Source-linked AI summary
Low Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss
Qingsong Yang, Pingkun Yan, Yanbo Zhang, Hengyong Yu, Yongyi Shi, Xuanqin Mou, Mannudeep K. Kalra, Ge Wang
TL;DR
Low-dose CT denoising must reduce noise without sacrificing diagnostically important details, while iterative and MSE-based approaches can suffer from computational cost, over-smoothing, or detail loss. The paper combines WGAN-based distribution matching with VGG perceptual loss, and clinical-image experiments report reduced noise, improved contrast, preserved details, and better lesion detection.
Problem
Low-dose CT denoising must reduce noise while preserving critical image details, but iterative and MSE-based approaches can incur high cost or over-smoothing.
Method
The method combines WGAN Wasserstein-distance matching between denoised LDCT and NDCT distributions with VGG-based perceptual loss.
Results
Clinical-image experiments showed WGAN-VGG reduced noise, increased contrast, avoided over-smoothing, and improved lesion detection.
Takeaways & Limitations
VGG-loss networks were desirable for visual quality with more details and fewer artifacts, despite MSE networks achieving higher traditional figures of merit.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
Low-dose CT reduces radiation exposure but increases noise and artifacts that can compromise diagnostic information. The paper proposes combining WGAN with perceptual loss to reduce noise while preserving clinically important image content.
- Lowering CT radiation dose increases reconstructed-image noise and artifacts, potentially compromising diagnostic information.
- Iterative reconstruction improves image quality but can lose details, retain artifacts, and impose high computational cost.
- Image post-processing is computationally efficient but can cause over-smoothing, residual errors, resolution loss, or edge blurring.
- End-to-end networks commonly use MSE, which is associated with over-smoothed edges and loss of details important for human perception.
- The proposed method combines Wasserstein distance with perceptual loss to align denoised LDCT distributions with NDCT while retaining image content.
- Perceptual loss compares VGG-extracted features rather than pixels, reflecting feature-based human visual comparison and image-manifold structure.
II. METHODS
The method formulates LDCT denoising as learning a mapping from low-dose to normal-dose CT images. The generator moves the LDCT distribution toward the NDCT distribution without requiring an explicit reconstructed-image noise model.
- The denoising function G maps an LDCT image z to an NDCT image x.
- The generator transforms samples from the LDCT distribution PL into a generated distribution Pg that is trained to approach the real NDCT distribution Pr.
- Reconstructed CT noise is complicated and nonuniform, making the relationship between LDCT and NDCT distributions difficult to specify with traditional methods.
B. WGAN
The WGAN component replaces the original GAN’s JS-divergence objective with Wasserstein-distance estimation and uses gradient penalty regularization. The discriminator and generator are trained alternately.
- WGAN compares generated and real data distributions using Wasserstein distance instead of JS divergence.
- The WGAN objective estimates Wasserstein distance and includes a gradient penalty for network regularization.
- The discriminator and generator are trained alternately by fixing one network while updating the other.
C. Perceptual Loss
The perceptual component supplements adversarial distribution matching with feature-based image-content preservation. It uses VGG features and combines the resulting loss with WGAN loss through a weighting parameter.
- MSE minimizes pixel-wise error but can generate blurry images and distort or lose details.
- The method replaces MSE with perceptual loss computed by comparing images in an established feature space.
- A pre-trained VGG-19 network extracts features for the perceptual loss, using its 16th convolutional-layer output.
- The joint objective combines WGAN loss and VGG perceptual loss using λ1 to control their trade-off.
D. Network Structures
WGAN-VGG combines an eight-layer CNN generator, a pre-trained VGG-based perceptual-loss calculator, and a discriminator. The network is trained on image patches and applied to entire images.
- Generator: The generator is an eight-layer CNN using stacked 3 × 3 convolutions, with 32 filters in its first seven hidden layers.Its final layer generates one feature map.
- Perceptual loss: The perceptual-loss module feeds the generated image and ground truth through a pre-trained VGG network and backpropagates feature reconstruction error only through the generator.VGG parameters remain fixed during this update.
- Discriminator: The discriminator contains six convolutional layers followed by two fully connected layers, with a single output and no sigmoid cross-entropy layer.Its convolutional filter counts progress from 64 to 128 to 256.
- Architecture notation: In the architecture notation, n denotes the number of convolutional kernels and s denotes convolutional stride; n32s1 means 32 kernels with stride 1.The notation is used in the network-structure figures.
- Deployment: The network is trained using image patches and then applied to entire images.
E. Other Networks
The experiments compare WGAN-VGG with networks using MSE, VGG, or WGAN objectives and evaluate them on paired clinical CT patches. Training includes an alternating discriminator procedure with gradient-penalty terms.
- Comparison networks: The comparison networks are CNN-MSE, CNN-VGG, WGAN-MSE, WGAN, and the original GAN.These variants isolate MSE loss, VGG loss, the WGAN framework, and additive-loss choices.
- Dataset: The clinical dataset contains normal-dose and simulated quarter-dose abdominal CT images from 10 anonymous patients.
- Dataset: Training uses 100,096 paired 64 × 64 image patches, with 5,056 additional paired patches extracted for validation.Mostly-air patches were excluded.
- Optimization: The WGAN optimization alternates discriminator and generator updates using interpolated samples and a gradient-penalty term.The algorithm specifies four discriminator iterations per epoch loop and batch size 128.
- Reference method: The study also compares the networks with a state-of-the-art 3D dictionary-learning reconstruction method using Mayo Clinic LDCT projection data.
B. Network Training
The networks are optimized with Adam under specified WGAN and loss-weight hyperparameters, and convergence is monitored through validation losses across five networks.
- Optimization: All networks were optimized using Adam with α = 1e−5, β1 = 0.5, and β2 = 0.9.The mini-batch size was 128.
- Monitoring: Validation MSE and VGG losses were calculated over 5,056 image patches after each epoch for the five networks.The corresponding loss curves are plotted against the number of epochs.
- Implementation: The networks were implemented in Python with TensorFlow and trained using an NVIDIA Titan XP GPU.
C. Network Convergence
Validation losses reveal different convergence behavior for pixel-wise and perceptual objectives, while increasing epochs reduces the estimated Wasserstein distance with a diminishing decay rate.
- Validation losses: MSE and VGG losses initially decrease together, but their rankings across networks subsequently run in opposite orders.CNN-MSE has the smallest MSE loss, whereas VGG-loss rankings differ.
- Validation losses: WGAN-MSE and WGAN-VGG show distinct convergence behavior, with WGAN perturbing MSE convergence but smoothly converging in VGG loss.
- Validation losses: The validation analysis compares MSE, VGG, and Wasserstein-related convergence curves across the five trained networks.
- Wasserstein convergence: Increasing the number of epochs reduces the estimated W-distance, although its decay rate becomes smaller.
- Wasserstein convergence: For WGAN-VGG, adding VGG loss improves perception and visibility at the cost of a compromised loss measure.
D. Denoising Results
The proposed WGAN-VGG denoising method produced visually improved CT images while preserving lesion and fine-structure visibility better than several comparison methods. Its post-processing design cannot recover information lost during FBP reconstruction.
- CNN-MSE blurred images and introduced waxy artifacts, whereas CNN-VGG and WGAN-VGG produced images more visually similar to NDCT.WGAN-MSE reduced over-smoothing but retained minor streak artifacts; WGAN or GAN alone generated stronger noise.
- All evaluated networks improved lesion visibility relative to the original noisy low-dose FBP images by reducing noise.The lesion examples include small low-attenuation liver lesions and a metastasis ROI.
- DictRecon reduced noise most aggressively but over-smoothed fine structures, including a vessel that remained identifiable in NDCT and WGAN-VGG images.DictRecon also preserved a bright spot visible in NDCT but not readily observable in LDCT or network-processed images.
- WGAN-VGG is a post-processing method, so information lost during FBP reconstruction cannot easily be recovered.DictRecon reconstructs from raw data, which contains more information than post-processing inputs.
E. Quantitative Analysis
Quantitative metrics favored methods optimized for those metrics, while regional statistics and visual assessment supported WGAN-VGG's balance of noise reduction and information preservation. The results also indicate that PSNR and SSIM alone may not fully capture CT image quality.
- CNN-MSE ranked first in PSNR, WGAN ranked worst, and DictRecon achieved the best SSIM with a high PSNR.The authors relate these rankings to the losses used during training and to DictRecon's strong regularization.
- DictRecon's strong quantitative performance coexisted with image blurring and blocky or waxy artifacts, suggesting PSNR and SSIM may be insufficient alone.The paper compares the metric results with visual observations from the reconstructed images.
- Prior cardiac CT and CS-MRI studies reported quantitative results consistent with the paper's pattern for MSE- and VGG-based networks.The cited studies also reported high PSNRs for MSE-based networks while discussing perceptual image quality.
- CNN-MSE and DictRecon produced smaller SDs than NDCT, indicating over-smoothing, while WGAN had a closer SD but smaller mean values.WGAN-VGG outperformed CNN-VGG, WGAN-MSE, and other selected methods in mean CT numbers, SDs, and visual impression.
- The reader study had two radiologists score noise suppression and artifact reduction on a five-point scale across 10 image groups.NDCT and LDCT images were included as references and were the only labeled images in each group.
F. VGG Feature Extractor
The paper uses VGG-based perceptual features alongside WGAN to make denoised CT images more similar to NDCT images while retaining details. Experiments report improved image quality, but the approach remains limited by noisy reference images, dataset dependence, and information loss during FBP reconstruction.
- VGG Feature Extractor: VGG extracts feature maps that emphasize CT boundaries, edges, and whole structures for perceptual comparison.The feature maps contain 512 images of size 32×32 arranged in a 32×16 array.
- VGG Feature Extractor: VGG loss compares denoised outputs with ground truth in feature space rather than pixel-by-pixel MSE.This follows the paper’s account of human vision as feature-based rather than purely pixel-wise.
- Experimental Findings: WGAN-VGG achieves higher PSNRs and more faithful statistical properties than CNN-VGG relative to NDCT images.The comparison is quantitative even though the two methods appear visually similar.
- Limitations: The WGAN-VGG loss combines feature-domain distance with adversarial loss, requiring careful balancing across datasets to limit image-content alteration.The experimental data contain only one noise setting, so networks require retraining or retuning for different noise properties.
- Limitations: The proposed network reduces noise and increases contrast, but its denoised images remain inferior to NDCT references and may retain noise features.NDCT images themselves contain noise, which may be captured by the VGG network and preserved in outputs.