Introduction

Many reports have addressed deep-learning-based noise removal^1-3, but most of them are simple slice-by-slice 2D image transformations, such as converting a slice image without averaging (number of acquisitions=1 [NOA1]) to an image-equivalent of an image obtained by averaging multiple acquisitions.
However, in a body diffusion-weighted image (DWI) with low signal-to-noise ratio (SNR), it is not difficult to apply this method directly because NOA1 often has insufficient information to reconstruct the appropriate image. This problem may be solved by using more than one acquisition for the input; however, such images tend to appear smoothened because of the motion-related-mismatch between acquisitions, which eventually blurs the small structures. In this study, we developed a neural network that utilizes 4D convolution to denoise the input image series. The concept was to incorporate all relevant information for each pixel, including the neighboring pixels in 3D (i.e. including the adjacent slices), and the same 3D area in another acquisition, instead of referring to only the neighboring pixels in the same slice in the same acquisition as done for the usual 2D convolution.
This study aimed to evaluate the usefulness of this novel strategy.

Methods

A total of 187 cases that included body DWI scans as part of the tumor screening, were extracted from the clinical database. The subject population and major DWI scanning parameters are summarized in Figure 1.
The outline of the proposed network is shown in Figure 2. The task given to this network was to estimate NOA5 as denoised images from the NOA2 equivalent data input that was derived from the original NOA5. The proposed network was designed to accept three adjacent slices and two acquisitions simultaneously and applied both 4D and 2D convolutions in combination. The control network (Figure 3) had a design similar to that of the proposed network, except that it accepted each slice and acquisition as separate inputs, and the results of different acquisitions were averaged at the end. Therefore, all convolutions were performed in 2D.
The subjects were randomly divided into 99 training and 38 testing cases. The networks were trained using the training cases: optimiser = Adam⁴ (lr=0.001, beta=0.5), epoch=50, and loss= mean absolute error. The Chainer⁵ platform was used for all deep-learning procedures.
The trained networks were tested using test cases. The denoised series obtained via the proposed network (4D-conv), control network (2D-conv), and NOA2 were compared both visually and numerically. For numerical comparisons, the mean absolute difference (MAE) and structural similarity index measure for DWI (b=800) were obtained for each case in comparison with NOA5 (MAE_DWI and SSIM). In addition, ADC maps were calculated for 4D-conv, 2D-conv, and NOA2, and MAE were obtained similarly (MAE_ADC). All indices were statistically compared between the noise-reduction methods using the Wilcoxon signed-rank test (P<.05 was considered significant).

Results

Examples of the denoised images are shown in Figure 4. Visually, the appearance of 4D-conv was closer to NOA5 compared to 2D-conv and NOA2.
The results of the statistical comparisons are shown in Figures 5. The MAE_DWI was significantly smaller in 4D-conv and 2D-conv than in NOA2 (P<.001), but the difference between 4D-conv and 2D-conv was not significant (P=0.56). The SSIM was significantly higher, and MAE_ADC was significantly smaller in 4D-conv than in 2D-conv and NOA2 (P<.001). However, the SSIM was significantly lower, and MAE_ADC was significantly higher in 2D-conv than in NOA2 (P<.001).

Discussion

In this study, we proposed a new neural network to adequately reduce noise from NOA2 input in low-SNR body DWI. The major underlying concept was to utilize 4D convolution to incorporate all relevant information for calculating the output value for a certain pixel. However, repeating many 4D convolutions was difficult because of the enormous memory consumption involved, and was applied only once to create the output. Instead, the pathway including repeated 2D convolutions with softmax activation at the end was merged with this layer so that the large optimized receptive field would be available, similar to our previous study⁶.4D-conv could not decrease MAE_DWI compared to 2D-conv, but successfully increased the SSIM. This could be because the proposed strategy was better in optimizing the convolution to fit the specific situation for each pixel, and thus the blurring was suppressed well. Furthermore, the fact that MAE_ADC was significantly smaller in 4D-conv compared to 2D-conv and NOA2 suggests that 4D-conv did not only make the DWI appearance closer to NOA5 but was appropriate to preserve the underlying diffusion information. However, the low SSIM and high MAE_ADC in 2D-conv supports the issue described in the introduction that blurring may occur when such a method is used.
As a limitation, this study only focused on the overall image quality and did not evaluate the 4D-conv’s ability to delineate individual lesions.

Conclusion

The proposed method using 4D convolution may be useful for denoising body DWI with low SNR because it can denoise images adequately without blurring.

Acknowledgements

This research was supported by a Grant-in-Aid for Scientific Research (Kakenhi #17K10385) from the Japan Society for the Promotion of Science (JSPS), and by QST President’s Strategic Grant QST Advanced Study Laboratory.

References

1. Kidoh, M., Shinoda, K., Kitajima, M., et al. Deep Learning Based Noise Reduction for Brain MR Imaging: Tests on Phantoms and Healthy Volunteers (2020) Magn Reson Med Sci. 19:195-206

2. Jose V. Manjon and Pierrick Coupe. MRI denoising using Deep Learning and Non-local averaging. (2019) arXiv:1911.04798 (https://arxiv.org/abs/1911.04798) [accessed Nov.10, 2021]

3. Koonjoo, N., Zhu, B., Bagnall, G. Cody, et al. Boosting the signal-to-noise of low-field MRI with deep learning image reconstruction. (2021) Sci Rep. 11:8248

4. Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization (2014) arXiv:1412.6980 (https://arxiv.org/abs/1412.6980) [accessed Nov.10, 2021]

5. Tokui, S., Okuta, R., Akiba, T., et al. Chainer: A Deep Learning Framework for Accelerating the Research Cycle. (2019) arXiv:1908.00213 (https://arxiv.org/abs/1908.00213) [accessed Nov.10, 2021]

6. Nozaki, H., Tachibana, Y., Otsuka, Y., et al. Deep learning-based DWI Denoising method that suppressed the "instability" problem. (2021) Proceedings of ISMRM 2021 #2442.