAN-MSE	3.55 $\pm$ 0.25	3.05 $\pm$ 0.27	3.30 $\pm$ 0.21
WGAN-VGG	3.20 $\pm$ 0.25	3.45 $\pm$ 0.25	3.70 $\pm$ 0.15
WGAN	2.90 $\pm$ 0.26	2.90 $\pm$ 0.28	3.05 $\pm$ 0.22
GAN	3.00 $\pm$ 0.21	3.05 $\pm$ 0.27	3.10 $\pm$ 0.21
DictRecon	4.65 $\pm$ 0.20	2.05 $\pm$ 0.27	2.05 $\pm$ 0.36

Although CNN-MSE and DictRecon achieved high noise suppression scores ( $4.35$ and $4.65$ ), their overall quality scores dropped ( $2.15$ and $2.05$ ) due to severe blurring and waxy artifact formation ( $1.70$ and $2.05$ artifact reduction). WGAN-VGG achieved the highest artifact reduction ( $3.45 \pm 0.25$ ) and highest overall diagnostic quality score ( $3.70 \pm 0.15$ ), closely approaching the NDCT reference ( $3.95 \pm 0.20$ ).

Knowl 10 — Limitations of Image-Domain Post-Processing Denoising
limitation
The WGAN-VGG framework has three principal limitations:
1. Post-processing information loss: Because the model operates exclusively in the image domain on reconstructed filtered backprojection (FBP) slices, subtle anatomical structures (such as fine high-density spots) that are destroyed or absent in the initial low-dose FBP reconstruction cannot be recovered by the post-processing network, unlike sinogram-based iterative reconstruction methods (such as DictRecon).
2. Transferability of natural image features: The VGG network was pre-trained on natural photographic images (ImageNet). While its intermediate feature maps effectively extract edges and structural textures in CT, it can inadvertently capture and preserve residual noise patterns present in the reference normal-dose CT training targets.
3. Sensitivity to noise level changes: Because WGAN aligns probability distributions between specific source and target datasets, the network parameters and the weighting parameter $\lambda_1$ must be re-calibrated when transferring the model to CT scans acquired with different radiation doses or noise characteristics.

Coverage note — None was omitted; all contributed models, loss formulations, training procedures, network architectures, quantitative evaluations, and reader study results are covered.

References

1.D. J. Brenner and E. J. Hall, “Computed tomography - an increasing source of radiation exposure,” New England J. Med., vol. 357, no. 22, pp. 2277–2284, Nov. 2007.
2.A. B. De Gonzalez and S. Darby, “Risk of cancer from diagnostic x-rays: estimates for the UK and 14 other countries,” The Lancet, vol. 363, no. 9406, pp. 345–351, Jan. 2004.
3.J. Wang, H. Lu, T. Li, and Z. Liang, “Sinogram noise reduction for low-dose CT by statistics-based nonlinear filters,” in Med. Imag. 2005: Image Process., vol. 5747. International Society for Optics and Photonics, Apr. 2005, pp. 2058–2067.
4.J. Wang, T. Li, H. Lu, and Z. Liang, “iterative reconstruction for low-dose x-ray computed tomography,” IEEE Trans. Med. Imag., vol. 25, no. 10, pp. 1272–1283, Oct. 2006.
5.A. Manduca, L. Yu, J. D. Trzasko, N. Khaylova, J. M. Kofler, C. M. McCollough, and J. G. Fletcher, “Projection space denoising with bilateral filtering and CT noise modeling for dose reduction in CT,” Med. Phys., vol. 36, no. 11, pp. 4911–4919, Nov. 2009.
6.M. Beister, D. Kolditz, and W. A. Kalender, “iterative reconstruction methods in x-ray ct,” Physica Medica: Eur. J. Med. Phys., vol. 28, no. 2, pp. 94–108, Apr. 2012.
7.A. K. Hara, R. G. Paden, A. C. Silva, J. L. Kujak, H. J. Lawder, and W. Pavlicek, “iterative reconstruction technique for reducing body radiation dose at CT: Feasibility study,” Am. J. Roentgenol., vol. 193, no. 3, pp. 764–771, Sep. 2009.
8.J. Ma, J. Huang, Q. Feng, H. Zhang, H. Lu, Z. Liang, and W. Chen, “low-dose computed tomography image restoration using previous normal-dose scan,” Med. Phys., vol. 38, no. 10, pp. 5713–5731, Oct. 2011.
9.Y. Chen, X. Yin, L. Shi, H. Shu, L. Luo, J.-L. Coatrieux, and C. Toumoulin, “Improving abdomen tumor low-dose CT images using a fast dictionary learning based processing,” Phys. Med. Biol., vol. 58, no. 16, p. 5803, Aug. 2013.
10.P. F. Feruglio, C. Vinegoni, J. Gros, A. Sbarbati, and R. Weissleder, “Block matching 3D random noise filtering for absorption optical projection tomography,” Phys. Med. Biol., vol. 55, no. 18, p. 5401, Sep. 2010.
11.B. De Man and S. Basu, “distance-driven projection and backprojection in three dimensions,” Phys. Med. Biol., vol. 49, no. 11, p. 2463, May. 2004.
12.R. M. Lewitt, “multidimensional digital image representations using generalized Kaiser–Bessel window functions,” J. Opt. Soc. Amer. A, vol. 7, no. 10, pp. 1834–1846, Oct. 1990.
13.B. R. Whiting, P. Massoumzadeh, O. A. Earl, J. A. O’Sullivan, D. L. Snyder, and J. F. Williamson, “properties of preprocessed sinogram data in x-ray computed tomography,” Med. Phys., vol. 33, no. 9, pp. 3290–3303, Sep. 2006.
14.I. A. Elbakri and J. A. Fessler, “statistical image reconstruction for polyenergetic x-ray computed tomography,” IEEE Trans. Med. Imag., vol. 21, no. 2, pp. 89–99, Feb. 2002.
15.S. Ramani and J. A. Fessler, “ A splitting-based iterative algorithm for accelerated statistical x-ray CT reconstruction,” IEEE Trans. Med. Imag., vol. 31, no. 3, pp. 677–688, Mar. 2012.
16.E. Y. Sidky and X. Pan, “Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,” Phys. Med. Biol., vol. 53, no. 17, p. 4777, Aug. 2008.
17.Y. Liu, J. Ma, Y. Fan, and Z. Liang, “Adaptive-weighted total variation minimization for sparse data toward low-dose x-ray computed tomography image reconstruction,” Phys. Med. Biol., vol. 57, no. 23, p. 7923, Nov. 2012.
18.Z. Tian, X. Jia, K. Yuan, T. Pan, and S. B. Jiang, “low-dose CT reconstruction via edge-preserving total variation regularization,” Phys. Med. Biol., vol. 56, no. 18, p. 5949, Nov. 2011.
19.Q. Xu, H. Yu, X. Mou, L. Zhang, J. Hsieh, and G. Wang, “low-dose x-ray CT reconstruction via dictionary learning,” IEEE Trans. Med. Imag., vol. 31, no. 9, pp. 1682–1697, Sep. 2012.
20.Y. Zhang, X. Mou, G. Wang, and H. Yu, “tensor-based dictionary learning for spectral CT reconstruction,” IEEE Trans. Med. Imag., vol. 36, no. 1, pp. 142–154, Jan. 2017.
21.D. Kang, P. Slomka, R. Nakazato, J. Woo, D. S. Berman, C.-C. J. Kuo, and D. Dey, “Image denoising of low-radiation dose coronary CT angiography by an adaptive block-matching 3D algorithm,” in SPIE Med. Imag. International Society for Optics and Photonics, Mar. 2013, pp. 86 692G–86 692G.
22.G. Wang, M. Kalra, and C. G. Orton, “machine learning will transform radiology significantly within the next 5 years,” Med. Phys., vol. 44, no. 6, pp. 2041–2044, Mar. 2017.
23.G. Wang, “ A perspective on deep imaging,” IEEE Access, vol. 4, pp. 8914–8924, Nov. 2016.
24.C. Dong, C. C. Loy, K. He, and X. Tang, “Image super-resolution using deep convolutional networks,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 38, no. 2, pp. 295–307, Feb 2016.
25.H. Chen, Y. Zhang, W. Zhang, P. Liao, K. Li, J. Zhou, and G. Wang, “low-dose CT denoising with convolutional neural network,” 2016. [Online]. Available: Figures:1610.00321
26.H. Chen, Y. Zhang, M. K. Kalra, F. Lin, Y. Chen, P. Liao, J. Zhou, and G. Wang, “low-dose CT with a residual encoder-decoder convolutional neural network,” IEEE Trans. Med. Imag., vol. 36, no. 12, pp. 2524–2535, Dec. 2017.
27.E. Kang, J. Min, and J. C. Ye, “ A deep convolutional neural network using directional wavelets for low-dose x-ray CT reconstruction,” 2016. [Online]. Available: arXiv:1610.09736
28.O. Ronneberger, P. Fischer, and T. Brox, “ U-net: Convolutional networks for biomedical image segmentation,” in Proc. Int. Conf. Med. Image Comput. Comput.-Assisted Intervention. Springer, Nov. 2015, pp. 234–241.
29.J. Johnson, A. Alahi, and L. Fei-Fei, “perceptual losses for real-time style transfer and super-resolution,” 2016. [Online]. Available: arXiv:1603.08155
30.C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and W. Shi, “Photo-realistic single image super-resolution using a generative adversarial network,” 2016. [Online]. Available: arXiv:1609.04802
31.M. Arjovsky, S. Chintala, and L. Bottou, “wasserstein GAN,” 2017. [Online]. Available: arXiv:1701.07875
32.I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “generative adversarial nets,” in Advances Neural Inform. Process. Syst., 2014, pp. 2672–2680.
33.I. Goodfellow, “ NIPS 2016 tutorial: Generative adversarial networks,” 2017. [Online]. Available: arXiv:1701.00160
34.A. Brock, T. Lim, J. M. Ritchie, and N. Weston, “Neural photo editing with introspective adversarial networks,” 2016. [Online]. Available: arXiv:1609.07093
35.J.-Y. Zhu, P. Krähenbùhl, E. Shechtman, and A. A. Efros, “generative visual manipulation on the natural image manifold,” in Eur. Conf. Comput. Vision. Springer, 2016, pp. 597–613.
36.P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” 2016. [Online]. Available: arXiv:1611.07004
37.D. Nie, R. Trullo, C. Petitjean, S. Ruan, and D. Shen, “medical image synthesis with context-aware generative adversarial networks,” 2016. [Online]. Available: arXiv:1612.05362
38.J. M. Wolterink, T. Leiner, M. A. Viergever, and I. Isgum, “generative adversarial networks for noise reduction in low-dose CT,” IEEE Trans. Med. Imag., Dec. 2017.
39.S. Yu, H. Dong, G. Yang, G. Slabaugh, P. L. Dragotti, X. Ye, F. Liu, S. Arridge, J. Keegan, D. Firmin et al., “Deep de-aliasing for fast compressive sensing mri,” 2017. [Online]. Available: arXiv:1705.07137
40.M. Arjovsky and L. Bottou, “ Towards principled methods for training generative adversarial networks,” in NIPS 2016 Workshop on Adversarial Training. In review for ICLR, vol. 2016, 2017.
41.I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville, “Improved training of wasserstein GANs,” 2017. [Online]. Available: arXiv:1704.00028
42.M. Nixon and A. S. Aguado, Feature Extraction & Image Process., 2nd ed. Academic Press, 2008.
43.K. Simonyan and A. Zisserman, “ Very deep convolutional networks for large-scale image recognition,” 2014. [Online]. Available: arXiv: 1409.1556
44.S. Srinivas, R. K. Sarvadevabhatla, K. R. Mopuri, N. Prabhu, S. S. Kruthiventi, and R. V. Babu, “ A taxonomy of deep convolutional neural nets for computer vision,” Frontiers Robot. AI, vol. 2, p. 36, 2016.
45.AAPM, “low dose CT grand challenge,” 2017. [Online]. Available: http://www.aapm.org/GrandChallenge/LowDoseCT/#
46.D. Kingma and J. Ba, “ Adam: A method for stochastic optimization,” 2014. [Online]. Available: arXiv:1412.6980
47.M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin et al., “ Tensorflow: Large-scale machine learning on heterogeneous distributed systems,” 2016. [Online]. Available: arXiv:1603.04467
48.J. Deng, W. Dong, R. Socher, L. J. Li, K. Li, and L. Fei-Fei, “ Imagenet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition, June 2009, pp. 248–255.

Citation

MLA

Yang, Q., et al. “Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss”. IEEE Transactions on Medical Imaging, vol. 37, no. 6, 2018, pp. 1348–57, https://doi.org/10.1109/TMI.2018.2827462.

APA

Yang, Q., Yan, P., Zhang, Y., Yu, H., Shi, Y., Mou, X., Kalra, M. K., Zhang, Y., Sun, L., & Wang, G. (2018). Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss. IEEE Transactions on Medical Imaging, 37(6), 1348–1357. https://doi.org/10.1109/TMI.2018.2827462

Chicago

Yang, Q., P. Yan, Y. Zhang, et al. 2018. “Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss”. IEEE Transactions on Medical Imaging 37 (6): 1348–57. https://doi.org/10.1109/TMI.2018.2827462.

Harvard

Yang, Q. et al. (2018) “Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss”, IEEE Transactions on Medical Imaging, 37(6), pp. 1348–1357. Available at: https://doi.org/10.1109/TMI.2018.2827462.

Vancouver

1. Yang Q, Yan P, Zhang Y, Yu H, Shi Y, Mou X, Kalra MK, Zhang Y, Sun L, Wang G (2018) Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss. IEEE Transactions on Medical Imaging 37:1348–1357

BibTeX

@article{Yang_2018, title={Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss}, volume={37}, ISSN={1558-254X}, url={http://dx.doi.org/10.1109/TMI.2018.2827462}, DOI={10.1109/tmi.2018.2827462}, number={6}, journal={IEEE Transactions on Medical Imaging}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Yang, Qingsong and Yan, Pingkun and Zhang, Yanbo and Yu, Hengyong and Shi, Yongyi and Mou, Xuanqin and Kalra, Mannudeep K. and Zhang, Yi and Sun, Ling and Wang, Ge}, year={2018}, month=June, pages={1348–1357} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — WGAN-VGG Framework for Low-Dose CT Denoising

Knowl 2 — Joint Loss Function of WGAN-VGG

Knowl 3 — VGG Perceptual Loss for Grayscale CT Images

Knowl 4 — Generator and Discriminator Network Architectures in WGAN-VGG

Knowl 5 — WGAN-VGG Training Optimization Algorithm

Knowl 6 — Experimental Setup and Baseline Network Configurations

Knowl 7 — Quantitative Evaluation: PSNR and SSIM Discrepancy with Visual Quality

Knowl 8 — Hounsfield Unit Statistical Fidelity in Uniform CT Regions

Knowl 9 — Radiologist Blind Reader Study on Denoising Quality

Knowl 10 — Limitations of Image-Domain Post-Processing Denoising

References

Citation

BibTeX

Access the Paper