A low-field magnetic resonance image super-resolution network

Abstract

The application discloses a low-field magnetic resonance image super-resolution network and relates to the field of magnetic resonance imaging. The application introduces a multi-level attention mechanism and a stable wavelet transform to perform image super-resolution in view of the problem of insufficient resolution of a low-field (Low-Field) magnetic resonance imaging (MRI) device. The low-field magnetic resonance image is firstly subjected to a shallow feature extraction network to extract shallow features, then is subjected to a U-shaped feature extraction network to extract deep-level features of local details, the features are subjected to a network containing an attention mechanism to perform feature integration, are subjected to pixel rearrangement up-sampling, and meanwhile, the original image structure is learned by using a residual design, and finally, high-resolution output is obtained. The application effectively improves the definition of the low-field magnetic resonance image, solves the problem of blurring in the high-frequency region of the traditional method, is superior in image quality indexes such as a peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and can realize efficient image reconstruction at a relatively low calculation cost.

Description

TECHNICAL FIELD

[0001] The application relates to the field of magnetic resonance imaging, in particular to a low-field magnetic resonance image super-resolution network. BACKGROUND

[0002] Magnetic resonance imaging (MRI) is an important medical imaging method, which has the advantages of non-ionizing, non-invasive and multi-parameter compared with other imaging methods. However, the superconducting MRI scanner (mainly 1.5T and 3T) under the existing standard is expensive, and needs to be installed in a place with magnetic shielding equipment, which limits the use of MRI by more people. The ultra-low field MRI (strength <0.1T) device may be able to enter the hospitals in low-income countries due to its low cost and strong flexibility, and to diagnose low-income groups. However, the resolution of the current low-field magnetic resonance image is low, and it is difficult to be directly used for clinical diagnosis. In order to improve the signal-to-noise ratio, the low-field magnetic resonance image is super-resolved.

[0003] Therefore, the technical personnel in the art are committed to developing a low-field magnetic resonance image super-resolution network. By introducing a multi-level attention mechanism and a stable wavelet transform, the clarity of the low-field magnetic resonance image is effectively improved, and the problem of blur in the high-frequency region of the traditional method is solved. SUMMARY

[0004] In view of the above defects of the prior art, the technical problem to be solved by the application is the insufficient resolution of the low-field MRI device, and the problem of low resolution of the low-field magnetic resonance image, especially the blur in the high-frequency region.

[0005] To achieve the above-mentioned purpose, the application provides a low-field magnetic resonance image super-resolution network, which comprises a multi-level attention mechanism module and a loss function based on a stable wavelet transform.

[0006] Further, the multi-level attention mechanism module extracts features of different granularities of the magnetic resonance image, and fuses and enhances the feature representation by using the attention mechanism.

[0007] Further, in the loss function based on the stable wavelet transform, the stable wavelet transform maintains the translational invariance of the signal at each scale by eliminating the downsampling step.

[0008] Further, the stable wavelet transform only performs filtering operation at each level of decomposition, and does not perform downsampling.

[0009] Further, a 3D network is used for training.

[0010] Further, the low-field magnetic resonance image is firstly extracted by a shallow feature extraction network to extract shallow features, then a U-shaped feature extraction network is used to extract deep features, the features are integrated through an attention mechanism module, and up-sampling is performed, and the output is obtained by using residual connection.

[0011] Further, a residual multi-level attention group module is further included, and the input is subjected to detailed feature extraction by the multi-level attention mechanism module, corresponding features are obtained by convolution, and the corresponding output is obtained by combining the original input through residual connection.

[0012] Further, the multi-level attention mechanism module integrates the input features by spatial attention and channel attention, extracts spatial and channel features, and finally combines the input through residual connection.

[0013] Further, the spatial attention part splits the channel into three features, which are respectively convolved by three different size convolution kernels to obtain spatial information of different levels, and after merging to obtain multi-level spatial features, high-order information is extracted by using maximum pooling and average pooling, and the attention weight is obtained by using a sigmoid activation function, and the weighted spatial feature information is obtained by multiplying the multi-level spatial feature information.

[0014] Further, the channel attention part obtains the corresponding channel attention weight after the average pooling, two convolution layers and a LeakyReLU activation layer, and then obtains the final overall feature by multiplying the spatial feature information.

[0015] The current low-field magnetic resonance image super-resolution algorithm has poor recovery ability in the high-frequency region, and the high-resolution image obtained is still relatively blurred in details (i.e. high-frequency region). The multi-level attention mechanism and the loss function based on stable wavelet transform proposed in the present application can effectively recover high-frequency information. The multi-level attention mechanism can extract features of different granularities of the magnetic resonance image, enhance the representation ability of the features through the fusion of attention, and thus enhance the super-resolution performance; the stable wavelet transform can obtain high-frequency and low-frequency information of the magnetic resonance image by eliminating the down-sampling step and maintaining the translation invariance of the signal at each scale, and can force the model to focus on high-frequency information after being added to the loss function, thereby improving the quality of the super-resolution image. The present application has achieved significant improvement compared with existing models on public data sets; and has excellent effect under real low-field data.

[0016] The simulated low-field data for training is quite different from the actual low-field data, and the model trained under the public data performs poorly under the real scene. The application improves the algorithm of simulating low-field by combining physical knowledge. The application improves the current low-field simulation algorithm based on the unique physical properties of low-field magnetic resonance images and the experience of self low-field equipment acquisition. The model trained by the public data set still performs well under real low-field data without much performance difference.

[0017] Many algorithms for super-resolution of magnetic resonance images only consider the resolution improvement of a certain plane, ignoring the resolution improvement between layers. The application uses a 3D network for training. The application not only designs the overall structure of the network as a 3D form, but also extends the existing stable wavelet transform to a 3D version. The results obtained by the application meet the requirements of super-resolution of real low-field data, i.e. the resolution of three dimensions is improved.

[0018] In terms of technical advantages, the application proposes a super-resolution network based on stable wavelet transform and multi-level attention mechanism to solve the problem of insufficient resolution of low-field MRI (<0.1T) equipment. The technical scheme effectively improves the high-frequency information recovery capability of low-field magnetic resonance images by introducing multi-level attention mechanism and stable wavelet transform, and solves the problem of blur in high-frequency region of traditional methods. It is the first algorithm to use stable wavelet transform to solve the super-resolution problem of 3D images. Compared with the prior art, the application performs better in image quality indicators such as PSNR and SSIM, and can realize efficient image reconstruction with relatively low computing cost.

[0019] In terms of performance indicators, the model of the application has excellent performance on the two most commonly used magnetic resonance imaging modalities (T1 and T2). On the Human Connectome Project public data set, the model of the application is significantly better than the best algorithm in terms of important indicators such as PSNR and SSIM for measuring super-resolution effect, and the model of the application also gets the clearest and most accurate super-resolution result under real low-field data test. As shown in Figure 1 .

[0020] Table 1. Comparison of super-resolution performance indicators of T1 modality in public data set

[0021]

[0022] Table 2. Comparison of super-resolution performance indicators of T2 modality in public data set

[0023]

[0024] [1] ZHAO Y, DING Y, LAU V, et al. Whole-body magnetic resonance imaging at 0.05 tesla[J]. Science, 2024.

[0025] The technical solution of the present application has good scalability and adaptability. First, the present application uses simulated public large-scale data for training, and after training, it can be integrated into existing low-field MRI equipment without the need for large-scale hardware modification, reducing the cost of equipment upgrade. Second, the present application relies on deep learning algorithms and GPU acceleration, and can be deployed on a large scale with existing computing resources. The breakthrough of the present application in improving the quality of low-field magnetic resonance images enables low-cost and flexible MRI equipment to better meet the needs of clinical diagnosis, especially in resource-limited areas. This not only reduces medical costs, but also improves the accessibility and quality of medical services. Therefore, the present application has broad market application prospects, especially in low-income countries and regions, as well as scenarios that require mobile or portable MRI equipment, such as emergency medical care, remote area diagnosis, etc. With the growing demand for efficient and low-cost diagnostic tools in the medical market, the industrialization prospects of the present application are very optimistic.

[0026] Compared with the prior art, the present application has the following obvious and substantial characteristics and significant advantages:

[0027] 1. The present application effectively improves the resolution of low-field magnetic resonance images, solving the problem of blurring in the high-frequency region of traditional methods.

[0028] 2. The present application can achieve efficient image reconstruction with relatively low computing cost.

[0029] 3. The present application has good scalability and adaptability. Without the need for large-scale hardware modification, it reduces the cost of equipment upgrade; it can be deployed on a large scale with existing computing resources; it can reduce medical costs and improve the accessibility and quality of medical services.

[0030] The concept, specific structure and technical effects of the present application will be further described below with reference to the accompanying drawings to fully understand the purpose, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is the super-resolution result of the unlabeled real low-field data of a preferred embodiment of the present application;

[0032] Figure 2 is the super-resolution overall flowchart of a preferred embodiment of the present application;

[0033] Figure 3This is a preferred embodiment of the super-resolution network overall structure of the present invention;

[0034] Figure 4 This is a preferred embodiment of the feature extraction module structure of the present invention;

[0035] Figure 5 This is the structure of the residual multi-level attention group and multi-level attention mechanism module in a preferred embodiment of the present invention;

[0036] Figure 6 This is a preferred embodiment of the present invention, which designs a loss function based on stable wavelet transform. Detailed Implementation

[0037] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0038] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0039] Figure 2 The overall super-resolution process involves inputting a low-field 3D magnetic resonance image (Low Field Input) into a super-resolution network (super-resolution process) to obtain the corresponding super-resolution output (SR Output). Simultaneously, a dual degradation process is constructed to better learn the simulation patterns of the low-field data. This involves inputting the super-resolution output into a degradation network composed of multiple convolutional layers (degradation process, detailed in step 2.6) to obtain the corresponding pseudo-low-field output (Degradation Output). When optimizing the network, let X and X* represent the low-field input and pseudo-low-field output, respectively, and Y and Y* represent the corresponding high-field ground truth image and super-resolution output, respectively. The loss function can then be:

[0040] L=λ1L1(Y,Y * )+λ2L1(X,X * )+λ3Wavelet(Y, Y * )

[0041] Here, λ1, λ2, and λ3 are the weights of the loss function, L1 is the mean absolute error, and Wavelet is a correlation loss function designed in this algorithm, which will be introduced later.

[0042] The overall structure of a super-resolution network is as follows:Figure 3 The low-field MRI image is first processed by a shallow feature extraction network to extract shallow features. Then, a U-shaped feature extraction network is used to extract deeper features of local details. These features are then integrated by a network with an attention mechanism, and finally upsampled after a convolutional layer using pixel shuffle. To preserve the information of the original input, a residual design is used to interpolate the original image using trilinear interpolation to obtain the dimensions of the corresponding high-resolution image. This dimension is then added to the pixel-shuffled features to obtain the final high-resolution output. During the training process described above, the high-resolution output is also processed by a degradation network to obtain a corresponding pseudo-low-field output.

[0043] Feature extraction module, such as Figure 4 As shown, after shallow information extraction through a convolutional (Conv) layer, the data is input into a U-shaped feature extraction network. Through downsampling followed by upsampling, deep features of local details are extracted. The feature extractor of each layer is a residual multi-scale attention group (RMSAG) designed in this algorithm to perform detailed feature extraction. Skip connections are used between the same layers to improve the feature recovery ability of the network, thereby improving the super-resolution performance. Finally, the features are output after a convolutional layer.

[0044] Figure 5 This paper demonstrates the structure of the residual multi-level attention group and its multi-level attention mechanism module. For the residual multi-level attention group module, the input undergoes detailed feature extraction through K multi-level attention modules (this invention actually uses 5), and finally, the corresponding features are obtained through convolution. Finally, these features are combined with the original input through residual connections to obtain the corresponding output. The multi-level attention mechanism is the core module of the algorithm. The input features are integrated through spatial attention and channel attention to extract spatial and channel features, which are then finally connected to the input through residual connections. For the spatial attention part, after point convolution, the channel is first split into three features, which are then convolved by three convolution kernels of different sizes. Different levels of spatial information are obtained through different receptive fields. After merging to obtain multi-level spatial features, high-order information is extracted using max pooling and average pooling. Attention weights are obtained through the sigmoid activation function and multiplied with the multi-level spatial features to obtain weighted spatial feature information. For channel attention, it is similar to the traditional channel attention module. After average pooling, it is passed through two convolution layers and one LeakyReLU activation layer to obtain the corresponding channel attention weights. These weights are then multiplied with the spatial features to obtain the final overall feature.

[0045] Currently, deep learning-based super-resolution algorithms have achieved good super-resolution capabilities for low-frequency (overall structure) data. The main challenge lies in recovering high-frequency information. While the commonly used L1 Loss can address the oversmoothing problem caused by the Mean Square Error Loss (MSE Loss), it still has issues with high-frequency information recovery. Wavelet Transform is a mathematical tool for signal processing, primarily used to analyze the local features of signals. Unlike the traditional Fourier Transform, wavelet Transform can provide signal information in both the time and frequency domains simultaneously, making it particularly suitable for processing non-stationary signals (such as abrupt changes and noise). In the traditional Discrete Wavelet Transform (DWT), the signal is decomposed into multiple scales through filtering and downsampling (usually a 2x downsampling). While effective, this method leads to the loss of positional information during downsampling, known as translation sensitivity. This means that if the signal is translated in the time domain, its wavelet coefficients will change significantly, which is undesirable for some applications (such as signal denoising and edge detection). The Stationary Wavelet Transform (SWT) preserves the translation invariance of a signal at each scale by eliminating the downsampling step. Specifically, SWT performs only filtering operations at each decomposition level, without downsampling. This means that the output signal length at each decomposition level is the same as the original signal length, thus preserving all positional information. This invention develops a stationary wavelet transform that can be used for 3D data and can be accelerated by a GPU. Figure 6 As shown, during the training phase, the obtained super-resolution output and the real image are subjected to stable wavelet transforms respectively, and the average absolute error of the high-frequency information is calculated. This allows the model to better focus on high-frequency information, thereby enhancing its ability to recover high-frequency information.

[0046] Specific training process:

[0047] Step 1: Establish a low-field image super-resolution dataset

[0048] Step 1.1: Obtain the true labels using an existing high-field public dataset. First, divide the dataset into training, validation, and test sets according to an 8:1:1 ratio. For each 224*224*224 3D high-field MR image, apply a degradation function to obtain a simulated low-field magnetic resonance image of size 112*112*112. To save computational resources, for each paired sample, crop multiple 32*32*32 magnetic resonance image patches (corresponding to a high-field image voxel size of 64*64*64) to avoid the inability to use GPU training due to excessively large image sizes, thus affecting training speed. The formula for simulating the low-field magnetic resonance image is as follows:

[0049]

[0050] In the formula: y is the low-resolution image; x is the high-resolution image corresponding to y; k represents the blur kernel; This refers to the convolution operation between the high-resolution image and the blur kernel; ↓s represents downsampling at a scaling scale of s, which in practical algorithms is partial truncation in the K-space; Rician is additive Ricean noise. Here, f(x, y, z) represents the magnetic resonance image, f final If (x, y, z) is a simulated low-field image, then the specific formula for degradation is:

[0051]

[0052] Among them, FFT3 and IFFT3 are the 3D Fast Fourier Transform and Inverse Fast Fourier Transform, respectively, pf line1 and pf line3 It is the boundary of the zero-filled region in a partial Fourier space (k-space). In the algorithm, the fill ratio is taken as 0.3, i.e., pf line1 =pf line3 =67.

[0053] Step 2: Establish a super-resolution network based on stable wavelet transform and multi-level attention mechanism.

[0054] Step 2.1: Use the degraded low-resolution image LR from Step 1.1 as the input image;

[0055] Step 2.2: Perform a convolution with a kernel size of 3*3*3, 64 channels, and a stride of 1 on the low-resolution image input to the network using LR. Use convolution (conv) to extract shallow features of the image.

[0056] Step 2.3: The features obtained in Step 2.2 are input into the "U"-shaped feature extraction network for detail learning, where the structure of the "U"-shaped feature extraction network is as follows: Figure 3 As shown. Specifically, the features are fed into the Residual Multilevel Attention Group (RMSAG) for feature extraction at the spatial and channel levels. The output features are downsampled by a convolution with a kernel of 3*3*3, 64 channels, and a stride of 2, and then fed into the next RMSAG layer. This process is repeated three times. For the features extracted from the third RMSAG, upsampling is performed using trilinear interpolation, and the features are concatenated with the features obtained from the previous layer using skip connections, and then fed into the RMSAG. The resulting features are further processed using the above method. Finally, the features are integrated by a convolution with a kernel of 3*3*3, 64 channels, and a stride of 1 to obtain the extracted features.

[0057] Step 2.3.1: Residual multi-level attention group, such as Figure 4As shown, each residual attention group includes three multi-level attention mechanism modules (MSA). The input features are processed... Figure 4 The process extracts detailed features from both channel and spatial aspects. For the spatial attention part, after point convolution, the channel is first split into three features, which are then convolved by three different-sized convolution kernels (3*3*3, 5*5*5, and 7*7*7). Different levels of spatial information are obtained through different receptive fields. After merging to obtain multi-level spatial features, high-order information is extracted using max pooling and average pooling. Attention weights are obtained through the sigmoid activation function and multiplied with the multi-level spatial features to obtain weighted spatial feature information. For channel attention, it is similar to the traditional channel attention module. After average pooling, it is passed through two layers of convolution with 3*3*3 kernels, 64 channels, and a stride of 1, and a LeakyReLU activation layer to obtain the corresponding channel attention weights. These weights are then multiplied with the spatial features to obtain the final overall feature.

[0058] Step 2.4: The extracted features are integrated and refined using a spatial attention mechanism. The features obtained in Step 2.3 are subjected to max pooling and average pooling along the channel dimension, compressing the channel dimension to 1 while preserving spatial information. Then, the features are concatenated and extracted using a convolution with a 3x3x3 kernel, 1 channel, and a stride of 1. Finally, sigmoid activation is applied to obtain spatial weights, which are multiplied by the original features to obtain weighted features. After validation on the test set, the optimal performance metrics and visual effects were achieved using only the spatial attention mechanism, without incorporating channel attention.

[0059] Step 2.5: Assuming the magnification factor is k, the features output in step 2.4 are processed by a convolution kernel of 3*3*3 with 64×2 channels. k The image is first convolved with a stride of 1, and then upsampled by pixel shaving. Compared with common deconvolution and trilinear interpolation, pixel shaving can avoid the mosaic effect, and the learnable parameters can be upsampled more accurately. In order to preserve the original spatial information, the low-resolution image LR obtained in step 1.1 is upsampled by trilinear interpolation and added to the image upsampled by pixel shaving to obtain the final super-resolution image.

[0060] Step 2.6: Perform the degradation process. First, input the super-resolution image obtained in Step 2.5 into a convolutional layer with a kernel of 3*3*3, 16 channels, a stride of 2, and padding of 1. Assuming the magnification factor is k (k should be a power of 2), the obtained super-resolution image is then input into log2(k)-1 downsampling blocks consisting of a 3*3*3 kernel, 16 channels, a stride of 2, and padding of 1, and LeakyReLU. Finally, the image is passed through a convolutional layer with a kernel of 3*3*3, 1 channel, and a stride of 1 to output the corresponding pseudo-low field output.

[0061] Step 3: Train the model. Use the training sample set obtained in Step 1 to train the super-resolution network based on stable wavelet transform and multi-level attention mechanism.

[0062] Step 3.1: Set the training parameters. Set the batch size to Batch = 16, the initial learning rate to lr = 0.001, and the number of training iterations to Epoch = 300.

[0063] Step 3.2: Feed the preprocessed training set from Step 1.1 into the super-resolution network constructed in Step 2 based on stable wavelet transform and multi-level attention mechanism;

[0064] Step 3.3: Objective function of the super-resolution network during training.

[0065] Step 3.3.1: Perform a stable wavelet transform on the generated super-resolution image and the real high-resolution image:

[0066] Step 3.3.1.1: Select a suitable mother wavelet (this invention uses the Haar wavelet).

[0067] Step 3.3.1.2: Perform a one-dimensional stable wavelet transform on each YZ plane (all slices in the x direction) of the 3D image to generate low-frequency and high-frequency coefficients.

[0068] Step 3.3.1.3: Perform 1D SWT on the result of the previous step along the Y-axis, processing each XZ plane (all slices in the y-direction).

[0069] Step 3.3.1.4: Perform 1D SWT on the result of the previous step along the Z-axis, processing each XY plane (all slices in the z-direction).

[0070] Step 3.3.1.5: Finally, low-frequency (LLL) and high-frequency (HLH, HHL, HLL, LHL, LHH, LLH, HHH) subbands are obtained. The high-frequency subbands are then concatenated along the channel dimension to obtain the high-frequency features. If the input image dimension is (x, y, z), then the dimensions of these subbands are also (x, y, z).

[0071] Step 3.3.2: Calculate the loss function of the super-resolution network. The formula for the loss function is as follows:

[0072] L=λ1L1(Y,Y * )+λ2L1(X,X * )+λ3W avelet(Y,Y * )

[0073] Wavelet(Y, Y) * = L1{SWT(Y) HF SWT(Y) * ) HF}

[0074] Here, X represents the low-field input, X* represents the pseudo-low-field output obtained in step 2.6, Y represents the high-field ground truth image, and Y* represents the super-resolution output obtained in step 2.5. λ1, λ2, and λ3 are the weights of the loss function, with weights of 0.8, 0.1, and 0.1 in the experiment. L1 is the mean absolute error, and L1(Y, Y*) = L*(Y*, Y*). * )=E|YY * |,L1(X,X * )=E|XX * For the calculation of frequency domain loss, i.e. Wavelet Loss, firstly, according to step 3.3.1.5, the high-frequency sub-band features of the generated super-resolution image and the real high-resolution image are obtained, and the average absolute error is calculated on these features;

[0075] Step 3.4: Based on the changing trends of the evaluation metrics Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Loss function, adjust the network parameters in a timely manner until the training reaches a stable state, and determine the final network parameters;

[0076] Step 3.5: Based on the network parameters determined in Step 3.4, complete the training of the model to obtain a well-converged super-resolution network based on stable wavelet transform and multi-level attention mechanism;

[0077] Step 4: Test the model; Input the low-resolution image to be reconstructed into the network trained in Step 3, note that it does not go through the degradation module in Step 2.6, and output the corresponding super-resolution image through Step 2.5.

[0078] Step 5: Evaluate the model. Based on the super-resolution images obtained in Step 4, evaluate the performance of the proposed super-resolution network in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

[0079] Step 6: Determine whether the performance metrics obtained in Step 5, namely Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), meet the actual application requirements. If the model meets the actual application requirements, proceed to Step 8; otherwise, proceed to Step 7.

[0080] Step 7: Correct the parameters of the super-resolution network based on stable wavelet transform and multi-level attention mechanism constructed in Step 2, and jump to Step 3 to retrain.

[0081] Step 8: Use the model from Step 6 that meets the practical application requirements for the restoration of low-resolution images.

[0082] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A low-field magnetic resonance imaging super-resolution network, characterized in that, It includes a multi-level attention mechanism module and a loss function based on stable wavelet transform; The low-field magnetic resonance image is first processed by a shallow feature extraction network to extract shallow features. Then, a "U"-shaped feature extraction network is used to extract deep features of local details. The features are integrated by a network with an attention mechanism, and finally upsampled by pixel rearrangement after the convolutional layer. At the same time, residual design is used to obtain the dimension of the corresponding high-resolution image by trilinear interpolation of the original image. This dimension is added to the features after pixel rearrangement to obtain the high-resolution output. The high-resolution output is also processed by a degradation network to obtain the corresponding pseudo-low-field output. The dual degradation process inputs the super-resolution output into a degradation network composed of multiple convolutional layers to obtain the corresponding pseudo-low-field output. The input is a U-shaped feature extraction network. By downsampling and then upsampling, deep features of local details are extracted. The feature extractor of each layer performs detailed feature extraction through residual multi-level attention groups. Skip connections are used between the same layers. The residual multi-level attention group splits the channel into three features, which are convolved by three convolutional kernels of different sizes. Different levels of spatial information are obtained through different receptive fields. After merging to obtain multi-level spatial features, high-order information is extracted using max pooling and average pooling. Attention weights are obtained through the sigmoid activation function and multiplied with the multi-level spatial features to obtain weighted spatial feature information. The channel attention is then average pooled and passed through two convolutional layers and LeakyReLU activation layers to obtain the corresponding channel attention weights. These weights are then multiplied with the spatial features to obtain the final overall feature. A 3D network is used for training; the overall structure of the network is in 3D form, and the stable wavelet transform is extended to a 3D version; the stable wavelet transform maintains the translation invariance of the signal at each scale by eliminating the downsampling step; during the training phase, the obtained super-resolution output and the real image are subjected to stable wavelet transform respectively, and the average absolute error of the high-frequency information is calculated. The formula for simulating low-field magnetic resonance images using GPU training is as follows; Where: y is the low-resolution image; x is the high-resolution image corresponding to y; k represents the blur kernel; This refers to the convolution operation between a high-resolution image and a blur kernel; ↓s represents downsampling at a scaling scale of s, which is a partial truncation in the K-space. It is additive Rice noise.

2. The low-field magnetic resonance imaging super-resolution network as described in claim 1, characterized in that, The multi-level attention mechanism module extracts features of different granularities from magnetic resonance images and uses attention fusion to enhance feature representation.

3. The low-field magnetic resonance imaging super-resolution network as described in claim 1, characterized in that, The stable wavelet transform performs only filtering operations in each decomposition stage, without performing downsampling.

Images