PaperlayerView exact revision on arXiv

Source-linked AI summary

Low-dose CT denoising with convolutional neural network

Hu Chen, Yi Zhang, Weihua Zhang, Peixi Liao, Ke Li, Jiliu Zhou, Ge Wang

arXiv:1610.00321v1physics.med-phcs.CV

TL;DR

Lowering CT radiation dose can degrade image quality, while raw projection data may be inaccessible and low-dose image noise is difficult to characterize. The paper trains a deep CNN to transform low-dose CT images toward normal-dose images, reporting best overall test-set measurements and potential for medical imaging.

  • Problem

    Lowering X-ray tube current can degrade image quality through excessive quantum noise, while encrypted projection data and hard-to-determine image-domain noise constrain some restoration approaches.

    Source: §1 · p. 1, §2.1 · p. 2

  • Method

    A deep CNN learns to transform low-dose CT images toward corresponding normal-dose images by processing image patches through coding, filtering, and reconstruction.

    Source: §1 · p. 2, §2.2.1 · p. 2, §2.2.3 · p. 3

  • Results

    Across all 100 test images, the CNN-based method had the best performance on all reported metrics.

    Source: §3.3.2 · p. 4

  • Takeaways & Limitations

    The pilot study’s results demonstrate the potential of CNN-based methods for medical imaging.

    Source: §4 · p. 4

    The evaluation is a pilot study, and the authors identify network optimization as future work.

    Source: §4 · p. 4

Abstract

from arXiv · show

To reduce the potential radiation risk, low-dose CT has attracted much attention. However, simply lowering the radiation dose will lead to significant deterioration of the image quality. In this paper, we propose a noise reduction method for low-dose CT via deep neural network without accessing original projection data. A deep convolutional neural network is trained to transform low-dose CT images towards normal-dose CT images, patch by patch. Visual and quantitative evaluation demonstrates a competing performance of the proposed method.

1. INTRODUCTION

Lowering CT radiation dose can degrade image quality, while access to projection data and the noise properties of reconstructed images complicate existing denoising approaches. The paper proposes a deep CNN that transforms low-dose CT images toward corresponding normal-dose images.

  • Lowering X-ray tube current can reduce radiation dose but causes excessive quantum noise and significantly degrades CT image quality.Source: §1 · p. 1
  • Sinogram filtering and iterative reconstruction can be constrained by unavailable well-formatted projection data, and iterative methods also have heavy computational costs.Source: §1 · p. 1
  • Image-processing methods work directly on low-dose CT images, but nonuniform noise makes effective noise and artifact removal difficult with traditional methods.Source: §1 · p. 1
  • The paper proposes a deep CNN to transform low-dose CT images toward corresponding normal-dose images, following a small number of prior deep-learning imaging studies.Source: §1 · pp. 1–2

2. MEHTHODS

The method learns a mapping from low-dose CT images to corresponding normal-dose images, then processes and recombines image patches. Its CNN uses patch coding, nonlinear filtering, reconstruction, and training on paired patches.

2.1. Noise reduction model

  • The model treats a low-dose image as a quantum-noise-corrupted version of its corresponding normal-dose image and seeks a function approximating the inverse corruption process.Source: §2.1 · p. 2

2.1. Convolutional neural network

  • The denoising pipeline consists of patch coding, nonlinear filtering, and reconstruction.Source: §2.1 · p. 2

2.2.1 Patch encoding

  • Patch coding applies convolution and ReLU to extracted patches, producing feature vectors for subsequent processing.Source: §2.2.1 · p. 2

2.2.2 Non-linear filtering

  • Nonlinear filtering maps the first layer’s n_1-dimensional features to n_2 dimensions; adding layers may increase network capacity but also computation and training time.Source: §2.2.2 · p. 2

2.2.3 Reconstruction

  • Reconstruction weights overlapping processed patches before summing them into the final image, using a convolutional operation designed for this step.Source: §2.2.3 · pp. 2–3

2.2.4 Training

  • Training estimates network parameters from paired normal-dose and noisy image patches by minimizing a loss function with stochastic gradient descent.Source: §2.2.4 · p. 3

3. EXPERIMENTS

The experiments compare CNN denoising with ASD-POCS, KSVD, and BM3D using PSNR, RMSE, and SSIM; the proposed method performs best across the 100-image test set.

3. EXPERIMENTS

  • ASD-POCS, KSVD, and BM3D served as comparison methods, with parameters set according to their original references; PSNR, RMSE, and SSIM were the evaluation metrics.Experiments used MATLAB 2015b on a PC with an Intel i7 6700K CPU, 16 GB RAM, and GTX 980 Ti GPU.Source: §3 · p. 3

3.1 Dataset preparation

  • The evaluation used 100 randomly selected low-dose test images and 200 training images from different patients, with simulated low-dose images generated by adding Poisson noise to normal-dose sinograms.The dataset contained 7,015 normal-dose images from 165 patients; 33 × 33 patches with a slide step of 4 were used to increase training samples.Source: §3.1 · pp. 3–4

3.2 Parameter setting

  • The three-layer network used 64 and 32 filters in its first two layers, filter sizes of 9, 3, and 5, and a learning rate decaying from 0.001 to 0.0001.Filter weights were initialized from a zero-mean Gaussian distribution with standard deviation 0.001.Source: §3.2 · p. 4

3.3.1 Visual inspection

  • Only the proposed CNN recovered the most small details and boundaries in the representative abdomen slice, while ASD-POCS produced blocky effects and KSVD and BM3D left streak artifacts near bone.The visual comparison used representative chest and abdomen slices.Source: §3.3.1 · p. 4

3.3.2 Quantitative measurement

  • Across all 100 test images, the proposed CNN method had the best average PSNR, RMSE, and SSIM results.For the representative abdomen image, CNN was best except for SSIM; BM3D was more preferable for the soft-tissue image, according to the authors.Source: §3.3.2 · p. 4

4. CONCLUSION

This pilot study evaluates a CNN for low-dose CT noise reduction. The authors identify network optimization and application to other CT tasks as future directions.

4. CONCLUSION

  • The pilot study evaluated a deep CNN for noise reduction in low-dose CT, and its results demonstrate potential for CNN-based medical imaging.The authors propose future network optimization and applications to few-view CT reconstruction, metal artifact reduction, and interior CT.Source: §4 · p. 4
  • Future work could investigate other network architectures for dynamic and spectral CT problems.Source: §4 · p. 4
  • The proposed network structure remains to be optimized, with few-view CT reconstruction, metal artifact reduction, and interior CT named as possible application areas.Source: §4 · p. 4
Loading 1610.00321v1…