A multi-domain optimization model construction method for reducing Rician noise in brain MRI images

Abstract

The present invention discloses a multi-domain optimization model construction method for reducing Rician noise in brain magnetic resonance imaging, comprising the following steps: establishing an MDNet denoising model including an image domain sub-model and a wavelet domain sub-model; concatenating the denoised data of the training set processed by the two sub-models and then passing it into a sampling-optimized non-local attention module to obtain the final denoised brain MR image data; setting the loss function MSE of the MDNet denoising model; using the loss value of the loss function MSE converging to a value lower than a preset value as the judgment standard for stopping the MDNet denoising model training, and saving the parameters of the MDNet denoising model. Based on the image domain, the present invention adds the wavelet domain to train the above network, so that the wavelet domain sub-model can pay attention to the structural similarity of the image, and the denoising result is further improved by the sampling-optimized non-local attention module, greatly reducing the noise pollution of the MR image.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic resonance imaging (MRI) medical image processing, and involves methods such as artificial intelligence, convolutional neural networks, image denoising, and the application of deep learning to remove Rice noise in brain MRI images. Specifically, it relates to a method for constructing a multi-domain optimization model for reducing Rice noise in brain MRI images. Background Art

[0002] Magnetic resonance imaging (MRI) is an imaging technique used in medical radiology to generate anatomical and physiological images. It offers numerous advantages: 1. It uses no ionizing radiation, making it safe and noninvasive; 2. It provides higher resolution images of the brain and soft tissue compared to other imaging techniques such as CT, ultrasound, and X-ray; and 3. Multi-directional imaging intuitively displays the spatial relationships between anatomical structures and lesions. Generally speaking, noise in MRI is primarily caused by factors such as the scanning system's components, the surrounding environment, and the scanned object. Furthermore, the Gaussian noise carried by the initial sampled signal, after transformation through the K-space data, transforms into a complex Rician distribution in the resulting image. Severely noisy brain MR image data can even adversely affect subsequent analysis and diagnosis by physicians. Therefore, Rician noise reduction in MR images is crucial.

[0003] Currently, the most popular denoising methods are those used for post-processing of MRI images. Traditional denoising methods are based on three categories: spatial domain, frequency domain, and statistical methods. Deep learning methods that utilize neural network schemes are relatively rare and are not widely used in practical denoising. Furthermore, due to the sensitivity and scarcity of medical data, deep learning-based MRI denoising research using limited samples is challenging. Summary of the Invention

[0004] The purpose of the present invention is to address the deficiencies of the existing technology and provide a multi-domain optimization model construction method for reducing Rician noise in brain magnetic resonance images, aiming to solve the problem of processing noise-contaminated pixels in MR images, especially to reduce the impact of Rician noise in brain magnetic resonance images.

[0005] The multi-domain optimization model construction method for reducing Rice noise in brain magnetic resonance imaging of the present invention comprises the following steps:

[0006] Step 1: Acquire brain MR image data with different noise levels, and perform cross-sectional slicing on the brain MR image data.

[0007] Step 2: De-redundancy is performed on the sliced ​​data, invalid slices are filtered out, and the data set is divided into training set, validation set and test set.

[0008] Step 3: Establish the MDNet denoising model, including the image domain sub-model and the wavelet domain sub-model.

[0009] Step 4: Input the brain MR image data with Rice noise levels of 3%, 5%, 7% and 9% in the training set into the image domain sub-model of the MDNet denoising model, and use the obtained output value as the denoising data of the image domain sub-model.

[0010] Step 5: Input the brain MR image data with Rice noise levels of 3%, 5%, 7% and 9% in the training set into the wavelet domain sub-model of the MDNet denoising model, and use the obtained output value as the denoising data of the wavelet domain sub-model.

[0011] Step 6: Concatenate the denoised data obtained in step 4 and the denoised data obtained in step 5 as preliminary denoised data, and pass them into the sampling-optimized non-local attention module to obtain the final denoised brain MR image data.

[0012] Step 7. Set the loss function MSE of the MDNet denoising model to measure the error between the denoised brain MR image data and the clean MR image data with a Rice noise level of 0% in the training set.

[0013] Step 8. Use the training set generated in step 2 to train the MDNet denoising model, and use the validation set to evaluate the training process. When the loss value of the loss function MSE converges to a value lower than the preset value, stop training and save the parameters of the MDNet denoising model.

[0014] Preferably, after executing step 8, the test set is input into the trained MDNet denoising model to test the performance of the MDNet denoising model.

[0015] Preferably, in step 2, the training set, validation set, and test set all contain clean MR image data with a Rice noise level of 0%, and noise-contaminated brain MR image data with Rice noise levels of 3%, 5%, 7%, and 9%. The noise-contaminated brain MR image data are paired with slices at the same spatial location in the clean MR image data to facilitate loss calculation during training. Furthermore, redundancy removal is performed by removing invalid slices below the neck.

[0016] Preferably, the image domain sub-model includes a feature correction unit and a multi-scale denoising module, the feature correction unit is used to constrain the channel and size parameters of the feature map, and the multi-scale denoising module is used to capture context information and noise distribution characteristics.

[0017] The feature correction unit consists of a 1*1 convolutional layer, a batch normalization layer and a parameter-corrected linear unit activation function.

[0018] The input of the multi-scale denoising module passes through a preliminary 3*3 convolution layer, a batch normalization layer, and a parameter-corrected linear unit activation function and is then passed into a three-branch structure. The three-branch structure combines the operations of two dilated convolutions and one 5*5 large-parameter convolution kernel, where the convolution kernels of the two dilated convolutions are 3*3, and their dilation rates are set to 2 and 3, respectively. Each branch also passes through a batch normalization layer and a parameter-corrected linear unit activation function and is merged at a confluence layer. Finally, the output passes through a 3*3 convolution layer, a batch normalization layer, and a parameter-corrected linear unit activation function and is residually connected to fuse noise features under multiple receptive fields.

[0019] Preferably, the wavelet domain submodel is composed of a wavelet transform unit, a feature correction unit, a multi-scale denoising module and a wavelet inverse transform unit. The input image data X is subjected to a discrete Haar wavelet transform of the wavelet transform unit to obtain four component data in the wavelet domain. Using the four components of the wavelet domain as training data has the obvious advantage over a single image domain in that low-frequency information improves the global topology and high-frequency information enhances the restoration of texture information. The discrete Haar wavelet transform obtains the wavelet domain component matrix The calculation formula is as follows:

[0020]

[0021] Among them, Q is the coefficient matrix of discrete Haar wavelet transform, Q T is the transposed matrix of Q. The discrete Haar wavelet transform is expanded as:

[0022]

[0023] in, A, H, V, and D represent the low-frequency smooth output component, horizontal high-frequency component, vertical high-frequency component, and diagonal high-frequency component, respectively. Different components contain different details and texture information.

[0024] The wavelet domain component matrix obtained by discrete Haar wavelet transform is passed to the feature correction unit and multi-scale denoising module, and finally passed to the inverse wavelet transform unit for inverse wavelet transform. The denoised data X′ is obtained.

[0025] Preferably, the calculation formula of step 6 is as follows:

[0026] F out =N(Concat(F1, F2))

[0027] Where N is the sampling-optimized non-local attention expression, Concat is the concatenation function, F1 is the denoised data tensor form of the wavelet domain submodel, and F2 is the denoised data tensor form of the image domain submodel. The sizes of F1 and F2 are equal. F1 and F2 are concatenated and converted into a visual format to obtain the denoised brain MR image data.

[0028] The non-local attention expression in the sampling-optimized non-local attention module is as follows:

[0029] N=softmax(W Q x·sm(W K x) T )sm(V(x))

[0030] V(x)=W V x

[0031] Among them, softmax is the normalized exponential function, sm() is the output of the optimized sampler, W Q 、W K and W V These are three independent feature maps after linear transformation of the three 1*1 convolutional layers in the sampling-optimized non-local attention module of the input preliminary denoised data x. V(x) represents the feature extraction information of the input preliminary denoised data x.

[0032] The optimized sampler specifically includes a feature mapping layer, a spatial mixing multilayer perceptron, a channel mixing multilayer perceptron and a pixel reorganization layer; the feature mapping layer divides the input data into blocks and converts them into a list, and then passes through a spatial mixing multilayer perceptron and a channel mixing multilayer perceptron in sequence to complete the fusion of inter-region information, and then passes through the pixel reorganization layer to realize the reorganization from channel to space to transform it into a feature map reduced by a given multiple, and finally passes through another channel mixing multilayer perceptron after fusion as the output of the optimized sampler.

[0033] Among them, the spatial hybrid multilayer perceptron consists of two fully connected layers, a Gaussian error linear unit activation function and a layer normalization function LayerNorm, which is expressed as:

[0034] MLP s =ω2σ(ω1LayerNorm(X γ )), γ=1…s

[0035] X γ Represents spatial hybrid multilayer perceptron MLP s The γth spatial information of the current input, ω1 and ω2 represent two different fully connected layers, σ is the Gaussian error linear unit activation function, and s represents the number of blocks.

[0036] The channel mixture multilayer perceptron consists of a fully connected layer and a layer normalization function, expressed as

[0037] MLP c =ω3LayerNorm(X η ), η=1…c

[0038] X η Represents a channel-mixed multilayer perceptron MLP c The nth channel information of the current input, ω3 is the MLP c In the fully connected layer, c represents the number of channels.

[0039] Preferably, the loss value of the loss function MSE in step 8 is loss MSE The calculation is as follows:

[0040]

[0041] Where M and N represent the length and width of the input image of the MDNet denoising model, respectively, and I tar and I gt They represent the brain MR image data after denoising in step 6 and the clean MR image data with a Rice noise level of 0% in the training set, respectively. The row number is i∈(1, M) and the column number is j∈(1, N).

[0042] Compared with the prior art, the present invention has the following specific beneficial effects: Based on the image domain, the wavelet domain is added to train the above network. Due to the addition of the four components of the wavelet domain, not only the learning samples of the model are enriched, but the characteristics of the high-frequency components in different directions also enable the wavelet domain sub-model to pay attention to the structural similarity of the image, making the edges and contours of the reconstructed image clearer; furthermore, the present invention innovatively proposes a sampling-optimized non-local attention module, which is a recalibration of each weight parameter of the feature map globally, and based on the sampling optimization, the feature map is reduced to a predetermined multiple. Therefore, the lightweight effect of the sampler can alleviate the computing power during training without destroying the global nature of the features. After the outputs of the image domain sub-model and the wavelet domain sub-model are feature-corrected, their denoising results are further improved by the sampling-optimized non-local attention module. Therefore, the present invention solves the problem of processing noise-contaminated pixels in MR images, especially reduces the influence of Rician noise in brain MRI images, and ultimately greatly improves the ability to restore high contrast between gray matter and white matter areas from noisy images. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the present invention;

[0044] Figure 2 Schematic diagram of the MDNet denoising model in the present invention;

[0045] Figure 3 Schematic diagram of the structure of the multi-scale denoising module in the present invention;

[0046] Figure 4 Schematic diagram of the discrete Haar wavelet transform structure of the wavelet transform unit in the present invention;

[0047] Figure 5 Schematic diagram of the structure of the sampling optimized non-local attention module in the present invention;

[0048] Figure 6 This is a schematic diagram of the structure of the optimized sampler in the present invention;

[0049] Figure 7 The figure is a comparison of images before and after noise reduction using the method of the present invention. DETAILED DESCRIPTION

[0050] In order to facilitate understanding by those skilled in the art and to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described below with reference to the following embodiments. It should be understood that the description herein is merely illustrative and does not limit the scope of the present invention.

[0051] like Figure 1 As shown, the multi-domain optimization model construction method for reducing Rice noise in brain magnetic resonance imaging of the present invention mainly includes the following steps:

[0052] S1. Acquire brain MR image data with different noise levels, and perform cross-sectional slicing processing on the raw data (RawData) of the brain MR image data.

[0053] S2. Perform redundancy removal on the sliced ​​data, filtering out invalid slices. Create a dataset of clean and noisy images, dividing the dataset into training, validation, and test sets in a ratio of 8:1:1. De-redundancy removal involves removing invalid slices below the neck. These removed slices exhibit characteristic white-gray stripes.

[0054] S3. Establish the MDNet denoising model, including the image domain sub-model and the wavelet domain sub-model.

[0055] like Figure 2 As shown in the figure, the image domain sub-model includes a feature correction unit and a multi-scale denoising module. The feature correction unit is used to constrain the channel and size parameters of the feature map, and the multi-scale denoising module is used to capture more contextual information and a wider range of noise distribution characteristics.

[0056] The feature correction unit consists of a 1*1 convolutional layer, a batch normalization layer, and a parameter-corrected linear unit activation function.

[0057] The multi-scale denoising module uses dilated convolutions with different dilation rates and parallel convolution operations with large parameter convolution kernels. Figure 3 As shown in the figure, the input is passed into the three-branch structure after the preliminary convolution operation. This structure combines the operations of two dilated convolutions and one 5*5 large parameter convolution kernel. The convolution kernels of the two dilated convolutions are 3*3, and their dilation rates are set to 2 and 3 respectively; the multi-scale denoising module is finally output through the confluence layer and subjected to convolution operation and residual connection to fuse the noise features under multiple receptive fields.

[0058] In the wavelet-domain sub-model, the input image data undergoes a discrete Haar wavelet transform in the wavelet transform unit, yielding four components in the wavelet domain. These components are then fed into the feature correction unit and the multi-scale denoising module, before being output as denoised data through the inverse wavelet transform unit. A, H, V, and D in the discrete Haar wavelet transform result represent the low-frequency smoothed output component, the horizontal high-frequency component, the vertical high-frequency component, and the diagonal high-frequency component, respectively. Different components contain different details and texture information. Using the four components in the wavelet domain as the model's input data offers a significant advantage over using a single image-domain input: low-frequency information improves global topology, while high-frequency information enhances the recovery of texture information. Figure 4 The discrete Haar wavelet transform process is shown in Figure 1. It is divided into two stages. The first stage is to pass the low-pass filter G L and high-pass filter G H Signal extraction is performed along the columns respectively. After the two extracted signals are downsampled by a factor of 2↓2, the second stage continues to perform signal extraction along the rows on each result of the first stage through a low-pass filter and a high-pass filter. The four extracted signals are then downsampled by a factor of 2 to finally obtain the four components A, H, V, and D.

[0059] S4. Input the brain MR image data with Rician noise levels of 3%, 5%, 7% and 9% in the training set into the image domain sub-model of the MDNet denoising model, and use the obtained output value as the denoising data of the image domain sub-model.

[0060] S5. Input the brain MR image data with Rice noise levels of 3%, 5%, 7% and 9% in the training set into the wavelet domain sub-model of the MDNet denoising model, and use the obtained output value as the denoising data of the wavelet domain sub-model.

[0061] S6. Concatenate the denoised data obtained in S4 and the denoised data obtained in S5 as preliminary denoised data, and pass them into the sampling-optimized non-local attention module to obtain the final denoised brain MR image data.

[0062] The formula of the non-local attention module for optimized sampling is as follows:

[0063] Fout =N(Concat(F1, F2))

[0064] Wherein, N represents the sampling optimized non-local attention expression, Concat is the connection function, F1 is the denoised data tensor form of the wavelet domain sub-model, F2 is the denoised data tensor form of the image domain sub-model, the sizes of F1 and F2 are equal, F1 and F2 are connected and converted into a visual format to obtain the denoised brain MR image data. The sampling optimized non-local attention module in the present invention is a recalibration of each weight parameter of the global feature map, and reduces the feature map to a given multiple based on sampling optimization. Therefore, the lightweight effect of the sampler can alleviate the computing power during training without destroying the global nature of the features. After the outputs of the image domain sub-model and the wavelet domain sub-model are feature-corrected, their denoising results are further improved by the sampling optimized non-local attention module.

[0065] The non-local attention expression in the sampling-optimized non-local attention module is as follows:

[0066] N=softmax(W Q x·sm(W K x) T )sm(V(x))

[0067] V(x)=W V x

[0068] Among them, softmax is the normalized exponential function, sm() is the output of the optimized sampler, W Q 、W K and W V These are three independent feature maps after linear transformation of the three 1*1 convolutional layers in the sampling-optimized non-local attention module of the input preliminary denoised data x. V(x) represents the feature extraction information of the input preliminary denoised data x.

[0069] like Figure 5 As shown in the non-local attention module with sampling optimization, the input is linearly transformed by three 1*1 convolutional layers, and W Q The shape parameter of the branch after deformation is HW*C / 2, where H represents the height of the feature map, W represents the width of the feature map, and C represents the number of feature map channels. K and W V After the branch is optimized and the sampler is used to reduce the feature map (from HW to sm(HW)) and convert the dimension, the shape parameters become C / 2*sm(HW) and sm(HW)*C / 2 respectively; then, the two feature maps corresponding to HW*C / 2 and C / 2*sm(HW) are multiplied to calculate the autocorrelation in the feature, and then normalized to obtain a non-local attention coefficient of 0 to 1; finally, this coefficient is multiplied by WV The branch feature map is then added to the original input x.

[0070] like Figure 6 As shown in the figure, the optimized sampler specifically includes a feature mapping layer, a spatial mixing multi-layer perceptron (Token-mixing MLP), a channel mixing multi-layer perceptron (Channel-mixing MLP) and a pixel reorganization layer (Pixelshuffle); a tensor data with a shape parameter of H*W*C is input into the optimized sampler, the feature mapping layer divides the input data into blocks (Patch) and converts it into a list, and then passes through a spatial mixing multi-layer perceptron and a channel mixing multi-layer perceptron to complete the fusion of inter-region information, and then passes through the pixel reorganization layer to realize the reorganization from channel to space to transform it into a feature map reduced by a given multiple i (H and W are reduced to H / i and W / i respectively), and finally passes through another channel mixing multi-layer perceptron after fusion as the output of the optimized sampler, whose shape parameter is H / i*W / i*C.

[0071] Among them, the spatial hybrid multilayer perceptron consists of two fully connected layers, a Gaussian error linear unit (GELU) activation function and a layer normalization function LayerNorm, which is expressed as:

[0072] MLP s =ω2σ(ω1LayerNorm(X γ )), γ=1…s

[0073] X γ Represents spatial hybrid multilayer perceptron MLP s The γth spatial information of the current input, ω1 and ω2 represent two different fully connected layers, σ is the Gaussian error linear unit activation function, and s represents the number of blocks.

[0074] The channel mixture multilayer perceptron consists of a fully connected layer and a layer normalization function, expressed as

[0075] MLP c =ω3LayerNorm(X η ), η=1…c

[0076] X η Represents a channel-mixed multilayer perceptron MLP c The nth channel information of the current input, ω3 is the MLP c In the fully connected layer, c represents the number of channels.

[0077] S7. Set the loss function MSE of the MDNet denoising model to measure the error between the denoised brain MR image data and the clean MR image data with a Rice noise level of 0% in the training set.

[0078] S8. Use the training set generated by S2 to train the MDNet denoising model, and use the validation set to evaluate the training process. When the loss value of the loss function MSE converges to a value lower than the preset value (for example, 0.001), stop training and save the parameters of the MDNet denoising model.

[0079] The present invention uses the loss function MSE to optimize the model, taking the denoised data output after the noise image in the training set passes through the image domain sub-model and the wavelet domain sub-model as the estimated value, and the original noise-free image in the data set as the GroundTruth. MSE The calculation is as follows:

[0080]

[0081] Where M and N represent the length and width of the input image of the MDNet denoising model, respectively, and I tar and I gt They represent the brain MR image data after denoising in step 6 and the clean MR image data with a Rice noise level of 0% in the training set, respectively. The row number is i∈(1, M) and the column number is j∈(1, N).

[0082] The present invention is built using the PyTorch1.4.0 platform. During the training process, the number of cycles is 100, the batch size is 12, Adam is set as the optimizer of the model, the initial learning rate is 0.001, the decay rate is 0.00001, and the training is stopped after the loss value of the loss function converges to a value lower than the preset value (0.001), and the weight file of the total network is saved.

[0083] Figure 7 The figure shows the denoising results of noisy brain MR image data in the test set. The left side shows the original noisy image, and the right side shows the brain MR image after noise reduction using the method of the present invention. It can be seen that the multi-domain optimization model construction method for reducing Rician noise in brain MRI images provided by the present invention can reconstruct high-quality MR images from data contaminated by Rician noise.

Claims

1. A multi-domain optimization model construction method for reducing Rice noise in brain magnetic resonance imaging, characterized by: The following steps are involved: Step 1: Acquire brain MR image data with different noise levels, and perform cross-sectional slicing on the brain MR image data; Step 2: De-redundancy processing is performed on the sliced ​​data, invalid slices are filtered out, and the data set is divided into training set, validation set and test set; Step 3: Establish an MDNet denoising model, including an image domain sub-model and a wavelet domain sub-model; Step 4: Input the brain MR image data with Rician noise levels of 3%, 5%, 7%, and 9% in the training set into the image domain sub-model of the MDNet denoising model, and use the obtained output values ​​as the denoised data of the image domain sub-model; Step 5: Input the brain MR image data with Rice noise levels of 3%, 5%, 7%, and 9% in the training set into the wavelet domain sub-model of the MDNet denoising model, and use the obtained output values ​​as the denoised data of the wavelet domain sub-model; Step 6: Concatenate the denoised data obtained in step 4 and the denoised data obtained in step 5 as preliminary denoised data, and pass them into the sampling-optimized non-local attention module to obtain the final denoised brain MR image data; Step 7: Set the loss function MSE of the MDNet denoising model to measure the error between the denoised brain MR image data and the clean MR image data with a Rice noise level of 0% in the training set; Step 8. Use the training set generated in step 2 to train the MDNet denoising model, and use the validation set to evaluate the training process. When the loss value of the loss function MSE converges to a value lower than the preset value, stop training and save the parameters of the MDNet denoising model.

2. The method for constructing a multi-domain optimization model for reducing Ricean noise in brain magnetic resonance imaging according to claim 1, characterized in that: After executing step 8, input the test set into the trained MDNet denoising model to test the performance of the MDNet denoising model.

3. The method for constructing a multi-domain optimization model for reducing Ricean noise in brain magnetic resonance imaging according to claim 1, characterized in that: In step 2, the training set, validation set, and test set all contain clean MR image data with a Rice noise level of 0%, and noise-contaminated brain MR image data with Rice noise levels of 3%, 5%, 7%, and 9%; the noise-contaminated brain MR image data are paired with slices at the same spatial position in the clean MR image data; The redundancy removal process is to remove invalid slices below the human neck.

4. The method for constructing a multi-domain optimization model for reducing Ricean noise in brain magnetic resonance imaging according to claim 1, characterized in that: The image domain sub-model includes a feature correction unit and a multi-scale denoising module. The feature correction unit is used to constrain the channel and size parameters of the feature map, and the multi-scale denoising module is used to capture context information and noise distribution characteristics. The feature correction unit consists of a 1*1 convolutional layer, a batch normalization layer and a parameter-corrected linear unit activation function; The input of the multi-scale denoising module passes through a preliminary 3*3 convolution layer, a batch normalization layer, and a parameter-corrected linear unit activation function and is then passed into a three-branch structure. The three-branch structure combines the operations of two dilated convolutions and one 5*5 large-parameter convolution kernel, where the convolution kernels of the two dilated convolutions are 3*3, and their dilation rates are set to 2 and 3, respectively. Each branch also passes through a batch normalization layer and a parameter-corrected linear unit activation function and is merged at a confluence layer. Finally, the output passes through a 3*3 convolution layer, a batch normalization layer, and a parameter-corrected linear unit activation function and is residually connected to fuse noise features under multiple receptive fields.

5. The method for constructing a multi-domain optimization model for reducing Ricean noise in brain magnetic resonance imaging according to claim 1, characterized in that: The wavelet domain sub-model is composed of a wavelet transform unit, a feature correction unit, a multi-scale denoising module and an inverse wavelet transform unit; the input image data X is subjected to a discrete Haar wavelet transform by the wavelet transform unit to obtain four component data in the wavelet domain, and the four components in the wavelet domain are used as training data; The discrete Haar wavelet transform obtains the wavelet domain component matrix The calculation formula is as follows: Among them, Q is the coefficient matrix of discrete Haar wavelet transform, Q T is the transposed matrix of Q; the discrete Haar wavelet transform is expanded as: in, A, H, V, and D represent the low-frequency smooth output component, horizontal high-frequency component, vertical high-frequency component, and diagonal high-frequency component, respectively. Different components contain different details and texture information; The wavelet domain component matrix obtained by discrete Haar wavelet transform is passed to the feature correction unit and multi-scale denoising module, and finally passed to the inverse wavelet transform unit for inverse wavelet transform. The denoised data X′ is obtained.

6. The method for constructing a multi-domain optimization model for reducing Rice noise in brain magnetic resonance imaging according to claim 1, characterized in that: The calculation formula for step 6 is as follows: F out =N(Concat(F1,F2)) Where N is the sampling-optimized non-local attention expression, Concat is the concatenation function, F1 is the denoised data tensor form of the wavelet domain sub-model, and F2 is the denoised data tensor form of the image domain sub-model. The sizes of F1 and F2 are equal. Concatenate F1 and F2 and convert them into a visual format to obtain the denoised brain MR image data. The non-local attention expression in the sampling-optimized non-local attention module is as follows: N=softmax(W Q x·sm(W K x) T )sm(V(x)) V(x)=W V x Among them, softmax is the normalized exponential function, sm() is the output of the optimized sampler, W Q 、W K and W V The three independent feature maps after linear transformation of the three 1*1 convolutional layers in the sampling-optimized non-local attention module after the initial denoising data x are input. V(x) represents the feature extraction information of the initial denoising data x. The optimized sampler includes a feature mapping layer, a spatial hybrid multilayer perceptron, a channel hybrid multilayer perceptron, and a pixel reassembly layer; the feature mapping layer divides the input data into blocks and converts them into a list, then sequentially passes through a spatial hybrid multilayer perceptron and a channel hybrid multilayer perceptron to complete the fusion of inter-region information, and then passes through the pixel reassembly layer to achieve channel-to-space reassembly to transform it into a feature map reduced by a predetermined multiple, and finally passes through another channel hybrid multilayer perceptron to fuse and serve as the output of the optimized sampler; Among them, the spatial hybrid multilayer perceptron consists of two fully connected layers, a Gaussian error linear unit activation function and a layer normalization function LayerNorm, which is expressed as: MLP s =ω2σ(ω1LayerNorm(X γ )),γ=1…s X γ Represents spatial hybrid multilayer perceptron MLP s The γth spatial information of the current input, ω1 and ω2 represent two different fully connected layers, σ is the Gaussian error linear unit activation function, and s represents the number of blocks; The channel mixture multilayer perceptron consists of a fully connected layer and a layer normalization function, expressed as MLP c =ω3LayerNorm(X η ),η=1…c X η Represents a channel-mixed multilayer perceptron MLP c The nth channel information of the current input, ω3 is the MLP c In the fully connected layer, c represents the number of channels.

7. The method for constructing a multi-domain optimization model for reducing Ricean noise in brain magnetic resonance imaging according to claim 1, characterized in that: The loss value of the loss function MSE in step 8 is loss MSE The calculation is as follows: Where M and N represent the length and width of the input image of the MDNet denoising model, respectively, and I tar and I gt They represent the brain MR image data after denoising in step 6 and the clean MR image data with a Rice noise level of 0% in the training set, respectively. The row number is i∈(1, M) and the column number is j∈(1, N).

Images