Introduction

Low-field MRI can promote accessibility of MRI scanners due to its potential low-cost. Additionally, compared with widely used high-field MRI, low-field systems benefit from reduced SAR and implants heating, and fewer susceptibility-induced artifacts^[1,2]. However, the image SNR decreases as the field strength becomes lower^[1,3]. Thus, a robust denoising method for low-field MRI is needed to maintain practical acquisition efficiency.
Recently, deep learning has been introduced to image denoising^[4-10]. Among them, methods with convolutional neural network like DnCNN^[4] have achieved competitive results. Meanwhile, multi-contrast MRI images share similar structures and thus provide additional information for joint denoising. As such, we investigate a deep learning based denoising method to improve SNRs of multi-contrast low-field MRI images jointly, and use a Noise2Noise^[11] strategy for training without noise-free targets.

Method

Sparse-Net and Joint-Sparse-Net
Neural network is widely used on image denoising. Meanwhile, model-driven deep learning networks, which exploit the sparsity of input images with iterative transform blocks, have shown great potential in image reconstruction^[12] and denoising^[6]. Thus, we develop a cascade model for low-field MRI images denoising, named Sparse-Net, as shown in Figure 1A. Several denoising blocks simulate an iterative denoising procedure. In each block, the general update step can be described as:
$$ x_{k+1}=DB(x_k)=x_k+\widehat{T}_k(\tau_k(T_k(x_k,DC(x_k)))) $$
Here $$$DC$$$ stands for data consistency. $$$T$$$ and $$$\widehat{T}$$$ are sparse transform consisting of 2D convolutions. Noisy images are filtered in transform domain using learnable soft thresholding function $$$\tau$$$.
To expand our model to multi-contrast images, we adopt the traditional group sparsity concept^[13,14]. The group sparsity of multi-contrast images $$$x_{k,i}$$$ with sparse transform $$$T$$$ is
$$ T_{Group}(x_k)=\sqrt{\sum_i(T(x_{k,i}))^2} $$
Thus, the update step of multi-contrast denoising model can be described as $$$ x_{k+1,i} = DB(x_{k,i})+\lambda_iDB\_G(x_k) $$$, where $$ DB\_G(x_k) = \widehat{T}_{k,Group}(\tau_k(T_{k,Group}(x_k,DC(x_k)))) $$
The model with group sparsity transform is denoted as Joint-Sparse-Net, and its structure is shown in Figure 1B. All proposed models are trained using a combination of Multiscale-Structural Similarity (MSSIM) and $$$L_1$$$ as the loss function:$$ Loss=0.84MSSIM(t-x)+(1-0.84)\parallel t-x\Vert_1 $$Where $$$t$$$ and $$$x$$$ denotes target and input. We also implement DnCNN^[4] trained on single contrast data for comparison.
Noise2Noise
Most deep learning denoising models use averaged images as noise-free targets. However, low-field MRI images acquired with up to 6 averages still contain considerable amount of noise. In view of that, we employ the Noise2Noise training strategy^[11]. According to the Noise2Noise theory, considering that the majority of the noise in MRI images has zero means, using a pair of noisy images from two separate acquisitions as input and target can theoretically produce results equal to using noise-free images as targets. Models using Noise2Noise are named as Sparse-Net-N2N and Joint-Sparse-N2N.
Data Acquisition
Low-field MRI data were acquired on a prototype 0.5T MR Scanner (Marvel Stone Healthcare Co., Ltd., Wuxi, China), including 2D T1w and T2w images from 17 subjects. 13 subjects are chosen randomly for training, while 2 for validation and 2 for testing. Each subject has 14 slices of T1w and T2w images with 6 NSAs and 4 NSAs. The single average images are used as noisy image sets for denoising. Averages of T1w and T2w images are used as targets for non-N2N model training. Images from different NSAs are grouped as noisy image pairs for N2N training. The resolution is 0.7$$$\times$$$0.7$$$\times$$$5mm3.

Result

Figure 2 shows the denoising results from DnCNN, Sparse-Net and Joint-Sparse-Net for one set of T1w images. Notice that the result from Joint-Sparse-Net comes from jointly denoising the T1w and T2w images of the same subject. All methods can produce comparable results to the NSA6 reference images. Compared to DnCNN, Sparse-Net shows an advantage in reducing noise, proving the effectiveness of the proposed sparsity-based denoising model. Meanwhile, Joint-Sparse-Net provides sharper edges and structures than Sparse-Net. Figure 3 shows the PSNR and SSIM values for DnCNN, Sparse-Net and Joint-Sparse-Net.
The results from Sparse-Net, Joint-Sparse-Net and their N2N-version are presented in Figure 4. Denoising results using the Noise2Noise strategy have a lower noise level compared with the reference NSA6 images, while maintaining the sharpness of edges and structures. Of all methods, the proposed Joint-Sparse-N2N model has the best SNR and preserves the most details, while Sparse-Net-N2N produces over-smoothed results. Figure 5 shows the results from another subject. The finding is consistent with that from Figure 4.

Conclusion and Discussion

In this work, we develop a deep learning denoising model, namely Joint-Sparse-N2N, for multi-contrast low-field MRI images. Experiments demonstrate that the proposed method can suppress noise while preserving detailed information of anatomical structures. We expect a promotion of the proposed method with larger training dataset. Further study on manually controlling the denoising levels will be explored in the future.

Acknowledgements

No acknowledgement found.