Low-field magnetic resonance image super-resolution network

Abstract

The invention discloses a low-field magnetic resonance image super-resolution network, and relates to the field of magnetic resonance imaging. In order to solve the problem of insufficient resolution of low-field magnetic resonance imaging (MRI) equipment, a multi-level attention mechanism and stable wavelet transform are introduced to carry out image super-resolution. According to a low-field magnetic resonance image, firstly, shallow features are extracted through a shallow feature extraction network, then, deep features of local details are extracted through a U-shaped feature extraction network, feature integration is carried out on the features through a network containing an attention mechanism, up-sampling is carried out through pixel rearrangement, meanwhile, an original image structure is learned through residual error design, and a low-field magnetic resonance image is obtained. And finally, high-resolution output is obtained. According to the method, the definition of the low-field magnetic resonance image is effectively improved, and the problem that a traditional method is fuzzy in a high-frequency area is solved; the method is excellent in image quality indexes such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), and efficient image reconstruction can be realized under the condition of relatively low calculation cost.

Description

Technical Field

[0001] The present invention relates to the field of magnetic resonance imaging, and in particular to a low-field magnetic resonance image super-resolution network. Background Art

[0002] Magnetic resonance imaging (MRI) is an important medical imaging method. Compared with other imaging methods, it has the advantages of non-ionization, non-invasiveness, and multi-parameters. However, superconducting MRI scanners under existing standards (mainly 1.5T and 3T) are expensive and need to be installed in places with magnetic shielding equipment, which limits the use of MRI by more people. Ultra-low-field MRI (intensity <0.1T) equipment may be able to enter hospitals in low-income countries and provide diagnosis for low-income people due to its low cost and high flexibility. However, the current low-field magnetic resonance images have a low resolution and are difficult to use directly for clinical diagnosis. In order to improve the signal-to-noise ratio, low-field magnetic resonance images are super-resolved.

[0003] Therefore, technicians in this field 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 low-field magnetic resonance images is effectively improved, solving the problem of blurring in high-frequency areas of traditional methods. Summary of the invention

[0004] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is the insufficient resolution of low-field MRI equipment, and the problem of very low resolution of low-field magnetic resonance images, especially the blurring of high-frequency areas.

[0005] To achieve the above objectives, the present invention provides a low-field magnetic resonance image super-resolution network, including a multi-level attention mechanism module and a loss function based on stable wavelet transform.

[0006] Furthermore, the multi-level attention mechanism module extracts features of different granularities of the magnetic resonance image and utilizes the attention mechanism to fuse enhanced feature representation.

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

[0008] Furthermore, the stable wavelet transform only performs filtering operations in each level of decomposition without performing downsampling.

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

[0010] Furthermore, the low-field magnetic resonance image first passes through a shallow feature extraction network to extract shallow features, and then uses a "U"-shaped feature extraction network to extract deep features. The features are integrated through an attention mechanism module, upsampled, and output using residual connections.

[0011] Furthermore, it also includes a residual multi-level attention group module, the input is subjected to detailed feature extraction by the multi-level attention mechanism module, the corresponding features are obtained by convolution, and are combined with the original input through residual connection to obtain the corresponding output.

[0012] Furthermore, in the multi-level attention mechanism module, the input features are integrated by spatial attention and channel attention, spatial and channel features are extracted, and finally connected with the input through a residual connection.

[0013] Furthermore, the spatial attention part splits the channel into three features, which are convolved by three convolution kernels of different sizes respectively to obtain spatial information at different levels. After merging the multi-level spatial features, the maximum pooling and average pooling are used to extract high-order information, and the attention weight is obtained by the sigmoid activation function, which is multiplied by the multi-level spatial features to obtain weighted spatial feature information.

[0014] Furthermore, the channel attention part, after average pooling, passes through two layers of convolution and one layer of LeakyReLU activation layer to obtain the corresponding channel attention weight, and then multiplies it with the spatial feature information to obtain the final overall feature.

[0015] The current super-resolution algorithm for low-field magnetic resonance images is not good enough in the recovery ability of high-frequency areas, and the obtained high-resolution images are still relatively fuzzy in details (i.e., high-frequency areas). The multi-level attention mechanism and the loss function based on the stable wavelet transform proposed in the present invention can effectively restore high-frequency information. The multi-level attention mechanism of the present invention can extract features of different granularities of magnetic resonance images, enhance the representation ability of 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 magnetic resonance images by eliminating the downsampling step and maintaining the translation invariance of the signal at each scale. After adding the loss function, the model can be forced to focus on high-frequency information, thereby improving the quality of super-resolution images. The present invention has achieved significant improvements in public data sets compared to existing models; it also has excellent results under real low-field data.

[0016] The simulated low-field data used for training is very different from the actual low-field data, and the performance of the model trained under the public data is significantly different in the real scene. The present invention improves the algorithm for simulating low fields by combining physical knowledge. The present invention improves the current low-field simulation algorithm based on the unique physical properties of low-field magnetic resonance images and the experience of collecting autonomous low-field equipment. The model trained by the present invention using 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 present invention uses a 3D network for training. The present invention not only designs the overall structure of the network into a 3D form, but also expands the existing stable wavelet transform into a 3D version. The results obtained by the present invention meet the requirements of real low-field data super-resolution, that is, the resolution is improved in all three dimensions.

[0018] In terms of technical advantages, the present invention aims at the problem of insufficient resolution of low-field MRI (<0.1T) equipment and proposes a super-resolution network based on stable wavelet transform and multi-level attention mechanism. The technical solution 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 blurring in high-frequency areas of traditional methods. This is also the first algorithm to use stable wavelet transform to solve the problem of 3D image super-resolution. Compared with the existing technology, the invention performs superiorly in image quality indicators such as PSNR and SSIM, and can achieve efficient image reconstruction at a relatively low computational cost.

[0019] In terms of performance indicators, the model of the present invention has excellent performance in the two most commonly used magnetic resonance imaging modalities (T1 and T2). In the Human Connectome Project public data set, the model of the present invention is significantly better than the best algorithm in terms of PSNR and SSIM, which are important indicators for measuring super-resolution effects. At the same time, under the test of real low-field data, the model of the present invention also obtains the clearest and most accurate super-resolution results. Figure 1 shown.

[0020] Table 1. Comparison of super-resolution performance indicators in the T1 mode of public datasets

[0021]

[0022] Table 2. Comparison of super-resolution performance indicators in T2 mode on public datasets

[0023]

[0024] [1]ZHAO Y, DING Y, LAU V, et al.Whole-body magnetic resonance imaging at0.05tesla[J].Science, 2024.

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

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

[0027] 1. The present invention effectively improves the resolution of low-field magnetic resonance images and solves the problem of blurring in high-frequency areas of traditional methods.

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

[0029] 3. The present invention has good scalability and adaptability. It does not require large-scale hardware modification, which reduces the cost of equipment upgrades; it can be deployed on a large scale with the help of 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 invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a label-free real low-field data super-resolution result of a preferred embodiment of the present invention;

[0032] Figure 2 It is a super-resolution overall flow chart of a preferred embodiment of the present invention;

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

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

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

[0036] Figure 6 It is a loss function design based on stable wavelet transform of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0038] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is appropriately exaggerated in some places in the drawings.

[0039] Figure 2 It is the overall super-resolution process. The low-field 3D magnetic resonance image (Low Field Input) is input into the super-resolution network (super-resolution process) to obtain the corresponding super-resolution output (SR Output). At the same time, a dual degradation process is constructed to better learn the simulation law of low-field data, that is, the super-resolution output is input into the degradation network composed of multiple convolutional layers (degradation process, see step 2.6 for detailed description) to obtain the corresponding pseudo low-field output (Degradation Output). When optimizing the network, let X, X* represent the low-field input and pseudo low-field output respectively, and Y, Y* represent the corresponding high-field real image and super-resolution output respectively. Then the loss function can be:

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

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

[0042] The overall structure of the super-resolution network is as follows Figure 3 The low-field MRI image first passes through a shallow feature extraction network to extract shallow features, and then uses a "U"-shaped feature extraction network to extract deep features of local details. The features are integrated through a network containing an attention mechanism, and finally upsampled through pixel shuffle after the convolution layer. At the same time, in order to maintain the original input information, the residual design is used to obtain the dimension of the corresponding high-resolution image through trilinear interpolation of the original image, and then add the features after pixel shuffle to finally obtain a high-resolution output. In the training process described above, the high-resolution output will also obtain the corresponding pseudo low-field output through the degradation network.

[0043] Feature extraction module such as Figure 4 As shown in the figure, after a layer of convolution (Conv) layer for shallow information extraction, it is input into a "U"-shaped feature extraction network. By downsampling and then upsampling, the deep features of local details are extracted. The feature extractor of each layer is a residual multi-scale attention group (RMSAG) designed by this algorithm for 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, after a layer of convolution layer, the features are output.

[0044] Figure 5 The structure of the residual multi-level attention group and the multi-level attention mechanism module therein is shown. For the residual multi-level attention group module, the input is subjected to detailed feature extraction by K multi-level attention modules (the present invention actually uses 5), and finally the corresponding features are obtained by convolution, and finally combined with the original input through residual connection to obtain the corresponding output. The multi-level attention mechanism is the core module of the algorithm. The input features are integrated by spatial attention and channel attention to extract spatial and channel features, and finally connected with the input through residual connection. For the spatial attention part, after the point convolution operation, the channel is first split into three features, which are convolved by three convolution kernels of different sizes respectively. Different levels of spatial information are obtained through different receptive fields. After merging to obtain multi-level spatial features, maximum pooling and average pooling are used to extract high-order information. The attention weight is obtained by 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, two layers of convolution and one layer of LeakyReLU activation layer are used to obtain the corresponding channel attention weight, which is then multiplied with the spatial feature to obtain the final overall feature.

[0045] At present, the super-resolution algorithm based on deep learning has a good super-resolution ability for low frequencies (overall structure), and the main problem is the recovery of high-frequency information. Although the commonly used L1 Loss can solve the over-smoothing problem caused by the mean square error loss function (MSE Loss), it still has problems in the recovery of high-frequency information. Wavelet transform is a mathematical tool for signal processing, which is mainly used to analyze the local characteristics of signals. Unlike the traditional Fourier transform, the wavelet transform can provide signal information in both the time domain and the frequency domain, so it is particularly suitable for processing non-stationary signals (such as mutations, noise, etc.). In the traditional discrete wavelet transform (DWT), the signal is decomposed step by step into multiple scales through filtering and downsampling (usually 2 times downsampling). Although this method is effective, it will lead to the loss of position information during the downsampling process, that is, translation sensitivity. This means that if the signal is translated in the time domain, its wavelet coefficients will change significantly, which is not desirable for some applications (such as signal denoising and edge detection). The stationary wavelet transform (SWT) maintains the translation invariance of the signal at each scale by eliminating the downsampling step. Specifically, SWT only performs filtering operations at each level of decomposition without performing downsampling. This means that the length of the output signal at each level of decomposition is the same as the length of the original signal, thereby retaining all position information. The present invention develops a stable wavelet transform that can be used for 3D data and can be accelerated by a GPU. Figure 6 As shown in the figure, during the training phase, the obtained super-resolution output and the real image are subjected to stable wavelet transform respectively, and the mean absolute error of the high-frequency information is calculated. This allows the model to better focus on the high-frequency information, thereby enhancing the ability to recover the high-frequency information.

[0046] Specific training process:

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

[0048] Step 1.1: Use the existing high-field public dataset to obtain the true label. First, divide the dataset into training set, validation set and test set according to the ratio of 8:1:1 of the total samples. For each 3D high-field MR image of size 224*224*224, a simulated low-field magnetic resonance image of size 112*112*112 is obtained through the degradation function. In order to save computing resources, for each paired sample, multiple magnetic resonance image blocks of size 32*32*32 (corresponding to the high-field image voxel size of 64*64*64) are cropped to avoid the inability to use GPU training due to the large image size, which affects the training speed. The formula for simulating low-field magnetic resonance images is:

[0049]

[0050] Where: y is the low-resolution image; x is the high-resolution image corresponding to y; k represents the blur kernel; Refers to the convolution operation between the high-resolution image and the blur kernel; ↓s represents the downsampling with a scale of s, which is a partial truncation of the K space in the actual algorithm; Rician is the additive Rician noise. Here, f(x, y, z) represents the magnetic resonance image, and f final (x, y, z) is the simulated low-field image, and the specific formula for degradation is:

[0051]

[0052] Among them, FFT3 and IFFT3 are 3D fast Fourier transform and inverse fast Fourier transform respectively, pf line1 and pf line3 It is the boundary of the area filled with zeros in some Fourier space (k space). The filling ratio is 0.3 in the algorithm, that is, 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 in step 1.1 as an input image;

[0055] Step 2.2: The low-resolution image LR input into the network undergoes a convolution with a kernel of 3*3*3, 64 channels, and a stride of 1, and uses 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 detailed learning, where the "U"-shaped feature extraction network structure is as follows Figure 3 Specifically, the features are sent to the residual multi-level attention group (RMSAG) for feature extraction from 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 sent to the RMSAG of the next layer. This process is performed three times. For the features extracted from the third RMSAG, they are upsampled by trilinear interpolation, concatenated with the features obtained from the previous layer using skip connections, and then sent to the RMSAG. The obtained features are further processed by the above method, and finally 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 in the figure, each residual attention group consists of three multi-level attention mechanism modules (MSA), and the input features are Figure 4 The process extracts detailed features from both channel and space aspects respectively. For the spatial attention part, after the point convolution operation, the channel is first split into three features, which are convolved by three convolution kernels of different sizes (3*3*3, 5*5*5, 7*7*7) respectively. Different levels of spatial information are obtained through different receptive fields. After merging to obtain multi-level spatial features, high-order information is extracted by maximum pooling and average pooling, and the attention weight is obtained by the sigmoid activation function. It is 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, two layers of convolution with a kernel of 3*3*3, 64 channels and a stride of 1 are passed through a layer of LeakyReLU activation layer to obtain the corresponding channel attention weight, which is then multiplied with the spatial features to obtain the final overall feature.

[0058] Step 2.4: The extracted features are integrated and improved through the spatial attention mechanism. The features obtained in step 2.3 are subjected to maximum pooling and average pooling in the channel dimension, and the channel dimension is compressed to 1 to retain the spatial information. The features are then concatenated and extracted through a convolution with a convolution kernel of 3*3*3, a channel of 1, and a step size of 1. Finally, after sigmoid activation, the spatial weight is obtained and multiplied with the original feature to obtain the weighted feature. After verification on the test set, the channel attention mechanism was not added, and only the spatial attention mechanism was used to obtain the optimal performance indicators and visual effects.

[0059] Step 2.5: Assuming the magnification is k, the features output from step 2.4 are convolved through a convolution kernel of 3*3*3 and a channel of 64×2. k , convolution with a step size of 1, and then upsampling through pixel shuffle. Compared with common deconvolution and trilinear interpolation, pixel shuffle can avoid the mosaic effect, and the learnable parameters can also upsample more accurately. In order to retain 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 shuffle to obtain the final super-resolution image.

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

[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, the training batch size is set to Batch = 16, the initial learning rate is set to lr = 0.001, and the number of training iterations Epoch = 300;

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

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

[0065] Step 3.3.1: Perform 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 (the present invention uses Harr 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 along the Y axis on the result of the previous step, processing each XZ plane (all slices in the y direction).

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

[0070] Step 3.3.1.5: Finally, we get the low-frequency (LLL) and high-frequency (HLH, HHL, HLL, LHL, LHH, LLH, HHH) subbands, and concatenate the high-frequency subbands according to the channel dimension to get the high-frequency features. If the input image dimension is (x, y, z), then the dimension of these subbands is also (x, y, z).

[0071] Step 3.3.2: Calculate the loss function of the super-resolution network. The formula of 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 real image, and Y* represents the super-resolution output obtained in step 2.5. λ1, λ2, and λ3 are the weight terms of the loss function, respectively. The weights in the experiment are 0.8, 0.1, and 0.1. L1 is the mean absolute error, L1(Y, Y * )=E|YY * |, L1(X, X * )=E|XX * |. For the calculation of frequency domain loss, namely 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 mean absolute error of the features is calculated;

[0075] Step 3.4: According to the changing trends of the evaluation indicators peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) and the loss function (Loss), adjust the network parameters in time until the training tends to be stable and determine the final network parameters;

[0076] Step 3.5: According to the network parameters determined in step 3.4, complete the model training and 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 outputs the corresponding super-resolution image through step 2.5.

[0078] Step 5: Evaluation model: Based on the super-resolution image obtained in step 4, the performance of the proposed super-resolution network is evaluated from the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM);

[0079] Step 6: Determine whether the performance indicators 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, execute step 8; otherwise, execute step 7;

[0080] Step 7: Modify 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 for retraining;

[0081] Step 8: Use the model in step 6 that meets the actual application requirements to restore low-resolution images.

[0082] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A low-field magnetic resonance image super-resolution network, characterized in that: It includes a multi-level attention mechanism module and a loss function based on stable wavelet transform.

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

3. The low-field magnetic resonance image super-resolution network according to claim 1, characterized in that: In the loss function based on stable wavelet transform, the stable wavelet transform maintains the translation invariance of the signal at each scale by eliminating the downsampling step.

4. The low-field magnetic resonance image super-resolution network according to claim 3, characterized in that: The stable wavelet transform only performs filtering operations in each level of decomposition without performing downsampling.

5. The low-field magnetic resonance image super-resolution network according to claim 1, characterized in that: A 3D network is used for training.

6. The low-field magnetic resonance image super-resolution network according to claim 1, characterized in that: The low-field magnetic resonance image is first passed through a shallow feature extraction network to extract shallow features, and then a "U"-shaped feature extraction network is used to extract deep features of local details. The features are integrated and up-sampled through a network containing the multi-level attention mechanism module, and the output is obtained using residual connections.

7. The low-field magnetic resonance image super-resolution network according to claim 1, characterized in that: It also includes a residual multi-level attention group module. The input is subjected to detailed feature extraction by the multi-level attention mechanism module, and the corresponding features are obtained by convolution. The features are combined with the original input through residual connection to obtain the corresponding output.

8. The low-field magnetic resonance image super-resolution network according to claim 1, characterized in that: In the multi-level attention mechanism module, the input features are integrated by spatial attention and channel attention, spatial and channel features are extracted, and finally connected with the input through a residual connection.

9. The low-field magnetic resonance image super-resolution network according to claim 8, characterized in that: The spatial attention part splits the channel into three features, which are convolved by three convolution kernels of different sizes respectively to obtain spatial information at different levels. After merging the multi-level spatial features, the maximum pooling and average pooling are used to extract high-order information, and the attention weight is obtained by the sigmoid activation function, which is multiplied by the multi-level spatial features to obtain weighted spatial feature information.

10. The low-field magnetic resonance image super-resolution network according to claim 9, characterized in that: The channel attention part, after average pooling, passes through two layers of convolution and one layer of LeakyReLU activation layer to obtain the corresponding channel attention weight, and then multiplies it with the spatial feature information to obtain the final overall feature.

Images