|Year : 2022 | Volume
| Issue : 4 | Page : 387-393
Denoising using Noise2Void for low-field magnetic resonance imaging: A phantom study
Shinya Kojima, Toshimune Ito, Tatsuya Hayashi
Department of Radiological Technology, Faculty of Medical Technology, Teikyo University, Itabashi-ku, Tokyo, Japan
|Date of Submission||04-Aug-2022|
|Date of Decision||14-Sep-2022|
|Date of Acceptance||26-Sep-2022|
|Date of Web Publication||10-Jan-2023|
Dr. Shinya Kojima
Department of Radiological Technology, Faculty of Medical Technology, Teikyo University, 2-11-1 Kaga, Itabashi-Ku, Tokyo, 173-8605
Source of Support: None, Conflict of Interest: None
| Abstract|| |
To reduce noise for low-field magnetic resonance imaging (MRI) using Noise2Void (N2V) and to demonstrate the N2V validity. N2V is one of the denoising convolutional neural network methods that allows the training of a model without a noiseless clean image. In this study, a kiwi fruit was scanned using a 0.35 Tesla MRI system, and the image qualities at pre- and postdenoising were evaluated. Structural similarity (SSIM), signal-to-noise ratio (SNR), and contrast ratio (CR) were measured, and visual assessment of noise and sharpness was observed. Both SSIM and SNR were significantly improved using N2V (P < 0.05). CR was unchanged between pre- and postdenoising images. The results of visual assessment for noise revealed higher scores in postdenoising images than that in predenoising images. The sharpness scores of postdenoising images were high when SNR was low. N2V provides effective noise reduction and is a useful denoising technique in low-field MRI.
Keywords: Denoising convolutional neural network, low-field magnetic resonance imaging, Noise2Void
|How to cite this article:|
Kojima S, Ito T, Hayashi T. Denoising using Noise2Void for low-field magnetic resonance imaging: A phantom study. J Med Phys 2022;47:387-93
|How to cite this URL:|
Kojima S, Ito T, Hayashi T. Denoising using Noise2Void for low-field magnetic resonance imaging: A phantom study. J Med Phys [serial online] 2022 [cited 2023 Feb 1];47:387-93. Available from: https://www.jmp.org.in/text.asp?2022/47/4/387/367427
| Introduction|| |
Recently, the advantages of low-field (<1.0 Tesla) magnetic resonance imaging (MRI) have been reevaluated and attracting attention:,,, reduced specific absorption rate (SAR), short T1 relaxation times, long T2* relaxation time, and reduced financial cost. Decreased SAR prevents heating of conductive devices or implants and contributes to the safety of MRI examination., In addition to decreased SAR, short T1 and long T2* allow more efficient and flexible use of pulse sequence., Typical low-field systems are equipped with a permanent magnet, without requiring cooling systems, large installation spaces, and low power consumption. To maximize these advantages in clinical practice, the low signal-to-noise ratio (SNR) should be overcome, the key drawback of low-field MRI.
Methods for improving SNR using deep learning based on the denoising convolutional neural network (DnCNN) have been reported.,,,,, These methods allow noise reduction; however, enormous images are required for training data and clean images without noise as the ground truth image. Magnetic resonance (MR) images with different contrasts are obtained by adjusting the scan parameters, and obtained images have various noise levels. To construct a robust DnCNN for images with different noise levels, numerous images acquired at various scan conditions are needed. However, preparing these image data is difficult, and acquiring noiseless images is almost impossible.
Noise2Void (N2V) has a novel DnCNN training scheme that overcomes the above-mentioned shortcomings, i.e., it does not require both enormous training images and clean ground truth images. Generally, multiple-slice images are acquired using MRI. Multiple-slice images have the same noise distribution; therefore, the training in the N2 is performed using these images, and the noise reduction is carried out. Although N2V is a feasible denoising technique, no reports applied N2V to MR images.
In this study, focusing on the convenience of both the low-field MRI system and N2V, the denoising effect is verified using N2V on MR images acquired from the low-field system.
| Materials and Methods|| |
This study was confirmed by the Institutional Review Broad of our facility.
Features of Noise2Void
The N2V network model can be trained without the clean ground truth image. The common convolutional neural network (CNN) takes one image as the input data and predicts another one as the output. In N2V, patches that segmented images are inputted to the CNN. The pixel in the output data from the CNN is influenced by all pixels in the patch. Thus, the CNN in N2V takes a patch that inputs and outputs a prediction for the single pixel located at the patch center. The true signal that reduced the noise is predicted using all pixels in the patch because of the similar noise distribution. Denoising of an entire image is achieved by extracting patches and supplying them to the CNN.
A blind-spot network is also one of the N2V features. The conventional network degenerates because of simply learning the identity when the training is conducted using the same noisy image as the input and target. To prevent network degeneration in the blind-spot network, the center pixel of the patch is excluded and predicted from the neighborhood pixel. Moreover, to efficiently implement a network, the pixel at the center of each input patch is replaced with a randomly selected pixel from the surrounding area.
N2V allows network training without requiring a large number of data because N2V is a self-supervised learning method, one of the N2V merits. Although enormous training data provides a robust DnCNN model, in cases of a small number of training data, the deterioration of the predicted image is caused by the influence of training images with various structures and textures. Therefore, the images in this study augmented the application of noise reduction, and the network training was conducted with the augmented data.
Magnetic resonance imaging
This study involved scanning of a kiwi fruit using the whole-body 0.35 Tesla MRI system (AIRIS Vento, Fuji Film Medical Co. Ltd.), and T1-weighted images (T1WI) and T2-weighted images (T2WI) were acquired. Kiwi fruit is suitable for evaluating spatial resolution and image contrast because of its fine structure and various components, and kiwi fruit was used as a prostate MRI phantom in a previous report., The scanning parameters were as follows: For T1WI, pulse sequence, spin-echo; repetition time, 450 ms; and echo time, 18.4 ms; and for T2WI, pulse sequence, fast spin-echo; repetition time 3000 ms; echo time, 90 ms; and echo train length, 5. The following parameters were the same for T1WI and T2WI: Matrix size, 256 × 256; field of view, 100 mm × 100 mm; slice thickness, 6 mm; and the number of slices, 8. For both T1WI and T2WI, a parallel imaging technique was not used, a single-channel knee coil was employed, and the number of excitations (NEX) was 1, 2, 4, 9, and 16. Images using the NEX of 16 were employed as the reference image, and the noise reduction by N2V was performed to each NEX image, excluding the reference image. Each scan was repeated ten times to avoid measurement errors.
Denoising flowchart based on Noise2Void
The denoising flowchart based on N2V is shown in [Figure 1]. The training was performed in 574 images augmented from eight images obtained on each slice location. During data augmentation, original images were rotated eight times by 45° and rotated images were translated in each of the eight directions: Up, down, left, right, and diagonal. Then, original images were inputted into the model after the training, and denoising on original images was implemented.
|Figure 1: Denoising flowchart by N2V in this study. First, 574 training images were obtained by the augmentation process, such as translation and rotation to eight slice images, and the network training model was performed using these augmented data. Then, the original eight slice images were inputted to the learned network model and the denoising was performed. N2V: Noise2Void|
Click here to view
The network architecture of N2V was as a U-net with the depth of 3, kernel size of 3, batch normalization, and a linear activation function in the last layer. The 32 feature maps on the initial layer were used and doubled as the layer got deeper.
The network model was trained for 100 epochs using the Adam optimizer, and a learning rate of 0.0004 was used. Other parameters were as follows: Batch size, 128; patch size, 24; and loss of function, mean square error. The training was run on Keras (version 2.3.1) with TensorFlow (version 1.14.0) using a workstation with an Intel Core i7-8700k 3.7 GHz and two Nvidia GeForce GTX 1080Ti.
In this study, quantitative and qualitative assessments were carried out to assess the image quality between pre and postprocessing in N2V.
Quantitative assessments were performed to estimate the image degradation structural similarity (SSIM) index:
where μx and μy are the local mean values of each NEX and reference image, [INSIDE:1] and [INSIDE:2] represent their respective variations, σxy is the covariance of each NEX and reference image, and C1 and C2 are the constant values used to avoid instability. In the SSIM assessments, the image with 16 NEX was used as the reference image, and the images with 1, 2, 3, 4, and 9 NEX were used as the evaluation images.
Furthermore, the region of interest (ROI) was measured, and the SNR and contrast ratio (CR) were calculated using the following formula:
where SIa and SIb are the average signal intensity within the ROIa and ROIb, respectively, and standard deviation (SD) is the SD of ROIBG. ROIa and ROIb were set on the center and periphery of a kiwi, respectively. ROIBG was positioned on background [Figure 2]. In addition, spatial frequency analysis was conducted using a radial direction intensity function (RDIF) acquired from a two-dimensional power spectrum.
|Figure 2: ROI positions. ROIa and ROIb were set at the center and peripheral of kiwi fruit, respectively. ROIBG was placed on the background region. ROI: Region of interest|
Click here to view
For qualitative assessments, two independent readers (TH and TI) with >10 years of experience as radiologic technologists visually evaluated the noise and sharpness of images using a 4-point scale. The criteria of the noise scales were as follows: 0, severe image deterioration due to noise; 1, conspicuous noise; 2, slight noise; and 3, no noise. Sharpness was scored as follows: 0, unrecognizable structure, 1, obscure structure; 2, a little unclear structure; and 3, distinct structure.
In the SSIM measurement, all slice images were used. The SNR and CR were investigated on four images near the center of the slice position, and a fifth slice image was employed for visual assessment.
In all assessments, the Wilcoxon signed-rank test was used to compare the image quality between pre and postprocessing using N2V, and a P < 0.05 was considered a statistically significant difference. Moreover, in CR measurements, the equivalence test was carried out. About 10% of the mean CR in the images without noise reduction was defined as the confidence interval, and a P < 0.05 was considered statistical equivalence of the image contrast. A JMP pro 16.0 (SAS Institute, Cary, NC, USA) was used for these analyses.
| Results|| |
Results of SSIMs on T1WI and T2WI are shown in [Figure 3]. In both T1WI and T2WI, SSIMs increased with NEX. From 1 NEX to 4 NEX, SSIMs in the postdenoising images were higher than that in the predenoising images; however, these values were reversed in 9 NEX. This SSIM tendency was similar in both T1WI and T2WI. A significant difference in the SSIM was observed between pre and postdenoising images in each NEX on both T1WI and T2WI (P < 0.001).
|Figure 3: SSIM of T1WI (a) and T2WI (b) at each NEX image. SSIM was high in the postdenoising images than in the predenoising images for both T1WI and T2WI. SSIM: Structural similarity, T1W1: T1 weighted image, T2WI: T2 weighted image, NEX: Number of excitations|
Click here to view
[Figure 4] and [Figure 5] show the results of the SNR and CR, respectively. In both T1WI and T2WI, the SNRs of postdenoising images were significantly improved (P < 0.001), and the postdenoising images >4 NEX had higher SNR than the reference images (16 NEX images). In CR measurements, no significant difference was observed between pre and postdenoising images on both T1WI and T2WI. Furthermore, the equivalence test indicated that CRs of pre and postdenoising images were statistically equivalent.
|Figure 4: SNR of T1WI (a) and T2WI (b) at each NEX value. The red, blue, and yellow boxes present the predenoising, postdenoising, and reference images (16 NEX), respectively. The SNRs of postdenoising images were higher than those of predenoising images on both T1WI and T2WI. SNR: Signal-to-noise ratio, T1W1: T1-weighted image, T2WI: T2-weighted image, NEX: Number of excitations|
Click here to view
|Figure 5: CR of T1WI (a) and T2WI (b) for each NEX value. The red, blue, and yellow boxes present the predenoising, post, image and reference images (16 NEX), respectively. The CRs of postdenoising images were similar to those of predenoising images on both T1WI and T2WI. CR: Contrast ratio, T1W1: T1-weighted image, T2WI: T2-weighted image, NEX: Number of excitations|
Click here to view
[Table 1] shows the results of the visual assessment based on noise in T1WI and T2WI. Scores of the postdenoising images were significantly high (P < 0.001), which greatly increased especially in the case of small NEX. A comparison of the reference image score and postdenoising images showed that postdenoising images had higher scores than the reference image, except for the 1 NEX image in T1WI. The scores based on sharpness are represented in [Table 2]. However, postdenoising images provided higher scores than predenoising images, but without significant differences in almost NEX. Particularly, in the T1WI, the score of postdenoising images that 9 NEX decreased rather than that of predenoising images. When comparing the reference and denoising images, the score of the postdenoising images in all NEX could not exceed that of the reference image.
|Table 1: Results of visual assessment based on noise in T1-weighted images and T2-weighted images|
Click here to view
|Table 2: Results of visual assessment based on the sharpness in T1-weighted images and T2-weighted images|
Click here to view
The RDIFs of the postdenoising and reference images (T1WI and T2WI) are presented in [Figure 6]. On the T1WI, the RDIFs decreased as NEX increased in the high spatial frequency region. In comparison to the reference image, the RDIF of each NEX image declined in the high spatial frequency domain except when NEX = 1. On the T2WI, similar tendency were observed for all NEX values, and the RDIFs of postdenoising images were lower than that of the reference image in the high-frequency domain.
|Figure 6: RDIF of denoising images on T1WI (a) and T2WI (b) at each NEX value. The yellow, blue, green, black, and gray lines denote NEXs of 1, 2, 3, 4, and 9, respectively, and the red dotted line represents the reference image (16 NEX). The vertical axis is on a logarithmic scale. RDIF: radial direction intensity function, T1W1: T1-weighted image, T2WI: T2-weighted image, NEX: Number of excitations|
Click here to view
The pre and postdenoising images in each NEX of T1WI and T2WI are represented in [Figure 7]. In both T1WI and T2WI, the noise was reduced on postdenoising images efficiently, and the effect of noise reduction was especially remarkable in the case of small NEX. Although noisy images were with 1 or 2 NEX, noise reduction more clarified the image details; whereas in images with 9 NEX, the depiction of the minute structure was slightly blurred on postdenoising images.
|Figure 7: Pre and postdenoising images at each NEX in T1WI and T2WI. In both T1WI and T2WI, the upper and lower rows represented pre and postdenoising images, respectively. In 16 NEX images, noise reduction using N2V was not applied because these images were used as the reference image. NEX: Number of excitations, T1W1: T1 weighted image, T2WI: T2 weighted image, N2V: Noise2Void|
Click here to view
| Discussion|| |
In this study, the network training data set was produced from only eight images that themselves were subjected to noise reduction. As the results of this study, the image quality was improved by N2V using a small data set. One of the merits of N2V is that network training can be performed using a small data set. This advantage is meaningful because noise reduction can be performed using only scanned images. In particular, because images with a low-field system tend to low SNR, noise reduction with N2V is more effective.
As the results of SSIMs, SSIMs of the postdenosing images improved as the NEX increased. However, the degree of SSIM improvement decreased in the case of large NEX, and the SSIMs of postdenoising images was lower than that of predenoising images in 9 NEX. This is because in the image with sufficient SNR, namely with little noise, the effect of N2V is predominantly to blur the fine structure of object rather than noise reduction. Since SSIM compares the structure with the reference image, SSIM improves due to noise reduction by the N2V in the case of noisy image, however when there is little noise on image SSIM plateaus without improving due to blurring. Therefore, the noise reduction by N2V should not be performed to images with sufficient SNR.
Results of the SNR measurement and visual assessment of the noise indicated that the noise reduction using N2V significantly improves the SNR. Although this result is similar to that of previous studies, N2V effectively removed the noise for images obtained at the low-field MRI system. Moreover, the image contrast between pre and postdenoising using N2V was remarkably invariable as shown in the results of CR measurement. Preventing variation of the image contrast during the denoising process is important in maintaining the diagnostic capability. Therefore, adapting N2V to the image at low-field MRI with disadvantages of relatively lower SNR than images at high-field MRI is useful in clinical practice.
In spatial frequency analysis, the RDIFs of T1WI and T2WI in the high spatial frequency region displayed different tendencies. Although the RDIF on T1WI decreased as NEX increased, that on T2WI decreased regardless of NEX. This is because T2WI has higher contrast than T1WI, and same degree of denoising was obtained independent of SNR in T2WI. This means that N2V provides higher noise reduction in images with high contrast.
When the SNR was low, such as from 1 NEX to 4 NEX, both the SSIM and visual assessment scores of sharpness in postdenoising images were superior to that in predenoising images because removing noise using N2V provided a clear depiction of the detail. However, in some extent of high SNR, such as 9 NEX, the blurring slightly occurred and the spatial resolution was degraded. During the noise reduction process, the use of CNN to prevent image deterioration due to blurring is significant. The super-resolution CNN (SRCNN) method is a technique that improves the spatial resolution with CNN. This method includes the following CNN processes: First, patches are extracted as future maps from low-resolution images, and each patch represents a high-dimensional vector. High-resolution images are reconstructed from each high-dimensional vector using nonlinear operation. We believe that the combination of N2V and SRCNN helps prevent blurring, and investigation of this technique is one of the meaningful future tasks.
Similar to N2V, the Noise2Noise (N2N) method does not require the ground truth image without noise. Instead, of the ground truth image, pairs of independently noisy images with the same probability distribution are employed in the N2N method training. Although the N2N method allows the practical noise reduction, a shortcoming is that the acquisition of such pairs is difficult in clinical practice. For this reason, N2V was adopted to remove the noise in this study.
One of the limitations of this study is that the investigation with the clinical data has not been performed. Although we expect that N2V also provides noise reduction of the clinical data, the evaluation of denoising effects using N2V to the clinical data is a priority task. Another limitation is the investigation of spatial resolution. In our results, degradation of the spatial resolution in postdenoising images was caused by N2V, when the SNR was high. Therefore, a more detailed measurement of spatial resolution, such as a modulation transfer function, is necessary because spatial resolution is an important factor to assess image quality.
In this study, N2V-based training was executed without both enormous training data and noiseless clean data, and the N2V effects were verified. Consequently, although this was an initial study using a phantom, N2V allows effective noise reduction for images obtained using low-field MRI, especially in the case of low SNR.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Marques JP, Simonis FF, Webb AG. Low-field MRI: An MR physics perspective. J Magn Reson Imaging 2019;49:1528-42.
Campbell-Washburn AE, Ramasawmy R, Restivo MC, Bhattacharya I, Basar B, Herzka DA, et al
. Opportunities in interventional and diagnostic imaging by using high-performance low-field-strength MRI. Radiology 2019;293:384-93.
Bhat SS, Fernandes TT, Poojar P, da Silva Ferreira M, Rao PC, Hanumantharaju MC, et al
. Low-field MRI of stroke: Challenges and opportunities. J Magn Reson Imaging 2021;54:372-90.
Rashid S, Han F, Gao Y, Sung K, Cao M, Yang Y, et al
. Cardiac balanced steady-state free precession MRI at 0.35 T: A comparison study with 1.5 T. Quant Imaging Med Surg 2018;8:627-36.
Tripathi PC, Bag S. CNN-DMRI: A convolutional neural network for denoising of magnetic resonance images. Pattern Recognit Lett 2020;135:57-63.
Manjón JV, Coupe P. MRI denoising using deep learning and non-local averaging. arXiv 2019:1911.04798v2.
Jiang D, Dou W, Vosters L, Xu X, Sun Y, Tan T. Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network. Jpn J Radiol 2018;36:566-74.
Li S, Zhou J, Liang D, Liu Q. MRI denoising using progressively distribution-based neural network. Magn Reson Imaging 2020;71:55-68.
Ueda T, Ohno Y, Yamamoto K, Iwase A, Fukuba T, Hanamatsu S, et al
. Compressed sensing and deep learning reconstruction for women's pelvic MRI denoising: Utility for improving image quality and examination time in routine clinical practice. Eur J Radiol 2021;134:109430.
Kidoh M, Shinoda K, Kitajima M, Isogawa K, Nambu M, Uetani H, et al
. Deep learning based noise reduction for brain MR imaging: Tests on phantoms and healthy volunteers. Magn Reson Med Sci 2020;19:195-206.
Krull A, Buchholz TO, Jug F. Noise2void-Learning Denoising from Single Noisy Images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2129-2137; 2019.
Mueller-Lisse UG, Murer S, Mueller-Lisse UL, Kuhn M, Scheidler J, Scherr M. Everyman's prostate phantom: Kiwi-fruit substitute for human prostates at magnetic resonance imaging, diffusion-weighted imaging and magnetic resonance spectroscopy. Eur Radiol 2017;27:3362-71.
Murer S, Scheidler J, Mueller-Lisse UL, Helling M, Scherr M, Mueller-Lisse UG. Two-centre comparative experimental study of biparametric MRI at 3.0 T with and without endorectal coil using kiwifruit (Actinidia deliciosa
) as a phantom for human prostate. Eur Radiol Exp 2019;3:30.
Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells W, Frangi A, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Vol. 9351. Cham: Springer; 2015. p. 234-41.
Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv 2014:1412.6980v9.
Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: From error visibility to structural similarity. IEEE Trans Image Process 2004;13:600-12.
Ito T, Matsusaka Y, Onoguchi M, Ichikawa H, Okuda K, Shibutani T, et al
. Experimental evaluation of the GE NM/CT 870 CZT clinical SPECT system equipped with WEHR and MEHRS collimator. J Appl Clin Med Phys 2021;22:165-77.
Dong C, Loy CC, He K, Tang X. Image super-resolution using deep convolutional networks. IEEE Trans Pattern Anal Mach Intell 2016;38:295-307.
Lehtinen J, Munkberg J, Hasselgren J, Laine S, Karras T, Aittala M, et al. Noise2noise: Learning image restoration without clean data. arXiv 2018:1803.04189v3.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7]
[Table 1], [Table 2]