3392

Noise2Noise MRI for High-resolution Diffusion-weighted Imaging of the Brain: Deep Learning-based denoising without need for Highly Averaged Ground-truth Images
Motohide Kawamura1, Daiki Tamada1, Satoshi Funayama1, Hiroshi Onishi1, and Utaroh Motosugi1

1Department of Radiology, University of Yamanashi, Chuo-shi, Japan

Synopsis

Deep learning (DL)-based denoising is promising to achieve high resolution diffusion-weighted imaging (HR-DWI) by improving SNR without signal averaging. Training supervised DL-based algorithm, however, requires thousands of teaching data, which need long acquisition time. In this study, we propose to use noise2noise (N2N) theory to develop DL-based denoising algorithm, which does not need teaching data with high SNR. In the results, the proposed method (N2N-MRI-based algorithm) outperformed conventional ground-truth-based algorithm in terms of maximum peak SNRs on validation sets during training. The image quality of HR-DWI denoised by N2N-MRI-based algorithm was equivalent to that denoised by conventional algorithm.

Introduction

Diffusion-weighted MRI (DWI) has been widely used for brain MRI given its high contrast resolution, especially for the early detection of acute stroke.1 Recent advances in multi-shot echo-planar imaging (EPI) including navigator-based reacquisition 2 and multiplexed sensitivity encoding 3 enable less image distortion and higher spatial resolution compared to conventional single-shot techniques. However, high resolution DWI (HR-DWI) with multi-shot EPI suffers from a limited signal-to-noise ratio (SNR) in a small voxel. Degraded image quality can be made up with long acquisition times for multiple signal averaging, but it is not a practical approach. To accelerate HR-DWI acquisition, denoising algorithms can be employed instead of time-consuming signal averaging, notably reducing the scan durations. Deep learning (DL)-based denoising algorithm is a promising approach to achieve outstanding denoising performance. One challenge of DL-based denoising is that network training requires thousands of ground-truth images as training inputs, which are scanned with exceptionally long acquisition time. So, it is often difficult to obtain a sufficient number of ground-truth images for training. Here we propose a novel ground-truth-free approach, called Noise2Noise (N2N) MRI, based on the N2N theory 4, in which networks learn without clean targets and achieve the same performance as those learning with clean targets. In this study, we compared the denoising performance of N2N-MRI with that of conventional ground-truth-based algorithm learning with the high SNR images acquired by 8 times averaging.

Methods

This study was approved by the institutional review board. DWI of the brain were acquired from eight healthy volunteers on a 3 Tesla MRI scanner (SIGNA Premier, GE Healthcare, Chicago, IL, USA) using a 48-channel head coil. For each volunteer, 24 axial slices were acquired using two-shot EPI sequence3 combined with parallel imaging (reduction factor = 2). Other MR parameters were as follows: slice thickness = 5 mm, TR/TE = 7000/74 ms, matrix = 320 × 320, b = 1000 s/mm2, and number of excitations (NEX) = 9. We adopted a deep convolutional neural network5 for denoising diffusion-weighted images. Data from two volunteers were used for training and from the other six for validation. In ground-truth-based learning, images with NEX 8 were regarded as clean targets and images with NEX 1 as pre-denoising inputs. In N2N-MRI, we used images with NEX 1 both as inputs and as targets and the remaining 7 NEX was not used for training. A schematic of both approaches is shown in Fig. 1. After data augmentation, training inputs are approximately 300,000 patches with size 40 × 40. Because infinitely averaged images are considered to be noise-free, noise is zero-mean. Thus, we used the L2 loss for training.4 Learning rates range from 10-3 to 10-6. The other parameters were as follows: network depth = 17, mini-batch size = 128, and epoch = 50. To evaluate the effectiveness of N2N-MRI, we plotted peak SNRs on the validation images of both methods as a function of training epoch.

Results

The learning curves are shown in Fig. 2. At learning rates of 10-3 and 3 × 10-4, maximum peak SNRs of N2N-MRI were comparable to those of the ground-truth-based learning. At the other learning rates, N2N-MRI clearly showed better maximum peak SNRs than the ground-truth-based learning. Figure 3 shows images from the validation set, that are denoised by ground-truth-based (conventional) algorithm (Fig.3B) and proposed N2N-MRI approach (Fig.3C).

Discussion

We proposed DL-based denoising algorithm called N2N-MRI. N2N-MRI does not require clean targets or ground-truth, which are difficult to obtain in clinical practice due to long scan times. Our N2N-MRI outperformed the ground-truth-based, or conventional learning, in terms of optimal peak SNRs during training. More learning should be necessary to establish the denoising algorithm. The acquisition time to obtain one training data for N2N-MRI was 1 min 59 s (2 NEX), which is much shorter than that for ground-truth-based learning (typically 9 – 10 NEX) 6,7. It is short enough to obtain training data in clinical situations, suggesting the feasibility of training deep neural networks with clinical images, i.e. images of patients with diseases. Although the theory states that denoising performance of N2N is equivalent to that of the conventional learning using clean targets, our results showed that N2N worked better in some cases. This unexpected behavior may be explained by reduced motion of the subject during the acquisition.

Conclusion

The proposed N2N-MRI is a promising approach to develop DL-based denoising algorithm without need for ground-truth data with high SNR. Our approach is helpful to collect training data and develop practically useful DL-based acceleration for high-resolution DWI.

Acknowledgements

No acknowledgement found.

References

  1. Moseley ME, Kucharczyk J, Mintorovitch J, et al. Diffusion-weighted MR imaging of acute stroke: correlation with T2-weighted and magnetic susceptibility-enhanced MR imaging in cats. AJNR Am J Neuroradiol. 1990;11(3):423–429.
  2. Porter DA, Heidemann RM. High resolution diffusion-weighted imaging using readout-segmented echo-planar imaging, parallel imaging and a two-dimensional navigator-based reacquisition. Magn Reson Med. 2009;62(2):468–475.
  3. Chen NK, Guidon A, Chang HC, Song AW. A robust multi-shot scan strategy for high-resolution diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE). Neuroimage. 2013;72:41–47.
  4. Lehtinen J, Munkberg J, Hasselgren J, et al. Noise2Noise: Learning image restoration without clean data. In Proceedings of the 35th International Conference on Machine Learning, Stockholm, Sweden, 2018;2965–2974.
  5. Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian denoiser: residual learning of deep CNN for image denoising. IEEE Trans Image Process. 2017;26(7):3142–3155.
  6. Kawamura M, Tamada D, Kromrey ML, et al. High resolution DWI of the brain using multi-shot EPI sequence: a study on accelerated acquisition with deep learning. In Proceedings of the 46th Annual meeting of the JSMRM, Kanazawa, Japan, 2018;197.
  7. Kitajima M, Kidoh M, Uetani H, et al. Deep Learning Convolutional Neural Network Reconstruction for Noise Reduction of High-Resolution MR Imaging: A Feasibility Study. In Proceedings of the 46th Annual meeting of the JSMRM, Kanazawa, Japan, 2018;252.

Figures

Schemas of the proposed Noise2Noise (N2N) MRI approach and the ground-truth-based learning with highly averaged images. In N2N MRI approach, the training was performed with sets of low SNR images (number of excitations [NEX] of 1). In our experiment, highly averaged ground-truth was reconstructed with NEX of 8. N2N theory guarantees the same network parameters for both methods.

Learning curves of the proposed Noise2Noise (N2N) MRI and the ground-truth-based approach at learning rates of 10-3, 3 × 10-4, 10-4, 3 × 10-5, 10-5, 3 × 10-6 and 10-6. Peak signal-to-noise ratios (PSNRs) on validation images are plotted against training epochs. Images with number of excitations of 8 were assumed to be noise-free for calculation of PSNRs.

Pre-denoising, denoised, and ground truth diffusion-weighted images. (A) Pre-denoising image. Images denoised using (B) the ground-truth-based method and (C) the proposed Noise2Noise (N2N) MRI. (D) Ground truth image. (E)–(H) Magnified images of the solid boxes in (A)–(D), respectively.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
3392