4856
Gibbs-Ringing Artifact reduction in MR images with varying sampling levels Via a Single Convolutional Neural Network1Guangdong Provincial Key Laboratory of Medical Image Processing & Key Laboratory of Mental Health of the Ministry of Education, School of Biomedical Engineering, Southern Medical University, Guangzhou, China, Guangzhou, China, 2Sun Yat-Sen University Cancer Center, Guangzhou, China, 3Philips Healthcare, Guangzhou, China, Guangzhou, China, 4Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China, Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong SAR, China, Hong Kong, China
Synopsis
Gibbs-ringing artifact is caused by the insufficient sampling of the high frequency data. And in clinical practice, the appearance of ringing artifact, i.e. the real sampling level, is not accurately obtained. To address this problem, a single convolutional neural network (CNN) has been trained for reducing Gibbs-ringing artifact in MR images under varying sampling levels. The experimental results demonstrate that Gibbs-ringing artifact can be effectively reduced by the proposed method without introducing noticeable blurring.
Purpose
To develop a deep learning based method to reduce Gibbs-ringing artifact in MR images under varying sampling levels.Introduction
Gibbs-ringing artifact is a common image distortion that manifests itself as a spurious ringing around sharp edges.1 This artifact in MR images is caused by insufficient sampling of the high frequency data in k-space domain. In clinical practice, the sampling level and pattern of the high frequency data is not accurately obtained. The varying appearance of ringing artifact may complicate diagnosis.2 Moreover, the artifact in diffusion-weighed (DW) images may affect the subsequent quantification of diffusion-related parameters.3 The ringing artifact can be reduced using low-pass filters with smooth roll-offs at expense of blurring.4 Machine learning based methods have been used to reduce the ringing artifact by predicting unknown high frequency components from the measured low frequency data5–8. This work propose a single CNN-based method for reducing the ringing artifact in MR images with multiple sampling levels.9Methods
Our proposed method employs a single CNN that learns an end-to-end mapping to predict artifact-free images from original images with varying sampling levels, and then replace the low-frequency k-space part with the measured data to obtain the final output images. Figure 1 demonstrates the overall architecture of the proposed CNN-based method. The proposed method takes the concatenated ringing MR image and degradation maps as the input. The degradation maps are made based on the dimensionality stretching strategy proposed in this paper9 and the difference with it is that we randomly sample the n values from the sinc signal obtained by applying Fourier transform to the corresponding sampling pattern and stretch them into degradation maps, rather than using PCA technique. In this work, n = 16 and these values are sampled from the beginning of the middle value of the sinc signal. There were 4 layers of the single CNN-based method. Rectified linear unit (ReLU, max(0,x))10 was applied as the activation function, and the batch normalization 11 was used to accelerate network training and boosts accuracy. The filter size was set f1 = 9, f2 = 7, f3 = 1, and f4 = 5; and the filter number was set n1 = 64, n2 = 32, n3 = 16 and n4 = 1. The mean-squared error (MSE) was used as the loss function and minimized by using stochastic gradient descent with the standard backpropagation.
The training data consist of 17532 T2-weighted (T2W) brain images of 136 healthy adult subjects from the human connectome project (http: //www.humanconnectome. org, 900 Subjects Data). Testing data consisted of twenty-four transverse T2W images of one normal subject, which were acquired on a 1.5 T MR scanner (OPTIMA MR360, GE, America) using the fast spin echo sequence with the following parameters: TR = 5900 ms, TE = 117 ms, field of view (FOV) = 240 × 240 mm2, pixel size = 0.47 × 0.47 mm2, slice thickness = 6.0 mm, and number of signal averages = 1.5.
We implanted the proposed CNN using the TensorFlow.10 In the training phase, multiply a randomly selected sampling pattern to the fully sampled data to obtain the artifact image, and then concatenating with the corresponding degradation maps as the input. The batch size is set to 4. The convolution parameters were randomly initialized from a Gaussian distribution with a standard deviation of 0.001, and the learning rate was set to 10-5. The root-mean-square error (RMSE), peak signal-to-noise ratio (PSNR) and structure similarity (SSIM) index were calculated for the quantitative evaluation.
Results
Figure 2 presents the results of Gibbs-ringing reduction on a representative T2W normal brain images with sampling levels of 70%，60% and 50% in phase-encoding direction. The proposed method effectively reduces the ringing artifact without noticeable blurring.
Table 1 presents the quantitative evaluation of the proposed method. The proposed method consistently generates images with improved RMSE, PSNR and SSIM values for T2W images.
Discussion & Conclusion
This work presents a novel and simple method that employs a single CNN as an end-to-end mapping that inputs the concatenated ringing MR image and degradation maps and outputs the artifact-free one to remove Gibbs-ring artifact from MR images under varying sampling levels. The results demonstrate that the proposed GRA-CNN method can effectively remove the Gibbs-ringing artifact without noticeable blurring.Acknowledgements
No acknowledgement found.References
1. Czervionke LF, Czervionke JM, Daniels DL, Haughton VM. Characteristic features of MR truncation artifacts. AJR Am J Roentgenol 1988;151(6):1219–1228.
2. Di Bella EV, Parker DL, Sinusas AJ. On the dark rim artifact in dynamic contrast-enhanced MRI myocardial perfusion studies. Magn Reson Med 2005;54(5):1295–1299.
3. Veraart, J., et al., Gibbs ringing in diffusion MRI. Magn Reson Med, 2015.
4. Parker DL, Gullberg GT, Frederick PR. Gibbs artifact removal in magnetic resonance imaging. Med Phys 1987;14(4):640–645.
5. Smith MR, Nichols ST, Henkelman RM, Wood ML. Application of autoregressive modelling in magnetic resonance imaging to remove noise and truncation artifacts. Magn Reson Med 1986;4(3):257–261.
6. Barone P, Sebastiani G. A new method of magnetic resonance image reconstruction with short acquisition time and truncation artifact reduction. IEEE Trans Med Imaging 1992;11(2):250–259.
7. Yan H, Mao J. Data truncation artifact reduction in MR imaging using a multilayer neural network. IEEE Trans Med Imaging 1993;12(1):73–77
8. Hui Y, Smith MR. Comments on "Data truncation artifact reduction in MR imaging using a multilayer neural network". IEEE Trans Med Imaging 1995;14(2):409–412.
9. K. Zhang, W. Zuo, L. Zhang. Learning a single convolutional super-resolution network for multiple degradations. In CVPR, 2018.
10. Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., ... & Kudlur, M. (2016, November). Tensorflow: a system for large-scale machine learning. In OSDI (Vol. 16, pp. 265–283).
Figures
