Magnetic resonance diffusion microstructure imaging method based on tensor learning
Through tensor learning-based methods, deep neural networks and adaptive parameter learning algorithms are used to solve the problems of long scanning time and poor image quality in magnetic resonance diffusion microstructure imaging, and fast, efficient and high-fidelity microstructure imaging is achieved, improving imaging accuracy and accuracy.
Patent Information
- Application Number
- CN202510540557.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2025-09-02
AI Technical Summary
The existing magnetic resonance diffusion microstructure imaging technology has problems with long scanning time, poor image quality and multi-solvency. The imaging accuracy decreases during sparse sampling, making it difficult to achieve fast, efficient and high-fidelity microstructure imaging.
Using a tensor learning method, a deep neural network model is constructed by obtaining densely sampled data, tensor decomposition is used to extract higher-order correlations between parameters, and adaptive parameter learning algorithm is designed to optimize regularized parameters to realize microstructure parameter imaging of multi-diffusion model.
It shortens the scanning time, improves the efficiency and accuracy of microstructure parameter imaging, achieves fast, efficient and high-fidelity microstructure imaging, reduces the computational burden, and improves imaging quality and accuracy.
Smart Images

Figure CN120580308A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of magnetic resonance imaging, and in particular to a magnetic resonance diffusion microstructure imaging method based on tensor learning. Background Art
[0002] Diffusion Magnetic Resonance Imaging (dMRI) is a technology that captures the diffusion characteristics of water molecules in biological tissues. By applying a specific diffusion-sensitive factor and diffusion gradient direction, it effectively encodes the diffusion information of water molecules in biological tissues along that direction into the contrast of the diffusion-weighted imaging (DWI), thereby reflecting the microstructural characteristics of brain tissue.
[0003] Microstructural imaging aims to establish a precise quantitative relationship between diffusion-weighted imaging (DWI) diffusion signals and tissue microstructure through systematic modeling and model fitting. Commonly used microstructural models include Diffusion Tensor Imaging (DTI), Diffusion Kurtosis Imaging (DKI), and Neurite Orientation Dispersion and Density Imaging (NODDI).
[0004] Microstructure imaging is essentially an inverse problem of inferring cause from effect for the acquired diffusion magnetic resonance signals. The accuracy of the solution is affected by several key factors: first, the number of diffusion gradient directions. Accurate microstructure imaging often requires a large amount of diffusion direction data for fitting, which results in excessively long acquisition times. Second, the image quality itself is low. Increased acquisition time is not only unfriendly to patients but also increases the proportion of motion artifacts. In addition, the inherent signal attenuation characteristics of diffusion magnetic resonance imaging result in a low image signal-to-noise ratio, limiting the accuracy of microstructure imaging. To address the problems of long scanning times and poor image quality, sparse sampling methods have emerged. The core idea of sparse sampling is to reduce the number of scanning directions by selecting only some key diffusion gradient directions for data acquisition, thereby significantly shortening the scanning time.
[0005] However, microstructural models are inherently pathological: microscopically specific parameters of different biophysical structures may exhibit the same macroscopic signal characteristics. This parameter coupling effect significantly reduces the specificity of the inversion results. This leads to multiple possible outcomes when inferring microstructural parameters from the signal, a phenomenon known as "multi-solution." When sparse sampling is used, this "multi-solution" problem becomes more prominent due to the significant reduction in the amount of collected data, significantly reducing imaging accuracy. Summary of the Invention
[0006] To address these shortcomings, the present invention proposes a tensor learning-based magnetic resonance diffusion microstructure imaging method. Starting from sparse sampling, it simultaneously images multiple microstructure parameters from multiple diffusion models. Using tensor decomposition to explicitly exploit the correlations between these multiple parameters, this method not only shortens acquisition time but also improves the efficiency and accuracy of microstructure parameter imaging, achieving fast, efficient, and high-fidelity microstructure imaging. Furthermore, the present invention designs an adaptive parameter learning algorithm based on gradient acceleration, which enables dynamic optimization of the regularization parameter, improves parameter tuning efficiency, and further enhances performance.
[0007] The technical solution adopted by the present invention is a method for magnetic resonance diffusion microstructure imaging based on tensor learning, comprising the following steps:
[0008] S100, acquiring dense sampling data of diffusion magnetic resonance imaging;
[0009] S200, fitting the densely sampled data through a diffusion model to generate a plurality of real microstructure parameters;
[0010] S300, generating sparse sampling data based on the dense sampling data;
[0011] S400, constructing a deep neural network model, inputting the sparse sampling data, and outputting a plurality of predicted microstructure parameters;
[0012] S500, performing tensor decomposition on the real microstructure parameters and the predicted microstructure parameters respectively, extracting high-order correlations between the parameters, and obtaining a real singular value tensor and a predicted singular value tensor containing the high-order correlations;
[0013] S600, constructing a joint training function Loss using the real singular value tensor containing high-order correlation and the predicted singular value tensor to train the deep neural network model;
[0014] S700: Input the real sparse sampling data into the trained deep neural network model.
[0015] Preferably, a plurality of diffusion models of S200 are provided.
[0016] Preferably, the S500 specifically includes:
[0017] S510, organizing the multiple real microstructure parameters obtained in S200 into a real multi-parameter tensor, and organizing the multiple predicted microstructure parameters obtained in S400 into a predicted multi-parameter tensor;
[0018] S520, performing Fourier transform on the real multi-parameter tensor and the predicted multi-parameter tensor to frequency domain respectively, and outputting a real frequency domain tensor and a predicted frequency domain tensor;
[0019] S530: Perform tensor decomposition on the real frequency domain tensor and the predicted frequency domain tensor in the frequency domain, extract spatial neighborhood information, and output the decomposed real frequency domain tensor and predicted frequency domain tensor;
[0020] S540. Restore the decomposed real frequency domain tensor and predicted frequency domain tensor to the original spatial domain through inverse Fourier transform to obtain a real singular value tensor and a predicted singular value tensor containing high-order correlations.
[0021] Preferably, the joint training function Loss of S600 includes a data fidelity term L Data And the regular term Reg, the formula of the regular term Reg is:
[0022]
[0023] Among them, λ is the regularization parameter, S GTi is the i-th eigencomponent of the true singular value tensor, is the i-th eigencomponent of the predicted singular value tensor.
[0024] Preferably, the method for obtaining the regularization parameter comprises the following steps:
[0025] Input the initial value of the regularization parameter λ0, the parameter learning rate α, the momentum coefficient β and the initial value of momentum m0 to the adaptive optimization algorithm;
[0026] By calculating the gradient of the loss function with respect to the regularization parameter, the network parameters and the regularization term weights are jointly updated to achieve automatic adjustment of the regularization parameter.
[0027] Preferably, the automatic adjustment of the regularization parameter is achieved by calculating the gradient of the loss function with respect to the regularization parameter and jointly updating the network parameters and the regularization term weight, specifically including:
[0028] S610, update iterative step t←t+1, where t is the time;
[0029] S620, by minimizing the loss function L(θ,λ t )Optimize network parameters θ t+1 ;
[0030] S630, calculate the momentum update term:
[0031] m t+1 =β·m t +R(θ t+1 )+β·[R(θ t+1)-R(θ t )]
[0032] Where m is the momentum, R(θ) is the regularization term related to the network parameter θ;
[0033] S640, update regularization parameters:
[0034] λ t+1 =λ t -α·m t+1
[0035] S650, loop through S610-S640 until convergence, and output the final regularization parameter λ t+1 .
[0036] Preferably, the momentum coefficient β is between 0.8 and 1, and the initial momentum value m0 is set to 0.
[0037] Preferably, the data fidelity item L Data The formula is:
[0038]
[0039] Among them, I GT are multiple real microstructure parameters of S200, I Pred are multiple predicted microstructure parameters of the S400.
[0040] Compared with the prior art, the present invention has the following beneficial effects:
[0041] 1. The present invention proposes a magnetic resonance diffusion microstructure imaging method based on tensor learning. A regularization term based on a high-order tensor is designed, which retains the high-dimensional information of the multi-parameter tensor and uses tensor decomposition to mine the parameter correlations from different diffusion models. This allows a single network to simultaneously image multiple microstructure parameters from multiple diffusion models, assisting in rapid and accurate microstructure imaging.
[0042] 2. The present invention also designs an adaptive learning algorithm for dynamically adjusting regularization hyperparameters to further improve the performance of microstructure parameter estimation. Furthermore, the proposed parameter adaptive learning algorithm only requires the calculation and storage of a single gradient value, resulting in a low computational burden. This invention accelerates diffusion MRI while improving the learning efficiency and imaging accuracy of multiple microstructure parameters. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] The present invention is described in detail below with reference to the embodiments and accompanying drawings, in which:
[0044] Figure 1 This is a flow chart of a tensor learning-based magnetic resonance diffusion microstructure imaging method;
[0045] Figure 2 This is a block diagram of a magnetic resonance diffusion microstructure imaging method based on tensor learning;
[0046] Figure 3 It is a step diagram of the algorithm for parameter adaptive learning;
[0047] Figure 4 This is a comparison chart of the results of the microstructure imaging method of the present invention and the existing microstructure imaging method. DETAILED DESCRIPTION
[0048] To make the objectives, technical solutions, and advantages of the present invention more apparent, embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar components or components having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.
[0049] Existing technologies do not fully utilize the high-dimensional characteristics of multiple microstructure parameters and the high correlation between parameters, resulting in limited efficiency and accuracy of multi-parameter imaging tasks.
[0050] In one embodiment, a tensor learning-based magnetic resonance diffusion microstructure imaging method is described. Figure 1 , including the following steps:
[0051] S100. Acquire dense sampling data of diffusion magnetic resonance imaging. Collect dense sampling data obtained by diffusion magnetic resonance imaging technology. This data contains rich tissue microstructural information and is the basis for subsequent analysis.
[0052] S200: Fit the densely sampled data using a diffusion model to generate multiple true microstructure parameters. The densely sampled data is fitted using a diffusion model (such as the neurite direction dispersion and density imaging NODDI model, the diffusion kurtosis imaging DKI model, etc.) to obtain multiple parameters reflecting the true microstructure of the tissue, such as neurite direction dispersion, intracellular volume fraction, average kurtosis, etc. These parameters serve as the "gold standard" for subsequent model training.
[0053] S300: Generate sparse sampling data based on the dense sampling data. Based on the dense sampling data, uniform undersampling is performed to simulate the sparse sampling situation commonly seen in clinical practice, generating sparse sampling data for training the prediction ability of a deep neural network under incomplete data.
[0054] S400: Construct a deep neural network model, input the sparsely sampled data, and output multiple predicted microstructure parameters. A deep neural network model is designed to input the sparsely sampled data into the model, learn the data features through the network's internal layer structure (such as convolutional layers, recurrent layers, etc.), and output predicted values for the microstructure parameters.
[0055] S500, respectively perform tensor decomposition on the real microstructure parameters and the predicted microstructure parameters, extract the high-order correlations between the parameters, and obtain a real singular value tensor and a predicted singular value tensor containing the high-order correlations. Perform tensor decomposition (such as singular value decomposition) on the real microstructure parameters and the predicted microstructure parameters respectively. Tensor decomposition can decompose high-dimensional parameter data into lower-dimensional components that are easier to analyze, extract high-order correlation features between the parameters, and obtain a real singular value tensor and a predicted singular value tensor containing these features. For example, the correlation patterns and shared features of different microstructure parameters in the spatial neighborhood can be captured.
[0056] S600: Utilizing the true singular value tensor containing high-order correlations and the predicted singular value tensor to construct a joint training function, Loss, and train the deep neural network model. By minimizing the training function Loss, the parameters of the deep neural network are adjusted so that the predicted microstructure parameters output by the model are not only numerically close to the true values but also match the true microstructure in terms of high-order correlation characteristics, thereby improving the accuracy and generalization ability of the model.
[0057] S700: Inputting real sparse sampling data into the trained deep neural network model. By inputting the actual collected real sparse sampling data into the trained deep neural network model and processing the real data using the trained model, multiple microstructure parameters under a high-fidelity multi-diffusion model can be quickly output, achieving efficient and accurate reconstruction of brain tissue microstructure.
[0058] Compared with existing technologies, this embodiment fully utilizes the high-dimensional nature of diffusion MRI signals, preserves the spatial structure of multiple microstructural parameters, and utilizes tensor decomposition to decouple the correlation of parameters in different dimensions. This allows for more accurate imaging of the microstructural parameters of multiple diffusion models from sparse signals. In actual imaging scenarios, highly accurate restoration and analysis of brain microstructure can be achieved based on less sampled data. This not only reduces sampling time and costs but also improves imaging quality, providing more efficient and accurate microstructural information for clinical diagnosis and scientific research. It demonstrates outstanding practicality and superiority in the application of diffusion MRI microstructural imaging.
[0059] Existing technologies typically only consider a single-model learning paradigm, learning the microstructural parameters of a single diffusion model at a time. Utilizing multiple microstructural parameters from multiple diffusion models can help provide a more comprehensive understanding of brain microstructure. For example, DTI parameters can capture the Gaussian properties of water molecule diffusion, DKI parameters can characterize non-Gaussian properties, and NODDI parameters can provide diffusion information specific to different cell tissues. However, current research has primarily focused on using separate signal models to estimate the metrics of each model separately. Learning the microstructure of multiple diffusion models requires multiple training sessions, resulting in low learning efficiency and a heavy computational burden. Furthermore, the "multi-solution" problem becomes more prominent, significantly reducing imaging accuracy.
[0060] In one embodiment, the diffusion model of S200 is configured to be multiple. For densely sampled signals, on the one hand, the signal is fitted by different diffusion models to obtain multiple microstructure parameters from the multiple diffusion models. These multiple parameters are used as the gold standard for network training and supervised model learning.
[0061] This embodiment utilizes multiple model parameters to jointly supervise network learning, alleviating the problem of "multi-solutions." Different model parameters constrain and complement each other, reducing the uncertainty of single model estimates, significantly improving imaging accuracy and avoiding the significant drop in accuracy caused by "multi-solutions." The integration of multiple diffusion model parameters can more comprehensively and multidimensionally reflect the characteristics of brain microstructures, enhancing the ability to characterize and analyze complex microstructures and contributing to a more precise understanding of brain tissue characteristics and functions.
[0062] In one embodiment, the S500 specifically includes:
[0063] S510 , organizing the multiple real microstructure parameters obtained in S200 into a real multi-parameter tensor, and organizing the multiple predicted microstructure parameters obtained in S400 into a predicted multi-parameter tensor.
[0064] S520: Perform Fourier transform on the real multi-parameter tensor and the predicted multi-parameter tensor to the frequency domain, and output a real frequency domain tensor and a predicted frequency domain tensor. This operation can decouple the dependency between the spatial structure and the parameter dimension.
[0065] S530 . In the frequency domain, perform tensor decomposition on the real frequency domain tensor and the predicted frequency domain tensor respectively, extract spatial neighborhood information, and output the decomposed real frequency domain tensor and predicted frequency domain tensor.
[0066] S540. Restore the decomposed real frequency domain tensor and predicted frequency domain tensor to the original spatial domain through inverse Fourier transform to obtain a real singular value tensor and a predicted singular value tensor containing high-order correlations.
[0067] Through tensor decomposition operations, various types of spatial and structural information in its high-dimensional structure are explicitly extracted, and alignment operations are performed at this information level, thereby guiding the network to more accurately learn the fine-grained structure in the target parameters, significantly improving the accuracy and structural fidelity of the prediction.
[0068] In one embodiment, the joint training function Loss of S600 includes a data fidelity term L Data The regularization term Reg uses the high-order correlations (singular value eigencomponents) between parameters extracted by tensor decomposition to constrain the model to learn the intrinsic structural relationship between the real parameters, thus preventing the model from focusing only on the surface features of the data and ignoring the deep correlation of the microstructure.
[0069] Specifically, the formula of the regularization term Reg is:
[0070]
[0071] Among them, λ is the regularization parameter, S GTi is the i-th eigencomponent of the true singular value tensor, is the i-th eigencomponent of the predicted singular value tensor.
[0072] Furthermore, the data fidelity term L Data By measuring the numerical difference between the predicted parameters and the actual parameters (after normalization), we ensure that the model output is close to the actual value in terms of value, and constrain the model's ability to fit the data itself. Data The formula can be:
[0073]
[0074] Among them, I GT are multiple real microstructure parameters of S200, I Pred are multiple predicted microstructure parameters of the S400.
[0075] Data fidelity item L Data The synergistic effect of the regularization term Reg not only enhances the model's ability to fit the data, but also improves the representation of the deep features of the microstructure, making the model more robust and generalizable in different scenarios, and can more accurately restore the microstructure of tissues, providing reliable support for applications such as medical diagnosis and scientific research analysis, balancing numerical accuracy and structural correlation, and optimizing the training effect and practical application value of deep neural network models in magnetic resonance diffusion microstructure imaging.
[0076] See also Figure 2The entire process can be divided into two branches. The upper branch represents the true microstructural parameters obtained from dense sampling, while the lower branch represents the output of the network prediction based on sparse sampling. The network input is sparsely sampled data obtained by uniformly undersampling from dense measurements. After the network prediction, tensor singular value decomposition is applied to the predicted values and the true data to further align them at the tensor singular values, thereby capturing the shared correlations between multiple microstructural parameters.
[0077] In the figure, Diffusivity Ellipsoid is a geometric representation of the diffusion characteristics of water molecules, used in related fields such as diffusion magnetic resonance imaging to describe differences in diffusion in different directions. Kurtosis Distribution is a kurtosis distribution, which describes the degree to which the data distribution deviates from a normal distribution. In this scenario, it can reflect relevant information such as the non-Gaussian characteristics of the diffusion signal. e1, e2, and e3 generally represent the three principal axis direction vectors of the ellipsoid, which are used to define the orientation and shape of the diffusivity ellipsoid in space. Stage1 and Stage2 represent the two main stages of the neural network, which are used to process input data step by step. SD-LSTM can be a variant of a specific long short-term memory (LSTM) unit. Layer1, Layer2, and Layer3 are three independent layers of Stage2, such as fully connected layers, convolutional layers, or other types of layers.
[0078] In view of the fact that the selection process of the regularization parameter λ in existing methods usually relies on experience or manual parameter adjustment, which has problems such as low efficiency and poor generalization, the present invention also designs an adaptive learning algorithm for dynamically adjusting the regularization hyperparameter to further improve the performance of microstructure parameter estimation. In one embodiment, the method for obtaining the regularization parameter is as follows: Figure 3 , including the following steps:
[0079] The adaptive optimization algorithm is fed with the initial value of the regularization parameter λ0, the parameter learning rate α, the momentum coefficient β, and the initial momentum value m0. Specifically, the momentum coefficient β is set between 0.8 and 1, allowing the algorithm to incorporate both historical and current gradient information when updating parameters. The initial momentum value m0 is set to 0, and as the algorithm iterates, the momentum is gradually updated based on the current gradient and β value. This approach is logically clear and easy to implement, ensuring consistency and standardization of the algorithm in the initial stages and avoiding the uncertainty caused by improper initial momentum settings.
[0080] By calculating the gradient of the loss function with respect to the regularization parameter, the network parameters and the regularization term weights are jointly updated to achieve automatic adjustment of the regularization parameter.
[0081] The specific steps include:
[0082] S610, update iterative step t←t+1, where t is the time;
[0083] S620, by minimizing the loss function L(θ,λ t )Optimize network parameters θ t+1 ;
[0084] S630, calculate the momentum update term:
[0085] m t+1 =β·m t +R(θ t+1 )+β·[R(θ t+1 )-R(θ t )]
[0086] Where m is the momentum, R(θ) is the regularization term related to the network parameter θ;
[0087] S640, update regularization parameters:
[0088] λ t+1 =λ t -α·m t+1
[0089] S650, loop through S610-S640 until convergence, and output the final regularization parameter λ t+1 .
[0090] The algorithm optimizes the network parameters θ and regularization parameter λ alternately on the training set and the validation set. Let λ t and θ t are the values at time t respectively. More specifically, the iterative process is as follows:
[0091]
[0092] Similar to the update of network parameters θ, λ should also move towards the loss function L(θ t+1 ,λ) updates the gradient direction of λ. One of the most direct ideas is to manually calculate the partial derivative of the function with respect to λ:
[0093]
[0094] Due to L Data (θ t+1 ) and R(θ t+1 ) are independent of λ and only related to θ, so it can be concluded that the loss function L(θ t+1 ,λ) with respect to the gradient of λ at time t is only related to the latest regularization term value R(θ t+1 ). Therefore, we introduce momentum m here:
[0095] m t+1 =β·m t +R(θ t+1)+β·[R(θ t+1 )-R(θ t )]
[0096] Where [R(θ t+1 )-R(θ t )] is the loss function L(θ t+1 ,λ) about the gradient of λ, that is, L(θ t+1 ,λ) is the approximate second-order derivative of λ. This foresight enables better utilization of curvature information and greater robustness. With a deep understanding of this, we introduce our algorithm here:
[0097] λ t+1 =λ t -α·m t+1 =λ t -αβ·m t -α·R(θ t+1 )-αβ·[R(θ t+1 )-R(θ t )]
[0098] By calculating the gradient of the loss function with respect to the regularization parameter λ and jointly updating the network parameters and the regularization term weights, the regularization parameters can be automatically adjusted, which improves the efficiency and stability of the training process and reduces the burden of manual parameter adjustment.
[0099] Through the above technical process, after the model training is completed, only real sparse sampling data needs to be input to quickly output multiple microstructure parameters under the high-fidelity multi-diffusion model, thereby achieving efficient and accurate reconstruction of the brain tissue microstructure.
[0100] In a more specific embodiment, the present invention uses preprocessed whole-brain diffusion magnetic resonance imaging data from a young adult dataset. 111 subjects were randomly selected in the experiment, 60 of whom were used for model training, 17 for validation, and the remaining 34 as a test set. In order to simulate clinical sparse sampling conditions, 6 encoded diffusion gradient directions were uniformly selected on each b-value layer, totaling 18 directions, to form sparse under-sampled data as the input of the network model; at the same time, all 270 diffusion gradient directions provided by the dataset were used for model fitting, and the resulting high-quality multi-microstructure parameter images were used as the gold standard for network training and evaluation. In the present invention, the diffusion kurtosis imaging model (DKI) and neurite orientation dispersion and density imaging (NODDI) were selected as representative examples to verify the versatility and effectiveness of the proposed method. Specifically, for the DKI model, the following four key parameters were fitted: Kurtosis Fractional Anisotropy (KFA), Mean Kurtosis (MK), Axial Kurtosis (AK) and Radial Kurtosis (RK); for the NODDI model, three parameters were estimated: Orientation Dispersion (OD), Volume Fraction of Intra-Cellular (Vic) and Volume Fraction of Isotropic (Viso).
[0101] To further verify the superiority of the method of the present invention, comparative experiments were carried out with the most advanced existing deep learning-based microstructure imaging methods, including q-DL (q-space deep learning method), U-Net (convolutional neural network), MESC-SD (tissue microstructure estimation deep network) and HGT (hybrid graph transformer). Figure 4 Comparison results between the proposed method and existing mainstream deep learning microstructure imaging methods under 15x accelerated sampling conditions. Seven microstructure parameters, including KFA, MK, AK, RK, OD, Vic, and Viso, are presented from two diffusion models, DKI and NODDI. The first row for each parameter shows the microstructure imaging result; the second row shows the error plot compared to the reference standard image.
[0102] Under sparse sampling conditions (approximately 15x accelerated sampling), the proposed method (the last column in the figure, labeled "Ours") outperforms the comparative methods in terms of reconstruction performance of various microstructural parameters. The reconstruction results are closest to the gold standard images obtained by fitting with the full sampling data in terms of both quantitative indicators and visual effects, demonstrating that excellent structural fidelity and parameter accuracy are maintained under high acceleration conditions.
[0103] In this specification, the use of terms such as "Embodiment 1," "this embodiment," and "in one embodiment" indicates that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in the invention or at least one embodiment or example of the invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example; furthermore, the specific features, structures, materials, or characteristics described may be appropriately combined in any one or more embodiments or examples.
[0104] In the description of this specification, the terms "connect," "install," "fix," "dispose," and "have" are to be understood in a broad sense. For example, "connect" can mean a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on specific circumstances.
[0105] In the description of this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.
[0106] The above description of the embodiments is to facilitate ordinary technicians in this technical field to understand and apply the technology of this case. People familiar with the technology in this field can obviously make various modifications to these examples easily and apply the general principles described here to other embodiments without having to go through creative work. Therefore, this case is not limited to the above embodiments. Modifications to the following situations should all be within the scope of protection of this case: ① A new technical solution implemented based on the technical solution of the present invention and combined with existing common knowledge, the technical effect produced by the new technical solution does not exceed the technical effect of the present invention; ② The equivalent replacement of some features of the technical solution of the present invention with the known technology, the technical effect produced is the same as the technical effect of the present invention; ③ The technical solution of the present invention is expandable, and the substantive content of the expanded technical solution does not exceed the technical solution of the present invention; ④ The equivalent transformation made by the content of the description and drawings of the present invention is directly or indirectly applied to other related technical fields.
Claims
1. A magnetic resonance diffusion microstructure imaging method based on tensor learning, characterized in that: The following steps are involved: S100, acquiring dense sampling data of diffusion magnetic resonance imaging; S200, fitting the densely sampled data through a diffusion model to generate a plurality of real microstructure parameters; S300, generating sparse sampling data based on the dense sampling data; S400, constructing a deep neural network model, inputting the sparse sampling data, and outputting a plurality of predicted microstructure parameters; S500, performing tensor decomposition on the real microstructure parameters and the predicted microstructure parameters respectively, extracting high-order correlations between the parameters, and obtaining a real singular value tensor and a predicted singular value tensor containing the high-order correlations; S600, constructing a joint training function Loss using the real singular value tensor containing high-order correlation and the predicted singular value tensor to train the deep neural network model; S700: Input the real sparse sampling data into the trained deep neural network model.
2. The imaging method according to claim 1, wherein There are multiple diffusion models in S200.
3. The imaging method according to claim 1 or 2, characterized in that The S500 specifically includes: S510, organizing the multiple real microstructure parameters obtained in S200 into a real multi-parameter tensor, and organizing the multiple predicted microstructure parameters obtained in S400 into a predicted multi-parameter tensor; S520, performing Fourier transform on the real multi-parameter tensor and the predicted multi-parameter tensor to frequency domain respectively, and outputting a real frequency domain tensor and a predicted frequency domain tensor; S530: Perform tensor decomposition on the real frequency domain tensor and the predicted frequency domain tensor in the frequency domain, extract spatial neighborhood information, and output the decomposed real frequency domain tensor and predicted frequency domain tensor; S540. Restore the decomposed real frequency domain tensor and predicted frequency domain tensor to the original spatial domain through inverse Fourier transform to obtain a real singular value tensor and a predicted singular value tensor containing high-order correlations.
4. The imaging method according to claim 3, wherein: The joint training function Loss of S600 includes the data fidelity term L Data And the regular term Reg, the formula of the regular term Reg is: Among them, λ is the regularization parameter, S GTi is the i-th eigencomponent of the true singular value tensor, is the i-th eigencomponent of the predicted singular value tensor.
5. The imaging method according to claim 4, characterized in that The method for obtaining the regularization parameter comprises the following steps: Input the initial value of the regularization parameter λ0, the parameter learning rate α, the momentum coefficient β and the initial value of momentum m0 to the adaptive optimization algorithm; By calculating the gradient of the loss function with respect to the regularization parameter, the network parameters and the regularization term weights are jointly updated to achieve automatic adjustment of the regularization parameter.
6. The imaging method according to claim 5, characterized in that The automatic adjustment of the regularization parameter is achieved by calculating the gradient of the loss function with respect to the regularization parameter and jointly updating the network parameters and the regularization term weight. Specifically, the following steps are performed: S610, update iterative step t←t+1, where t is the time; S620, by minimizing the loss function L(θ,λ t )Optimize network parameters θ t+1 ; S630, calculate the momentum update term: m t+1 =β·m t +R(θ t+1 )+β·[R(θ t+1 )―R(θ t )] Where m is the momentum, R(θ) is the regularization term related to the network parameter θ; S640, update regularization parameters: l t+1 =λ t ―a·m t+1 S650, loop through S610-S640 until convergence, and output the final regularization parameter λ t+1 .
7. The imaging method according to claim 5 or 6, characterized in that: The momentum coefficient β is between 0.8 and 1, and the initial momentum value m0 is set to 0.
8. The imaging method according to claim 4, wherein: The data fidelity term L Data The formula is: Among them, I GT are multiple real microstructure parameters of S200, I Pred are multiple predicted microstructure parameters of the S400.
Citation Information
Cited By
Craniocerebral multi-frequency conductivity tensor modeling method and system, medium and electronic equipment
CN121459911A
Skull-brain multi-frequency conductivity tensor modeling method, system, medium and electronic device
CN121459911B
