PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Low-Dose CT Denoising

Maryam Gholizadeh-Ansari, Javad Alirezaie, Paul Babyn

arXiv:1902.10127v1eess.IVcs.CVcs.LG

TL;DR

This paper addresses denoising low-dose CT images, a setting motivated by concern about CT radiation exposure. It combines dilated convolutions, residual connections, edge detection, and a joint MSE–perceptual objective; the authors report that the joint objective improves visual results while avoiding the respective smoothing and grid artifacts of the individual losses.

  • Problem

    CT radiation exposure motivates low-dose imaging, while the paper addresses denoising low-dose CT images.

    Source: §1 · pp. 2, 4

  • Method

    The network combines dilated convolutions, residual learning, an edge-detection layer, and a joint mean-square-error and perceptual-loss objective.

    Source: §1 · p. 4, §5 · p. 14

  • Results

    The joint objective provides better visual results than either loss alone, avoiding the smoothing and detail loss associated with per-pixel optimization and grid-like artifacts associated with perceptual loss.

    Source: §5 · p. 15

  • Takeaways & Limitations

    The reported combined objective retains image texture details without the stated artifacts of either loss used alone.

    Source: §5 · p. 15

    The receptive-field comparison assumes equal filter counts across layers and matching input and output channel counts.

    Source: §2.2 · p. 7

Abstract

from arXiv · show

Low-dose CT denoising is a challenging task that has been studied by many researchers. Some studies have used deep neural networks to improve the quality of low-dose CT images and achieved fruitful results. In this paper, we propose a deep neural network that uses dilated convolutions with different dilation rates instead of standard convolution helping to capture more contextual information in fewer layers. Also, we have employed residual learning by creating shortcut connections to transmit image information from the early layers to later ones. To further improve the performance of the network, we have introduced a non-trainable edge detection layer that extracts edges in horizontal, vertical, and diagonal directions. Finally, we demonstrate that optimizing the network by a combination of mean-square error loss and perceptual loss preserves many structural details in the CT image. This objective function does not suffer from over smoothing and blurring effects caused by per-pixel loss and grid-like artifacts resulting from perceptual loss. The experiments show that each modification to the network improves the outcome while only minimally changing the complexity of the network.

1 Introduction

Reducing CT radiation exposure is important, but lower X-ray current can produce noisier images that may be less diagnostic. The paper situates its denoising network among existing approaches and introduces architectural and loss-function choices intended to improve image quality without substantially increasing complexity.

1 Introduction

  • Lowering X-ray current can reduce radiation exposure, but the resulting CT images have increased noise and may be less diagnostic.The introduction notes particular concern about radiation exposure in patients requiring multiple scans and in children.Source: §1 · p. 2
  • Projection-space denoising uses sinogram data but can be vendor dependent, require scanner access and integration, and increase implementation cost.Iterative reconstruction can outperform projection-space methods, but also requires projection data and has high computational cost.Source: §1 · pp. 2–3
  • Image-space denoising works directly on reconstructed CT images without projection data and is generally fast, vendor independent, and easier to integrate.The introduction surveys KSVD, non-local means, and BM3D as methods used for low-dose CT denoising.Source: §1 · p. 3
  • The proposed network combines batch normalization, residual learning, dilated convolutions, and an edge-detection layer to improve denoising without substantially increasing complexity.Its objective combines MSE with perceptual loss, comparing feature maps from four blocks of pretrained VGG-16 [33].Source: §1 · p. 4

2 Methods

Because paired normal-dose and low-dose training data are difficult to obtain, the methods include a simulated low-dose dataset generated from normal-dose CT images. The simulation adds Poisson noise in the projection domain.

2.1 Low-Dose CT Simulation

  • Paired normal-dose and low-dose CT training samples are difficult to obtain, so the authors generate a simulated low-dose dataset alongside two other datasets.The passage identifies training-sample scarcity as a challenge for machine learning in medical imaging.Source: §2.1 · p. 5
  • The simulation adds Poisson noise to normal-dose sinogram data because low-dose projection-domain noise is described as predominantly Poisson distributed.The stated simulation procedure is attributed to.Source: §2.1 · p. 5
  • The procedure derives Hounsfield units and attenuation coefficients from the normal-dose image, then applies a Radon transform to obtain projection data before generating noisy transmission data.The passage also specifies multiplying projection data by voxel size to eliminate the size factor.Source: §2.1 · p. 5

7. Find the projection of the added Poisson noise,

The section develops an eight-layer dilated-convolution denoiser, adding symmetric shortcuts and fixed Sobel edge maps, then combines MSE and perceptual losses.

7. Find the projection of the added Poisson noise,

  • Low-dose attenuation is formed by adding inverse-Radon-transformed noise to normal-dose attenuation, then converted back to Hounsfield units.The subsection describes the added noise as Poisson noise.Source: §7 · p. 6

2.2 Dilated Convolution

  • The eight-layer network uses dilation rates 1, 2, 3, 4, 3, 2, 1, and 1 to capture more context with fewer weights.Dilated convolution enlarges the receptive field with a fraction of the weights needed by larger filters or more layers.Source: §2.2 · pp. 6–7

2.3 Residual Learning

  • Symmetric shortcuts concatenate the input and outputs of layers 2 and 3 with outputs of layers 7, 6, and 5, passing early image details forward.The authors report that these connections boost performance.Source: §2.3 · p. 7

2.4 Edge Detection Layer

  • A non-trainable layer applies four Sobel kernels and concatenates the resulting edge maps with the input image.The authors report that the edge-detection layer improves network performance.Source: §2.4 · p. 8

2.5 Objective Function

  • The objective combines pixelwise MSE with VGG-16 perceptual loss computed from feature maps in blocks 1–4.MSE produces blur and omits textural detail, while perceptual loss alone produces grid-like artifacts.Source: §2.5 · p. 8, Figure 3 · p. 9
  • Combining perceptual loss with MSE greatly improves the visual characteristics of the denoised images.The authors report blur and missing textural detail with MSE and grid-like artifacts with perceptual loss alone.Source: §2.5 · pp. 8–9

3 Experiments Setup

The evaluation uses simulated low-dose lung CT, real piglet CT, and thoracic phantom data, with held-out contiguous images for testing.

3 Experiments Setup

  • Evaluation uses simulated lung CT (663 slices), real piglet CT (900 images), and 407 thoracic phantom image pairs.The simulated scans used 330 mAs, 120 kVp, and 1.25 mm slices, with simulated low-dose incident flux I0 = 2 × 10^3; piglet scans used 300 mAs normal-dose and 15 mAs low-dose, 100 kVp, and 0.625 mm slices; thoracic scans used 480/60 mAs, 120 kVp, and 0.75 mm slices.Source: §3 · p. 9
  • Each dataset was split 70% for training and 30% for testing, with the final 30% of the original image sequence held out.Pixel values were divided by 4095 to map them between 0 and 1.Source: §3 · pp. 9–10
  • Training used overlapping 40 × 40 patches; full-size 512 × 512 test images were processed by the fully convolutional network.The receptive field is 37 pixels per direction, and zero padding was used to avoid boundary artifacts.Source: §3 · p. 10
  • The network used 64 filters per convolutional layer except layers 7 and 8, which used one filter; Adam training ran for two 20-epoch stages.The learning rates were 1e−3 and then 1e−4.Source: §3 · p. 10

4 Results

The experiments compare architectural and loss-function variants with BM3D and CNN200. They report that the proposed DRL-E network outperforms the alternatives, while combined MSE/perceptual training improves visual detail.

4 Results

  • The evaluation compares BM3D and CNN200 with network variants that test residual shortcuts, edge detection, and objective-function choice.The DRL variant adds shortcuts between layers 2/3 and 6/5; DRL-E adds an edge-detection layer, and DRL-E-MP uses combined MSE and perceptual loss.Source: §4 · p. 10
  • The proposed DRL-E network with dilated convolutions, shortcut connections, and edge detection outperforms the other tested networks, while combined-loss training improves perceptual image quality and preserves details.The authors report quantitative comparisons and visual confirmation of the perceptual improvement.Source: §4 · p. 11

4.1 Denoising results on simulated lung dataset

  • On simulated Lung data, MSE training smooths regions and affects texture, perceptual-only training creates grid artifacts, and DRL-E-MP preserves most texture detail with a better visual outcome.The results are shown in abdomen and lung windows; Table 2 reports average PSNR and SSIM.Source: §4.1 · p. 11, Table 2 · p. 11
  • MSE-trained networks have higher PSNR than perceptual-loss-trained networks, and perceptual loss does not improve PSNR.The paper attributes this to PSNR being higher when MSE is lower.Source: §4.1 · p. 11

4.2 Denoising results on real Piglet dataset

  • On real Piglet images, residual learning with edge detection outperforms the compared networks under MSE, while combined per-pixel and perceptual loss improves fine-detail reconstruction visually.Table 3 reports average PSNR and SSIM for the Piglet dataset.Source: §4.2 · p. 12, Table 3 · p. 13

4.3 Denoising results on phantom Thoracic dataset

  • Thoracic-dataset results are consistent with the other experiments, and comparison of DRL with DRL-E indicates that edge detection helps produce sharper, more precise edges.Table 4 reports average PSNR and SSIM for the Thoracic dataset.Source: §4.3 · p. 13, Table 4 · p. 14

5 Conclusion

The proposed network combines dilated convolutions, residual shortcuts, and a Sobel-like edge layer to denoise low-dose CT images. A joint MSE and perceptual loss gives better visual results than either loss alone while avoiding their reported artifacts.

5 Conclusion

  • The network uses dilated instead of standard convolutions and symmetric residual shortcuts to produce normal-dose-appearing images from low-dose CT inputs.Source: §5 · p. 14
  • A Sobel-like edge detection layer is included to capture image boundaries more effectively.Source: §5 · p. 14
  • Joint MSE and perceptual loss gives better visual results than either loss alone, avoiding oversmoothing and detail loss from per-pixel optimization and grid artifacts from perceptual-only optimization.Source: §5 · p. 15
Loading 1902.10127v1…