Mixed domain diffusion magnetic resonance image denoising method and related device

By adopting a hybrid domain denoising method in dMRI images, using parallel connections between the frequency domain and the spatial domain and the fusion subnet module, the dMRI image noise problem is solved, significantly improving the denoising performance and image quality, and improving clinical and scientific research data reliability.

CN119090757APending Publication Date: 2024-12-06SHAANXI NORMAL UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411007256.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In magnetic resonance imaging technology, diffusion-weighted imaging (dMRI) images increase background noise due to signal scattering, which reduces the image clarity and signal-to-noise ratio, affects the diagnostic accuracy and reliability of scientific research data. The existing repeated scan superimposed average Strategies are inefficient in clinical practice.

Method used

Using the hybrid domain denoising method, the dMRI denoising model is trained to recover the high signal-to-noise image from the low signal-to-noise image through parallel connection between the frequency domain and the spatial domain and the fusion subnet module, combined with the multi-scale structure similarity loss function. .

Benefits of technology

It significantly improves the denoising performance of diffuse magnetic resonance images, reduces noise, improves the clarity and signal-to-noise ratio of the image, and improves the reliability of diagnostic and scientific research data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090757A_ABST
    Figure CN119090757A_ABST
Patent Text Reader

Abstract

The invention discloses a mixed domain diffusion magnetic resonance image denoising method and related device, and the method comprises the steps: obtaining diffusion magnetic resonance imaging dMRI data, and carrying out the preprocessing, and obtaining a dMRI image needed by model training; de-noising the dMRI image required by model training obtained after preprocessing through a pre-constructed frequency domain de-noising sub-network module, a pre-constructed spatial domain de-noising sub-network module and a pre-constructed frequency domain-spatial domain fusion sub-network module; the frequency domain denoising subnet module and the spatial domain denoising subnet module are connected in parallel, and after combined output, the frequency domain denoising subnet module and the spatial domain denoising subnet module are connected in series with the frequency domain-spatial domain fusion subnet module to construct a dMRI denoising model; and a multi-scale structure similarity loss function is constructed, the constructed dMRI denoising model is trained, and an output image with a high signal-to-noise ratio is recovered from the noise image with the low signal-to-noise ratio through the dMRI denoising model. The frequency domain subnet and the spatial domain subnet respectively process data and extract features in respective domains, the fusion subnet can fully combine the advantages of the two subnet modules, and the denoising performance of the whole model is further improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of magnetic resonance medical image processing, and in particular relates to a mixed domain diffusion magnetic resonance image denoising method and a related device. Background Art

[0002] Diffusion-weighted imaging (dMRI) in the field of magnetic resonance imaging is a key technology for exploring the microstructural characteristics of biological tissues, especially in neuroscience research. With its high sensitivity to capture the dynamics of water molecule diffusion, it realizes non-invasive visualization of the complex structure of the human body, especially the white matter of the brain, and promotes our deep understanding of the functional organization and development process of the brain. dMRI technology enables scientists to describe the microscopic wiring diagram of the brain in detail, track the path of white matter fiber bundles, monitor microstructural changes associated with age, disease or recovery, and accurately identify abnormal connection patterns, becoming an indispensable tool for studying neurological diseases. In dMRI practice, a core parameter, the B value, plays a key role in regulating diffusion sensitivity. Although enhancing the B value can improve the image's resolution of subtle diffusion characteristics, it also amplifies the signal scattering phenomenon caused by the natural random movement of water molecules. This physical process is reflected in the final image as increased background noise, thereby reducing the image clarity and signal-to-noise ratio. Noise not only blurs tissue boundaries and subtle features, but may also mislead the quantitative analysis of diffusion parameters, affecting diagnostic accuracy and the reliability of scientific research data. In view of this, effectively suppressing noise pollution in dMRI images has become an urgent issue to be solved in this field. Currently, the repeated scanning stacking average strategy is the most widely adopted noise reduction method, which reduces random noise in a statistical way by accumulating multiple scanning data, but its time-consuming and resource-intensive characteristics limit its wide application and efficiency in clinical practice. Summary of the invention

[0003] The object of the present invention is to provide a method and related device for denoising mixed-domain diffusion magnetic resonance images, so as to solve the problem of poor data availability caused by severe noise influence on image data acquired in clinical environment.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] In a first aspect, the present invention provides a method for denoising a mixed-domain diffusion magnetic resonance image, comprising:

[0006] Acquire diffusion magnetic resonance imaging (dMRI) data and obtain the dMRI images required for model training after preprocessing;

[0007] The dMRI images required for model training obtained after preprocessing are denoised through the pre-built frequency domain denoising subnet module, space domain denoising subnet module and frequency domain-space domain fusion subnet module;

[0008] The frequency domain denoising subnet module is connected in parallel with the spatial domain denoising subnet module, and the combined output is then connected in series with the frequency domain-spatial domain fusion subnet module to construct a dMRI denoising model.

[0009] A multi-scale structural similarity loss function is constructed, and the constructed dMRI denoising model is trained. The dMRI denoising model is used to recover the output image with high signal-to-noise ratio from the noisy image with low signal-to-noise ratio.

[0010] Furthermore, diffusion magnetic resonance imaging (dMRI) data is obtained and preprocessed to obtain the dMRI images required for model training, including:

[0011] The original multi-shell data of the Human Connectome Project (HCP) were first extracted according to the size of the b value to obtain a single-shell data set aggregated by the b value; then, the K-mean clustering algorithm was used in each b-value data set, and 4 typical b-vector directions were selected as the central vectors of the clustering to obtain 4 clusters containing multiple single-shell data. Then, geometric averaging was performed in each cluster to obtain a cleaner image for model training; during training, according to different b values, noise in the range of [0,20]% and [0,50]% was added to the averaged data respectively. During testing, noise was added to the original HCP data set in the same way and used as the input data of the network;

[0012] For clinical quality dMRI data, the dMRI data first needs to be processed by motion correction, spatial registration, distortion correction, and eddy current correction. When restoring low signal-to-noise ratio noisy data to high signal-to-noise ratio output data, the three-dimensional volume data needs to be sliced ​​and the obtained slices are input into the model in batches of 32. Two-dimensional data needs to pass through an additional convolutional layer when inputting the model and before outputting the results to convert the number of feature channels to match the data format required by the model.

[0013] Furthermore, a frequency domain denoising subnet module is constructed to enable the model to learn the frequency domain feature information of the image and separate the effective signal from the noise signal in the frequency domain:

[0014] The dual discrete cosine transform DDCT is composed of two scales of DCT blocks, which perform DCT transforms in different ranges and obtain the corresponding frequency coefficient matrix; the obtained multi-scale parallel data are input into different branches for feature learning, and each branch is composed of four frequency module blocks FMB; each FMB is composed of several four-expanded convolutions with a receptive field of 3×3 and conventional convolutions and a multi-scale fast channel attention, which is used to efficiently extract data features; then, the results obtained on each branch are fused through another convolution layer and input into the next DDCT structure; after all modules are processed, a frequency domain subnet model that has learned the frequency domain feature information is obtained, and the output result of the frequency domain subnet is obtained.

[0015] Furthermore, a spatial domain denoising subnet module is constructed to remove noise from the image in the spatial domain:

[0016] The spatial domain subnet module adopts a convolutional neural network structure, which consists of 12 spatial module blocks SMB; each SMB consists of a 3×3 convolution, a four-expanded convolution and a multi-scale spatial attention; the input of the SMB is subjected to a 3×3 convolution and a four-expanded convolution for feature extraction, and feature weighting is performed through a multi-scale spatial attention, and then connected to the input of the SMB through a residual connection. Finally, the final output result of the SMB is obtained through another 3×3 convolution, batch normalization and activation function; after processing through all SMBs, the final output result of the spatial domain subnet is obtained.

[0017] Furthermore, a frequency domain-spatial domain fusion subnet module is constructed to fuse the results of the frequency domain subnet and the spatial domain subnet:

[0018] The frequency-space fusion subnet module adopts a convolutional neural network and consists of two fusion blocks FB; each FB is composed of a four-dilated convolution and several conventional convolutions in series, and a residual connection is integrated; the input of each FB will first be subjected to multi-scale feature extraction through four-dilated convolution to enhance the ability to capture features; then, the input will be further subjected to three conventional convolutions for further feature extraction and fusion, and the result will be connected to the input of the FB through a residual connection; finally, the final output result of the FB is obtained through 3×3 convolution, batch normalization and activation function.

[0019] Furthermore, the multi-scale fast channel attention module and the multi-scale spatial attention module are integrated into the frequency domain denoising subnet module and the spatial domain denoising subnet module respectively:

[0020] The multi-scale fast channel attention will be integrated into FMB, and the multi-scale spatial attention will be integrated into SMB; the multi-scale fast channel attention module consists of three parts: a multi-scale feature extraction part, a channel attention calculation part, and a channel attention weighting part; the multi-scale feature extraction part is mainly composed of multiple parallel convolution kernels of different sizes, which are used to extract feature information of different scales; the channel attention calculation part needs to reduce the dimensionality of two-dimensional feature maps of different scales into one-dimensional feature vectors through adaptive average pooling, and learn the interaction between feature vectors through 1D convolution, and then concatenate the feature vectors of each scale and perform nonlinear mapping through the Sigmoid activation function to obtain the calculated channel Attention; the channel attention weighting part concatenates the feature information of each scale and multiplies it with the channel attention between channels to obtain the feature cube after the channel feature is weighted; four branches are designed, each of which consists of two 3×3 convolutions with expansion rates d=1, 2, 3, and 4 respectively; the input of multi-scale spatial attention will be input into each branch respectively, and feature extraction and learning will be performed respectively, and finally the outputs of these branches will be concatenated, and feature fusion will be performed through another convolution to obtain a two-dimensional feature distribution map; the obtained two-dimensional feature distribution map is nonlinearly mapped through the Sigmoid activation function to obtain the final weight distribution map; the weight distribution map is multiplied with the input to obtain the spatially weighted feature cube.

[0021] Furthermore, the overall dMRI denoising model is constructed:

[0022] The frequency domain denoising subnet module and the space domain denoising subnet module are connected in parallel to construct the denoising main network. The input of the model will be sent to the frequency domain subnet module and the space domain subnet module for feature extraction and learning in the frequency domain and space domain respectively; then, for the output of the frequency domain subnet module and the space domain subnet module, the two are spliced ​​with the input to realize residual connection, and then the spliced ​​result is input into the feature fusion subnet for feature fusion; finally, before outputting the final result, the residual connection is introduced again with the input to obtain the final output result; the clean "ground truth" is obtained by extracting, clustering and geometrically averaging the HCP dMRI data set, and Rice noise within a certain horizontal range is added as noise data to obtain a data set containing clean-noise image pairs as the training set of the model; the model takes 32 two-dimensional noise images as a batch as input data each time, and outputs denoised high signal-to-noise ratio two-dimensional images of the corresponding batch number;

[0023] The output of the model and the "ground truth" data are respectively compared through the multi-scale structural similarity loss function to calculate the loss, and during the training process, the parameters in the model are adjusted to minimize the loss, thereby optimizing the model and improving the denoising performance of the model.

[0024] In a second aspect, the present invention provides a system for denoising a diffusion magnetic resonance image, comprising:

[0025] A data acquisition module is used to acquire diffusion magnetic resonance imaging (dMRI) data and obtain dMRI images required for model training after preprocessing;

[0026] A denoising module is used to denoise the dMRI images required for model training obtained after preprocessing through a pre-built frequency domain denoising subnet module, a space domain denoising subnet module and a frequency domain-space domain fusion subnet module;

[0027] A model building module is used to connect the frequency domain denoising subnet module and the space domain denoising subnet module in parallel, and then connect the combined output to the frequency domain-space domain fusion subnet module in series to build a dMRI denoising model;

[0028] The output module is used to construct a multi-scale structural similarity loss function, train the constructed dMRI denoising model, and recover a high signal-to-noise ratio output image from a low signal-to-noise ratio noisy image through the dMRI denoising model.

[0029] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of a method for denoising mixed-domain diffusion magnetic resonance images when executing the computer program.

[0030] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for denoising mixed-domain diffusion magnetic resonance images.

[0031] Compared with the prior art, the present invention has the following technical effects:

[0032] The present invention uses deep learning technology to denoise diffuse magnetic resonance images in a hybrid domain composed of a spatial domain and a domain. Compared with the traditional method of denoising diffuse magnetic resonance images, the method has significant denoising performance advantages. First, the present invention designs a new spatial domain subnet module, which is composed of several spatial domain blocks (SBs). Each SB contains several convolution blocks, residual connections and multi-scale spatial attention mechanisms, which can make full use of the effective signal and noise characteristics of diffuse magnetic resonance images in the spatial domain, thereby effectively reducing noise. Secondly, the present invention proposes a new frequency domain subnet module, which is composed of multiple dual discrete cosine transform modules (DDCTBs). Each DDCTB contains two branches, each of which includes a discrete cosine transform (DCT), several convolution modules, residual connections and a multi-scale fast channel attention mechanism. The design converts the original data in the spatial domain to the frequency domain, so that the model can efficiently learn and distinguish the effective information and noise characteristics in the frequency domain, and realize noise removal in the frequency domain. Finally, the present invention designs a fusion network, which is composed of a fusion block (FB) through several convolution blocks and residual connections, and several FBs are connected in series in the fusion subnet. Since the frequency domain subnet and the spatial domain subnet process data and extract features in their respective domains, the fusion subnet can fully combine the advantages of these two subnet modules to further improve the denoising performance of the overall model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The overall logic block diagram of the present invention.

[0034] Figure 2 Schematic diagram of the overall network structure of the present invention.

[0035] Figure 3 Schematic diagram of the frequency domain subnet module of the present invention.

[0036] Figure 4 Schematic diagram of the spatial domain subnet module of the present invention.

[0037] Figure 5 Schematic diagram of the frequency domain-spatial domain feature fusion subnet module of the present invention.

[0038] Figure 6 Schematic diagram of the multi-scale fast-path attention module of the present invention.

[0039] Figure 7 Schematic diagram of the multi-scale spatial attention module of the present invention.

[0040] Figure 8 Schematic diagram of four extended convolution modules of the present invention.

[0041] Fig. 9 The present invention is at b = 1000s / mm 2 Schematic diagram of the comparison between the denoising results of diffusion magnetic resonance images and other methods.

[0042] Fig.10 The present invention is at b = 3000s / mm 2 Schematic diagram of the comparison between the denoising results of diffusion magnetic resonance images and other methods.

[0043] Fig.11 Flowchart of the present invention. DETAILED DESCRIPTION

[0044] This section will describe the exemplary embodiments in detail, examples of which are shown in the accompanying drawings. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are only examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0045] Example 1, please refer to Fig.11 , a mixed domain diffusion magnetic resonance image denoising method, comprising:

[0046] Acquire diffusion magnetic resonance imaging (dMRI) data and obtain the dMRI images required for model training after preprocessing;

[0047] The dMRI images required for model training obtained after preprocessing are denoised through the pre-built frequency domain denoising subnet module, space domain denoising subnet module and frequency domain-space domain fusion subnet module;

[0048] The frequency domain denoising subnet module is connected in parallel with the spatial domain denoising subnet module, and the combined output is then connected in series with the frequency domain-spatial domain fusion subnet module to construct a dMRI denoising model.

[0049] A multi-scale structural similarity loss function is constructed, and the constructed dMRI denoising model is trained. The dMRI denoising model is used to recover the output image with high signal-to-noise ratio from the noisy image with low signal-to-noise ratio.

[0050] Example 2, please refer to Figures 1 to 8 , a mixed domain diffusion magnetic resonance image denoising method, comprising:

[0051] Acquire diffusion magnetic resonance imaging (dMRI) data and perform corresponding preprocessing to obtain dMRI images required for model training;

[0052] Construct a frequency domain denoising subnet module to enable the model to learn the frequency domain feature information of the image and separate the effective signal from the noise signal in the frequency domain;

[0053] Construct a spatial domain denoising subnet module so that the model can learn the noise characteristics in the spatial domain and then remove the noise in the image in the spatial domain;

[0054] Construct a frequency domain-spatial domain fusion subnet module to fuse the results of the frequency domain subnet and the spatial domain subnet, so that the model can give full play to the unique advantages of the frequency domain denoising network and the spatial domain denoising network;

[0055] The multi-scale fast channel attention module and the multi-scale spatial attention module are integrated into the frequency domain subnet and the spatial domain subnet respectively to increase the model's attention to key features, thereby improving the model's denoising performance.

[0056] The above three sub-network modules are combined to realize the parallel connection of frequency domain sub-network and space domain sub-network, and their joint output is then connected in series with the fusion sub-network to construct the overall dMRI denoising model;

[0057] A multi-scale structural similarity loss function is constructed to train the constructed model, through which a high signal-to-noise ratio output image can be recovered from a low signal-to-noise ratio noisy image.

[0058] Specific:

[0059] The present invention is an example of denoising a diffusion magnetic resonance image. The overall logic flow is as follows: Figure 1 As shown in Figure 2, the overall architecture of the model is as follows: Figure 2 As shown, there are the following steps:

[0060] Data preprocessing:

[0061] The datasets used in training the model and testing were downloaded from the Human Connectome Project (HCP) dataset. The HCP dataset contains data collected using a 1.25mm×1.25mm×1.25mm 3T scanner. The collected data has undergone a rigorous data preprocessing process to provide high-quality, highly consistent, and highly comparable brain imaging data, which is very suitable as a training set and test set for the model. The data of each subject consists of three shells, b=1000s / mm 2 , b=2000s / mm 2 and b = 3000s / mm 2 , and each shell collected signals in 90 gradient directions. In addition, at b = 0s / mm 2 The shell also collects signals in 18 gradient directions. Therefore, the dMRI data size of each subject in the HCP dataset is 145×174×145×288.

[0062] Based on this data, we used b = 1000s / mm 2 and b = 3000s / mm 1The model was trained and experimentally verified using the volume data of the HCP. We chose to extract the original multi-shell data according to the size of the b value to obtain a single-shell data set aggregated by b value. Then, we used the K-mean clustering algorithm in each b value data set and selected 4 typical b-vector directions as the center vectors of the cluster for clustering. In this way, 4 clusters containing multiple single-shell data were obtained. Then, we performed geometric averaging in each cluster to obtain a cleaner image that can be used as model training, namely "ground-truth". During training, depending on the b value, we added noise in the range of [0,20]% and [0,50]% to the averaged data, respectively, to correspond to the fact that the noise level contained in the dMRI data with different b values ​​is different. In this way, clean-noise image pairs can be obtained for training the denoising model. During testing, noise was added to the HCP original data set in the same way and used as the input data of the network.

[0063] For clinical quality dMRI data, the dMRI data first needs to be processed by motion correction, spatial registration, distortion correction, and eddy current correction. When restoring low signal-to-noise ratio noisy data to high signal-to-noise ratio output data, the three-dimensional volume data needs to be sliced ​​and the obtained slices are input into the model in batches of 32. Two-dimensional data needs to pass through an additional convolutional layer when inputting the model and before outputting the results to convert the number of feature channels to match the data format required by the model.

[0064] Construct frequency domain subnet module and integrate multi-scale fast spatial attention mechanism:

[0065] Frequency domain subnet module such as Figure 3 As shown. We use DDCT to convert the frequency domain to the spatial domain in parallel under two-scale DCT blocks, and obtain the corresponding frequency coefficient data; for the obtained multi-scale parallel data, we input them into different branches for feature learning, each branch consists of four frequency module blocks (Frequency ModuleBlock, FMB); each FMB is composed of several four-expanded convolutions with a receptive field of 3×3 and conventional convolutions and a multi-scale fast channel attention combination, which is used to efficiently extract data features; then, for the results obtained on each branch, we fuse them through another convolution layer and input them into the next DDCT structure; after all modules are processed, a frequency domain subnetwork model that has learned the frequency domain feature information will be obtained, and the output result of the frequency domain subnetwork can be obtained. The multi-scale fast channel attention module is shown as follows Figure 6As shown. This module consists of three parts: multi-scale feature extraction part, channel attention calculation part and channel attention weighting part. The multi-scale feature extraction part is mainly composed of multiple parallel convolution kernels of different sizes, and the convolution kernel sizes are set to 1×1, 3×3, 5×5, and 7×7 respectively, which are used to extract feature information of different scales; the channel attention calculation part needs to reduce the dimension of the two-dimensional feature maps of different scales obtained from the multi-scale feature extraction part to one-dimensional feature vectors through adaptive average pooling and learn the interaction between feature vectors through 1D convolution, and then splice the feature vectors of each scale processed by 1D convolution, and perform nonlinear mapping through Sigmoid activation function to obtain the calculated channel attention; the channel attention weighting part splices the feature information of each scale and multiplies the channel attention between channels to obtain the feature cube after the channel feature weighting. This module is used to learn the feature information between multi-scale feature channels and emphasize the feature channels that are beneficial to the results. The output weighted feature cube will be used as the input of the next module.

[0066] Construct a spatial domain subnet module and integrate a multi-scale spatial attention mechanism:

[0067] The spatial domain subnet module is as follows Figure 4 As shown. The spatial domain subnet module also adopts a convolutional neural network structure, which consists of 12 spatial module blocks (SMBs); each SMB consists of a 3×3 convolution, a four-expanded convolution, and a multi-scale spatial attention; the input of the SMB is sequentially subjected to a 3×3 convolution and a four-expanded convolution for feature extraction, and feature weighting is performed through a multi-scale spatial attention, and then connected to the input of the SMB through a residual connection. Finally, the final output result of the SMB is obtained through another 3×3 convolution, batch normalization, and activation function; after all SMBs are processed, the final output result of the spatial domain subnet can be obtained for subsequent processing. In addition, we have also integrated a multi-scale spatial attention in the SMB of the spatial domain subnet, such as Figure 7 As shown in the figure, it is a multi-branch result. We designed four branches, each of which consists of two 3×3 convolutions with expansion rates of d=1, 2, 3, and 4 respectively; the input of the multi-scale spatial attention will be input into each branch respectively, and feature extraction and learning will be performed respectively. Finally, the outputs of these branches are concatenated and feature fusion is performed through another convolution to obtain a two-dimensional feature distribution map; the obtained two-dimensional feature distribution map is nonlinearly mapped through the Sigmoid activation function to obtain the final weight distribution map; the weight distribution map is multiplied by the input to obtain the spatially weighted feature cube.

[0068] Construct frequency domain-spatial domain feature fusion subnet module:

[0069] Feature fusion subnet module such as Figure 5 As shown. The feature fusion subnet adopts the classic convolutional neural network (CNN) architecture, which is composed of multiple fusion blocks connected in series in sequence. In the present invention, a total of 2 such fusion blocks are deployed. The fusion block integrates dilated convolution, standard convolution and residual connection technology. The input of each fusion block will first be subjected to multi-scale feature extraction through four dilated convolutions to enhance the ability to capture features; then, the input will be further subjected to three conventional convolutions for further feature extraction and fusion, and the result will be connected to the input of the fusion block through a residual connection; finally, the final output result of the fusion block is obtained through 3×3 convolution, batch normalization and activation function; and the residual calculation is performed with the input through a jump connection to obtain the final denoising output.

[0070] Construct a multi-scale structural similarity loss function and perform model training:

[0071] The frequency domain subnet module and the spatial domain subnet module are connected in parallel to construct the denoising main network. The input of the model will be sent to the frequency domain subnet module and the spatial domain subnet module for feature extraction and learning in the frequency domain and spatial domain respectively; then, for the output of the frequency domain subnet module and the spatial domain subnet module, we will splice the two with the input to achieve residual connection, and then input the spliced ​​result into the feature fusion subnet for feature fusion; finally, before outputting the final result, we introduce residual connection with the input again and obtain the final output result; the clean "ground truth" is obtained by extracting, clustering and geometrically averaging the dMRI dataset of HCP, and Rice noise within a certain range of levels is added to it as noise data. After performing the same operation on all the data in the original dataset, a dataset containing clean-noise image pairs can be obtained, which can be used as the training set of the model. The model takes 32 two-dimensional noise images as a batch as input data each time, and outputs denoised high signal-to-noise ratio two-dimensional images of the corresponding batch number. The output results of the model and the "ground truth" data are respectively calculated through the multi-scale structural similarity loss function, and the parameters in the model are adjusted during the training process to minimize the loss, thereby optimizing the model and improving the denoising performance of the model.

[0072] The model in the embodiment of the present invention is set as follows:

[0073] 1. Model initialization: When training the model, the normal random initialization method is used to initialize the model. The initialization weight coefficient can be used to help the reasonable initialization of the neural network parameters to promote network training and performance improvement.

[0074] 2. Loss function: Use the multi-scale structural similarity-total variation loss function to train the model. The loss function mainly consists of multi-scale structural similarity loss and total variation loss. Let i and j represent the index of the pixel in the row and column of the image respectively. The multi-scale structural similarity loss at the pixel p_(i,j) is calculated as follows:

[0075]

[0076] Among them, MS-SSIM(p i,j ) is in p i,k The multi-scale structural similarity at , P is the set of all pixels in the image.

[0077] 3. Training details: Adam is used as the optimizer for model training. MultiStepLR is used with a step size of 20 and a decay rate of 0.2 to adjust the learning rate of the model. The total number of training rounds is 100.

[0078] In another embodiment of the present invention, a system for denoising a diffusion magnetic resonance image in a mixed domain is provided, which can be used to implement the above-mentioned method for denoising a diffusion magnetic resonance image. Specifically, the system includes:

[0079] A data acquisition module is used to acquire diffusion magnetic resonance imaging (dMRI) data and obtain dMRI images required for model training after preprocessing;

[0080] A denoising module is used to denoise the dMRI images required for model training obtained after preprocessing through a pre-built frequency domain denoising subnet module, a space domain denoising subnet module and a frequency domain-space domain fusion subnet module;

[0081] A model building module is used to connect the frequency domain denoising subnet module and the space domain denoising subnet module in parallel, and then connect the combined output to the frequency domain-space domain fusion subnet module in series to build a dMRI denoising model;

[0082] The output module is used to construct a multi-scale structural similarity loss function, train the constructed dMRI denoising model, and recover a high signal-to-noise ratio output image from a low signal-to-noise ratio noisy image through the dMRI denoising model.

[0083] In order to verify the effectiveness of the present invention, the inventors used the method of the present invention to compare the denoising effect with eight advanced deep learning denoising methods, namely DnCNN, RDN, RIDNet, RDDCNN, DudeNet, ECNDNet, SADNet and ADNet. Our method and the comparison methods were qualitatively compared on the HCP public dataset. We used b = 1000s / mm 2With b = 3000s / mm 2 The data were qualitatively compared, such as Figure 6 and Figure 7 As shown in Figure 2. Through qualitative analysis, it can be found that our model shows better performance than other deep learning methods and achieves more effective noise removal. In the quantitative experiment, when b = 1000s / mm 2 With b = 3000s / mm 2 In the case of , a complete multi-shell data of the subject in the test set was used as the evaluation unit to calculate and compare the evaluation indicators. The evaluation indicators used were Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Mean Absolute Error (MAE): This is an indicator that directly reflects the deviation between the predicted value and the true value. It calculates the average of the absolute values ​​of all pixel differences. The smaller the MAE value, the smaller the difference between the predicted image and the true image (ground truth), and the better the denoising effect. PSNR is used to measure the ratio of image signal strength to noise level, usually in decibels (dB). The higher the PSNR value, the higher the fidelity of the image and the better the denoising effect. SSIM evaluates the similarity of two images in structural information. The value ranges from 0 to 1. The closer to 1, the better the structural information of the image content is preserved, and the denoising also maintains the original structural characteristics of the image. The quantitative results are shown in Table 1. The present invention has excellent performance in all three quantitative indicators.

[0084] Table 1 Evaluation results using mean absolute error (MAE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The best results are in bold and the suboptimal results are underlined.

[0085]

[0086] In another embodiment of the present invention, a computer device is provided, the computer device comprising a processor and a memory, the memory being used to store a computer program, the computer program comprising program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in a computer storage medium to implement a corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of a method for denoising a diffusion magnetic resonance image.

[0087] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It can be understood that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for denoising a mixed domain diffusion magnetic resonance image in the above embodiment.

[0088] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for denoising mixed-domain diffusion magnetic resonance images, characterized in that: include: Acquire diffusion magnetic resonance imaging (dMRI) data and obtain the dMRI images required for model training after preprocessing; The dMRI images required for model training obtained after preprocessing are denoised through the pre-built frequency domain denoising subnet module, space domain denoising subnet module and frequency domain-space domain fusion subnet module; The frequency domain denoising subnet module is connected in parallel with the spatial domain denoising subnet module, and the combined output is then connected in series with the frequency domain-spatial domain fusion subnet module to construct a dMRI denoising model. A multi-scale structural similarity loss function is constructed, and the constructed dMRI denoising model is trained. The dMRI denoising model is used to recover the output image with high signal-to-noise ratio from the noisy image with low signal-to-noise ratio.

2. The method for denoising mixed-domain diffusion magnetic resonance images according to claim 1, characterized in that: Acquire diffusion magnetic resonance imaging (dMRI) data and obtain the dMRI images required for model training after preprocessing, including: The original multi-shell data of the Human Connectome Project (HCP) were first extracted according to the size of the b value to obtain a single-shell data set aggregated by the b value; then, the K-mean clustering algorithm was used in each b-value data set, and 4 typical b-vector directions were selected as the central vectors of the clustering to obtain 4 clusters containing multiple single-shell data. Then, geometric averaging was performed in each cluster to obtain a cleaner image for model training; during training, according to different b values, noise in the range of [0,20]% and [0,50]% was added to the averaged data respectively. During testing, noise was added to the original HCP data set in the same way and used as the input data of the network; For clinical quality dMRI data, the dMRI data first needs to be processed by motion correction, spatial registration, distortion correction, and eddy current correction. When restoring low signal-to-noise ratio noisy data to high signal-to-noise ratio output data, the three-dimensional volume data needs to be sliced ​​and the obtained slices are input into the model in batches of 32. Two-dimensional data needs to pass through an additional convolutional layer when inputting the model and before outputting the results to convert the number of feature channels to match the data format required by the model.

3. The method for denoising mixed-domain diffusion magnetic resonance images according to claim 1, characterized in that: Construct a frequency domain denoising subnet module to enable the model to learn the frequency domain feature information of the image and separate the effective signal from the noise signal in the frequency domain: The dual discrete cosine transform (DDCT) is composed of two scales of DCT blocks, which perform DCT transforms in different ranges and obtain the corresponding frequency coefficient matrix. The obtained multi-scale parallel data are input into different branches for feature learning, and each branch is composed of four frequency module blocks (FMBs). Each FMB is composed of several four-expanded convolutions with a receptive field of 3×3 and conventional convolutions, as well as a multi-scale fast channel attention, which is used to efficiently extract data features. Then, the results obtained on each branch are fused through another convolutional layer and input into the next DDCT structure; after all modules are processed, a frequency domain subnet model that has learned the frequency domain feature information is obtained, and the output result of the frequency domain subnet is obtained.

4. The method for denoising mixed-domain diffusion magnetic resonance images according to claim 1, characterized in that: Construct a spatial domain denoising subnet module to remove noise from the image in the spatial domain: The spatial domain subnet module adopts a convolutional neural network structure, which consists of 12 spatial module blocks SMB; each SMB consists of a 3×3 convolution, a four-expanded convolution and a multi-scale spatial attention; the input of the SMB is subjected to a 3×3 convolution and a four-expanded convolution for feature extraction, and feature weighting is performed through a multi-scale spatial attention, and then connected to the input of the SMB through a residual connection. Finally, the final output result of the SMB is obtained through another 3×3 convolution, batch normalization and activation function; after processing through all SMBs, the final output result of the spatial domain subnet is obtained.

5. The method for denoising mixed-domain diffusion magnetic resonance images according to claim 1, characterized in that: Construct a frequency domain-spatial domain fusion subnet module to fuse the results of the frequency domain subnet and the spatial domain subnet: The frequency-spatial fusion subnet module adopts a convolutional neural network, which consists of two fusion blocks FB. Each FB is composed of a four-dilated convolution and several conventional convolutions in series, and integrates residual connections. The input of each FB will first be subjected to multi-scale feature extraction through four-dilated convolution to enhance the ability to capture features. Then, the input is passed through three conventional convolutions in sequence for further feature extraction and fusion, and the result is connected to the input of FB through a residual connection; finally, the final output result of FB is obtained through 3×3 convolution, batch normalization and activation function.

6. A mixed domain diffusion magnetic resonance image denoising method according to claim 3 or 4, characterized in that: The multi-scale fast channel attention module and the multi-scale spatial attention module are integrated into the frequency domain denoising subnet module and the spatial domain denoising subnet module respectively: Multi-scale fast-path attention will be integrated into FMB, and multi-scale spatial attention will be integrated into SMB; The multi-scale fast channel attention module consists of three parts: a multi-scale feature extraction part, a channel attention calculation part, and a channel attention weighting part; The multi-scale feature extraction part is mainly composed of multiple parallel convolution kernels with different sizes, which are used to extract feature information of different scales; the channel attention calculation part needs to reduce the dimensionality of the two-dimensional feature maps of different scales into one-dimensional feature vectors through adaptive average pooling and learn the interaction between feature vectors through 1D convolution, and then concatenate the feature vectors of each scale and perform nonlinear mapping through the Sigmoid activation function to obtain the calculated channel attention; the channel attention weighting part concatenates the feature information of each scale and multiplies the channel attention between channels to obtain the feature cube after the channel feature is weighted; four branches are designed, each of which is composed of two 3×3 convolutions with expansion rates of d=1, 2, 3, and 4 respectively; the input of the multi-scale spatial attention will be input into each branch respectively, and feature extraction and learning will be performed respectively, and finally the outputs of these branches will be concatenated, and feature fusion will be performed through another convolution to obtain a two-dimensional feature distribution map; the obtained two-dimensional feature distribution map is nonlinearly mapped through the Sigmoid activation function to obtain the final weight distribution map; the weight distribution map is multiplied with the input to obtain the spatially weighted feature cube.

7. The method for denoising mixed-domain diffusion magnetic resonance images according to claim 1, characterized in that: Constructing the overall dMRI denoising model: The frequency domain denoising subnet module and the space domain denoising subnet module are connected in parallel to construct the denoising main network. The input of the model will be sent to the frequency domain subnet module and the space domain subnet module for feature extraction and learning in the frequency domain and space domain respectively. Then, the outputs of the frequency domain subnet module and the spatial domain subnet module are concatenated with the input to realize residual connection, and then the concatenated results are input into the feature fusion subnet for feature fusion. Finally, before outputting the final result, residual connection is introduced again with the input to obtain the final output result. The clean "ground truth" is obtained by extracting, clustering and geometrically averaging the HCP dMRI dataset, and Rice noise within a certain range of levels is added as noise data to obtain a dataset containing clean-noise image pairs as the training set of the model. The model takes 32 two-dimensional noise images as a batch as input data each time, and outputs denoised high signal-to-noise ratio two-dimensional images of the corresponding batch number. The output of the model and the "ground truth" data are respectively compared through the multi-scale structural similarity loss function to calculate the loss. During the training process, the parameters in the model are adjusted to minimize the loss, thereby optimizing the model and improving the denoising performance of the model.

8. A system for denoising a diffusion magnetic resonance image, characterized in that: include: A data acquisition module is used to acquire diffusion magnetic resonance imaging (dMRI) data and obtain dMRI images required for model training after preprocessing; A denoising module is used to denoise the dMRI images required for model training obtained after preprocessing through a pre-built frequency domain denoising subnet module, a space domain denoising subnet module and a frequency domain-space domain fusion subnet module; A model building module is used to connect the frequency domain denoising subnet module and the space domain denoising subnet module in parallel, and then connect the combined output to the frequency domain-space domain fusion subnet module in series to build a dMRI denoising model; The output module is used to construct a multi-scale structural similarity loss function, train the constructed dMRI denoising model, and recover a high signal-to-noise ratio output image from a low signal-to-noise ratio noisy image through the dMRI denoising model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for mixed-domain diffusion magnetic resonance image denoising as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for mixed-domain diffusion magnetic resonance image denoising as claimed in any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Multi-scale underground nuclear magnetic resonance detection signal denoising method and system

    CN120428343A