Diffusion magnetic resonance imaging microstructure parameter estimation method and device and electronic equipment

Abstract

The invention relates to the technical field of medical image processing, in particular to a diffusion magnetic resonance imaging microstructure parameter estimation method and device and electronic equipment, and the method comprises the steps: obtaining diffusion magnetic resonance imaging data of an imaging target; diffusion magnetic resonance imaging data are input into a microstructure model, the microstructure model outputs diffusion indexes of an imaging target, the microstructure model comprises a plurality of stacked hybrid neural network layers, and the stacked hybrid neural network layers are connected in a jumper connection mode. The mixed neural network layer carries out linear compression, nonlinear spline activation and linear expansion on diffusion magnetic resonance imaging data; and estimating microstructure parameters of the imaging target according to the diffusion index. Therefore, the problems of low calculation efficiency, high calculation complexity, low calculation precision and the like in related technologies are solved.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a method, device and electronic equipment for estimating microstructure parameters of diffusion magnetic resonance imaging. Background Art

[0002] Diffusion magnetic resonance imaging (MRI) is a key technology for noninvasively mapping the microstructure of the human brain. By measuring and simulating the diffusion patterns of water molecules in brain tissue, it can characterize brain tissue microstructure in vivo. Early biophysical models, such as the diffusion tensor model and the Neurite Directed Diffusion and Density Imaging (NODDI) model, quantified relevant parameters and provided key biomarkers for research on cognitive function, aging, and mental health.

[0003] Traditional diffusion MRI modeling relies on iterative nonlinear optimization methods to process large numbers of voxels, requiring computational time of up to several hours and significantly impacted by model complexity, imaging resolution, and hardware capabilities. Despite improvements in acceleration algorithms and GPU parallelization, computational efficiency remains a bottleneck in large-scale imaging studies and real-time clinical processing.

[0004] Related technologies use deep learning methods such as multilayer perceptrons and convolutional neural networks to improve computing speed. However, due to limitations in network architecture design, fixed activation functions and linear convolution operations make it difficult to capture the complex nonlinear physical relationships in diffusion signals (such as fiber crossing and axonal microstructure heterogeneity), resulting in a decrease in parameter estimation accuracy in complex scenarios. Summary of the Invention

[0005] The present application provides a method, device and electronic device for estimating microstructure parameters using diffusion magnetic resonance imaging to solve the problems of low computational efficiency, high computational complexity and low computational accuracy in related technologies.

[0006] A first aspect embodiment of the present application provides a method for estimating microstructure parameters of diffusion magnetic resonance imaging, comprising the following steps: acquiring diffusion magnetic resonance imaging data of an imaging target; inputting the diffusion magnetic resonance imaging data into a microstructure model, and the microstructure model outputting a diffusion index of the imaging target, wherein the microstructure model includes multiple stacked hybrid neural network layers, and the multiple stacked hybrid neural network layers are connected by jump connections, and the hybrid neural network layers perform linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data; and estimating the microstructure parameters of the imaging target based on the diffusion index.

[0007] Optionally, the hybrid neural network layer includes a bottleneck Kolmogorov-Arnold convolution layer and a jump connection structure, wherein the bottleneck Kolmogorov-Arnold convolution layer includes a first convolution, linear compression, a second convolution and a linear expansion, the first convolution and the second convolution are different, the second convolution is a Kolmogorov-Arnold convolution, the bottleneck Kolmogorov-Arnold convolution layer performs linear compression on the input data to obtain compressed features, the second convolution performs nonlinear spline activation on the compressed features according to pixels to obtain nonlinear features, and the bottleneck Kolmogorov-Arnold convolution layer performs linear expansion on the nonlinear features.

[0008] Alternatively, the linear compression formula in the bottleneck Kolmogorov-Arnold convolutional layer is:

[0009]

[0010] Among them, Z b,d,h,w is the eigenvalue after linear compression, with channel b and spatial position (d, h, w); d is the depth dimension; h is the height dimension; w is the width dimension; C in is the number of channels of input data; W b,c is the learnable weight of the 1×1×1 convolution during the linear compression process; Y c,d,h,w The input to the bottleneck structure layer is channel c, and the eigenvalue in the three-dimensional space position is (d, h, w);

[0011] The nonlinear spline activation formula for the second convolution in the bottleneck Kolmogorov-Arnold convolution layer is:

[0012]

[0013] in, For Z b,d,h,w Eigenvalue after nonlinear spline activation; α b,k is the learnable coefficient in the nonlinear spline activation function; B k represents the kth spline basis function; φ b is a nonlinear spline activation function;

[0014] The linear expansion formula of the bottleneck Kolmogorov-Arnold convolutional layer is:

[0015]

[0016] in, is the output eigenvalue after linear expansion, with channel c and spatial position (d, h, w); C b is the bottleneck dimension; is the learnable weight of the 1×1×1 convolution during the linear expansion process; For Z b,d,h,w Eigenvalues after nonlinear spline activation.

[0017] Optionally, before inputting the diffusion magnetic resonance imaging data into the microstructure model, the method further includes: acquiring training data, wherein the training data includes a plurality of diffusion nuclear magnetic resonance imaging data; training the microstructure model using the training data, and during the training process, optimizing the network parameters of the microstructure model based on a loss function established based on a forward diffusion model function.

[0018] Optionally, the loss function is:

[0019]

[0020] Among them, E is the mathematical expectation; Y is the true value; is the predicted value; For the model according to the parameters The process of processing input data and obtaining prediction results.

[0021] Optionally, estimating the microstructure parameters of the imaging target according to the diffusion index includes: using the diffusion index to input a forward diffusion model function, and the forward diffusion model function outputting the microstructure parameters of the imaging target.

[0022] Optionally, the forward diffusion model function includes a tensor model and a neurite directional diffusion and density imaging model, wherein,

[0023] The formula of the tensor model is:

[0024] Tensor(S0,D)=S0e -bAD ;

[0025] Where S0 is the magnetic resonance signal intensity without diffusion weighting; D is the diffusion coefficient; b is the diffusion sensitivity factor, both D and b are determined by the acquisition protocol; A is the diffusion tensor transformation vector, which is determined by the diffusion gradient encoding direction; Tensor is the tensor model;

[0026] The formula for the neurite directional diffusion and density imaging model is:

[0027] NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso );

[0028] Where S0 is the magnetic resonance signal intensity without diffusion weighting; f iso is the volume fraction of the isotropic diffusion component; f ic is the volume fraction of the intracellular diffusive component in the non-isotropic part; κ is the dispersion parameter of neurite orientation; μ is the average orientation of neurites; A ic is the signal attenuation function of intracellular diffusion; A ec is the signal attenuation function of extracellular diffusion; A iso is the signal attenuation function of isotropic diffusion; NODDI is the neurite directional diffusion and density imaging model.

[0029] A second aspect of the present application provides a device for estimating microstructure parameters of diffusion magnetic resonance imaging, including: an acquisition module for acquiring diffusion magnetic resonance imaging data of an imaging target; an input module for inputting the diffusion magnetic resonance imaging data into a microstructure model, and the microstructure model outputs a diffusion index of the imaging target, wherein the microstructure model includes multiple stacked hybrid neural network layers, and the multiple stacked hybrid neural network layers are connected by jump connections, and the hybrid neural network layers perform linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data; an estimation module for estimating the microstructure parameters of the imaging target based on the diffusion index.

[0030] Optionally, the hybrid neural network layer includes a bottleneck Kolmogorov-Arnold convolution layer and a jump connection structure, wherein the bottleneck Kolmogorov-Arnold convolution layer includes a first convolution, linear compression, a second convolution and a linear expansion, the first convolution and the second convolution are different, the second convolution is a Kolmogorov-Arnold convolution, the bottleneck Kolmogorov-Arnold convolution layer performs linear compression on the input data to obtain compressed features, the second convolution performs nonlinear spline activation on the compressed features according to pixels to obtain nonlinear features, and the bottleneck Kolmogorov-Arnold convolution layer performs linear expansion on the nonlinear features.

[0031] Alternatively, the linear compression formula in the bottleneck Kolmogorov-Arnold convolutional layer is:

[0032]

[0033] Among them, Z b,d,h,w is the eigenvalue after linear compression, with channel b and spatial position (d, h, w); d is the depth dimension; h is the height dimension; w is the width dimension; C in is the number of channels of input data; W b,c is the learnable weight of the 1×1×1 convolution during the linear compression process; Y c,d,h,wThe input to the bottleneck structure layer is channel c, and the eigenvalue in the three-dimensional space position is (d, h, w);

[0034] The nonlinear spline activation formula for the second convolution in the bottleneck Kolmogorov-Arnold convolution layer is:

[0035]

[0036] in, For Z b,d,h,w Eigenvalue after nonlinear spline activation; α b,k is the learnable coefficient in the nonlinear spline activation function; b k represents the kth spline basis function; φ b is a nonlinear spline activation function;

[0037] The linear expansion formula of the bottleneck Kolmogorov-Arnold convolutional layer is:

[0038]

[0039] in, is the output eigenvalue after linear expansion, with channel c and spatial position (d, h, w); C b is the bottleneck dimension; is the learnable weight of the 1×1×1 convolution during the linear expansion process; For Z b,d,h,w Eigenvalues after nonlinear spline activation.

[0040] Optionally, it also includes: a training module, which is used to obtain training data before inputting the diffusion magnetic resonance imaging data into the microstructure model, wherein the training data includes multiple diffusion nuclear magnetic resonance imaging data; use the training data to train the microstructure model, and during the training process, optimize the network parameters of the microstructure model based on the loss function established by the forward diffusion model function.

[0041] Optionally, the loss function is:

[0042]

[0043] Among them, E is the mathematical expectation; Y is the true value; is the predicted value; For the model according to the parameters The process of processing input data and obtaining prediction results.

[0044] Optionally, the estimation module is further configured to: utilize the diffusion index to input a forward diffusion model function, and the forward diffusion model function outputs microstructure parameters of the imaging target.

[0045] Optionally, the forward diffusion model function includes a tensor model and a neurite directional diffusion and density imaging model, wherein,

[0046] The formula of the tensor model is:

[0047] Tensor(S0,D)=S0e -bAD ;

[0048] Where S0 is the magnetic resonance signal intensity without diffusion weighting; D is the diffusion coefficient; b is the diffusion sensitivity factor, both D and b are determined by the acquisition protocol; A is the diffusion tensor transformation vector, which is determined by the diffusion gradient encoding direction; Tensor is the tensor model;

[0049] The formula for the neurite directional diffusion and density imaging model is:

[0050] NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso );

[0051] Where S0 is the magnetic resonance signal intensity without diffusion weighting; f iso is the volume fraction of the isotropic diffusion component; f ic is the volume fraction of the intracellular diffusive component in the non-isotropic part; κ is the dispersion parameter of neurite orientation; μ is the average orientation of neurites; A ic is the signal attenuation function of intracellular diffusion; A ec is the signal attenuation function of extracellular diffusion; A iso is the signal attenuation function of isotropic diffusion; NODDI is the neurite directional diffusion and density imaging model.

[0052] A third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the diffusion magnetic resonance imaging microstructure parameter estimation method as described in the above embodiment.

[0053] A fourth aspect of the present application provides a computer-readable storage medium having a computer program or product stored thereon, which is executed by a processor to implement the diffusion magnetic resonance imaging microstructure parameter estimation method as described in the above embodiment.

[0054] Therefore, this application has the following beneficial effects:

[0055] In an embodiment of the present application, diffusion magnetic resonance imaging data can be input into a microstructure model, and the microstructure model outputs a diffusion index of the imaging target, and the microstructure parameters of the imaging target are estimated based on the diffusion index. The microstructure model includes multiple stacked hybrid neural network layers, and jump connections are used between the multiple stacked hybrid neural network layers. The hybrid neural network layer performs linear compression, nonlinear spline activation, and linear expansion on the diffusion magnetic resonance imaging data. Through the nonlinear convolution operation of the adaptive spline, the inherent nonlinear relationship in the diffusion magnetic resonance imaging data is effectively utilized, which significantly improves the modeling accuracy, reduces the computational complexity, and improves the computational efficiency. The jump connection method can promote efficient gradient propagation and detailed spatial information retention, thereby improving the prediction accuracy and training stability. Thus, the technical problems of slow computational efficiency, high computational complexity, and low computational accuracy in the related art are solved.

[0056] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0058] Figure 1 This is a flow chart of a method for estimating microstructure parameters by diffusion magnetic resonance imaging according to an embodiment of the present application;

[0059] Figure 2 A schematic diagram of a supervised learning architecture provided according to an embodiment of the present application;

[0060] Figure 3 A schematic diagram of a self-supervised learning architecture provided according to an embodiment of the present application;

[0061] Figure 4 A DTI (Diffusion Tensor Imaging) metric map for supervising Connect-KAN according to an embodiment of the present application;

[0062] Figure 5 A NODDI (Neurite Orientation Dispersion and Density Imaging) metric based on self-supervised learning Connect-KAN provided in an embodiment of the present application;

[0063] Figure 6 A DTI metric map based on self-supervised learning DIMOND-KAN provided according to an embodiment of the present application;

[0064] Figure 7 A NODDI metric mapping diagram based on self-supervised learning DIMOND-KAN provided according to an embodiment of the present application;

[0065] Figure 8 This is an example diagram of a device for estimating microstructure parameters by diffusion magnetic resonance imaging according to an embodiment of the present application;

[0066] Figure 9 Schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0067] The following describes in detail embodiments of the present application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0068] In order to assist in understanding the solution of this application, the relevant technology is first introduced.

[0069] Traditional diffusion MRI modeling relies on iterative nonlinear optimization methods to process millions of voxels per subject, and the computational process takes hours (affected by model complexity, imaging resolution, and hardware capabilities). Although existing acceleration algorithms and GPU parallelization technologies have improved, large-scale imaging studies (such as HCP, UKB, ADNI) and clinical real-time processing still face computational efficiency bottlenecks.

[0070] In recent years, DL (Deep Learning) techniques have been introduced to diffusion MRI analysis, demonstrating significant potential. Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs) transform iterative optimization into a signal-parameter mapping through a single forward pass, significantly accelerating parameter estimation. For example, the q-DL framework utilizes MLPs to reduce the number of DWIs (Diffusion Weighted Images) required for reliable parameter estimation by a factor of 12; DeepDTI uses CNNs to recover DTI indices from just six DWIs, leveraging spatial contextual information to improve estimation performance. While these methods reduce data dependency and accelerate processing, they are still based on fixed activation functions and linear convolution operations, and have inherent limitations in simulating the complex nonlinear relationships between diffusion signals and microstructural parameters.

[0071] Specifically, the core flaws of traditional diffusion MRI modeling methods lie in their computational inefficiency and lack of adaptability. They rely on iterative nonlinear optimization, which requires voxel-by-voxel processing. The computational burden increases significantly with imaging resolution and model complexity, making large-scale data analysis (such as 10,000-person cohort studies) prohibitively time-consuming and difficult to integrate into real-time clinical workflows. Despite improvements in optimization algorithms and hardware acceleration techniques, the results are highly dependent on model-specific tuning, limiting the universality and rapid deployment of new technologies.

[0072] Although deep learning methods based on MLP and CNN in related technologies have improved processing speed, they are limited by network architecture design: fixed activation functions (such as ReLU, Sigmoid) and linear convolution operations make it difficult to capture the complex nonlinear physical relationships in diffusion signals (such as fiber crossing and axonal microstructural heterogeneity), resulting in a decrease in parameter estimation accuracy in complex scenarios. For example, when faced with imaging data with multiple fiber orientations or high noise, the linear convolution of traditional CNN cannot adaptively adjust the nonlinear mapping, and the MLP lacks the ability to integrate spatial context information, further exacerbating modeling errors. These limitations make it impossible for existing DL methods to completely replace traditional optimization methods in high-precision microstructural modeling (such as NODDI parameter estimation). A more flexible nonlinear modeling framework is urgently needed to break through the performance bottleneck.

[0073] To this end, the present application provides a method for estimating microstructure parameters of diffusion magnetic resonance imaging, in which diffusion magnetic resonance imaging data can be input into a microstructure model, the microstructure model outputs a diffusion index of the imaging target, and the microstructure parameters of the imaging target are estimated based on the diffusion index, wherein the microstructure model includes multiple stacked hybrid neural network layers, and jump connections are used between the multiple stacked hybrid neural network layers. The hybrid neural network layer performs linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data. Through the nonlinear convolution operation of the adaptive spline, the inherent nonlinear relationship in the diffusion magnetic resonance imaging data is effectively utilized, which significantly improves the modeling accuracy, reduces the computational complexity, and improves the computational efficiency. In addition, the jump connection method can promote efficient gradient propagation and detailed spatial information retention, thereby improving the prediction accuracy and training stability.

[0074] Specifically, Figure 1 A schematic flow chart of a diffusion magnetic resonance imaging microstructure parameter estimation method provided in an embodiment of the present application.

[0075] like Figure 1 As shown, the diffusion magnetic resonance imaging microstructure parameter estimation method includes the following steps:

[0076] In step S101 , diffusion magnetic resonance imaging data of an imaging target is acquired.

[0077] Among them, the imaging target is the object targeted by diffusion magnetic resonance imaging, such as the brain, that is, the specific target object whose microstructure is to be observed and analyzed through diffusion magnetic resonance imaging technology.

[0078] It is understandable that the embodiments of the present application can acquire diffusion magnetic resonance imaging data of an imaging target, so as to be subsequently used to estimate microstructure parameters of the imaging target.

[0079] Among them, diffusion magnetic resonance imaging is an important technology for non-invasively mapping the microstructure of the human brain. By measuring and simulating the diffusion pattern of water molecules in brain tissue, it can characterize the microstructural characteristics of brain tissue in vivo.

[0080] In step S102, the diffusion magnetic resonance imaging data is input into the microstructure model, and the microstructure model outputs the diffusion index of the imaging target, wherein the microstructure model includes multiple stacked hybrid neural network layers, and the multiple stacked hybrid neural network layers are connected by jump connections. The hybrid neural network layers perform linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data.

[0081] Among them, diffusion indices are some quantitative parameters that reflect the diffusion characteristics of water molecules in the imaging target, such as apparent diffusion coefficient (ADC), fractional anisotropy (FA), radial diffusion coefficient (DR) and axial diffusion coefficient (DA).

[0082] It can be understood that the embodiment of the present application can input diffusion magnetic resonance imaging data into the microstructure model, and the microstructure model outputs the diffusion index of the imaging target, wherein the microstructure model includes multiple stacked hybrid neural network layers, and jump connections are used between the multiple stacked hybrid neural network layers. The hybrid neural network layer performs linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data. Through the nonlinear convolution operation of the adaptive spline, the inherent nonlinear relationship in the diffusion magnetic resonance imaging data is effectively utilized, which significantly improves the modeling accuracy and reduces the computational complexity. In addition, the jump connection method can promote efficient gradient propagation and detailed spatial information retention, thereby improving the prediction accuracy and training stability.

[0083] In an embodiment of the present application, the hybrid neural network layer includes a bottleneck Kolmogorov-Arnold convolution layer and a jump connection structure, wherein the bottleneck Kolmogorov-Arnold convolution layer includes a first convolution, linear compression, a second convolution and a linear expansion, the first convolution and the second convolution are different, the second convolution is a Kolmogorov-Arnold convolution, the bottleneck Kolmogorov-Arnold convolution layer performs linear compression on the input data to obtain compressed features, the second convolution performs nonlinear spline activation on the compressed features according to pixels to obtain nonlinear features, and the bottleneck Kolmogorov-Arnold convolution layer performs linear expansion on the nonlinear features.

[0084] Among them, the first convolution can be a normal convolution, and the second convolution can be a Kolmogorov-Arnold convolution. The Kolmogorov-Arnold convolution contains nonlinear spline activation, which is different from the normal convolution.

[0085] It can be understood that the hybrid neural network layer of the embodiment of the present application includes a bottleneck Kolmogorov-Arnold convolution layer and a jump connection structure, wherein the bottleneck Kolmogorov-Arnold convolution layer includes a first convolution, linear compression, a second convolution and a linear expansion, the second convolution is a Kolmogorov-Arnold convolution, the bottleneck Kolmogorov-Arnold convolution layer performs linear compression on the input data to obtain compressed features, the second convolution performs nonlinear spline activation on the compressed features according to pixels to obtain nonlinear features, and the bottleneck Kolmogorov-Arnold convolution layer performs linear expansion on the nonlinear features.

[0086] In the embodiment of the present application, the linear compression formula in the bottleneck Kolmogorov-Arnold convolutional layer is:

[0087]

[0088] Among them, Z b,d,h,w is the eigenvalue after linear compression, with channel b and spatial position (d, h, w); d is the depth dimension; h is the height dimension; w is the width dimension; C in is the number of channels of input data; W b,c is the learnable weight of the 1×1×1 convolution during the linear compression process; Y c,d,h,w The input to the bottleneck structure layer is channel c, and the eigenvalue in the three-dimensional space position is (d, h, w);

[0089] The nonlinear spline activation formula for the second convolution in the bottleneck Kolmogorov-Arnold convolution layer is:

[0090]

[0091] in, For Z b,d,h,w Eigenvalue after nonlinear spline activation; α b,k is the learnable coefficient in the nonlinear spline activation function; B k represents the kth spline basis function; φ b is a nonlinear spline activation function;

[0092] The linear expansion formula of the bottleneck Kolmogorov-Arnold convolutional layer is:

[0093]

[0094] in, is the output eigenvalue after linear expansion, with channel c and spatial position (d, h, w); C b is the bottleneck dimension; is the learnable weight of the 1×1×1 convolution during the linear expansion process; For Z b,d,h,w Eigenvalues after nonlinear spline activation.

[0095] Unlike traditional convolutional neural networks that rely on linear convolution operations and fixed activation functions, the neural network layer of this application implements convolution operations based on the Kolmogorov-Arnold representation theorem. According to this theorem, any continuous multivariate function can be approximated using a finite combination and superposition of single-variable functions. Using this theorem, Conv-KANs replaces traditional linear convolution with an activation function based on adaptive splines, significantly enhancing the representation capability of convolution operations.

[0096] Among them, the 3D Conv-KAN operation form for input volume data is Among them C in Denotes the input channel and represents the spatial dimension, D, H, W represent the spatial dimension:

[0097]

[0098] in, It is for each spatial offset (a, b, e) and channel pair in the convolution kernel Separately learned nonlinear spline activation functions. Each spline activation function φ(z) is explicitly parameterized as a linear combination of spline basis functions B k (z):

[0099]

[0100] α kAs a trainable spline parameter, K represents the number of spline nodes.

[0101] In order to effectively reduce the computational complexity introduced by numerous spline parameters, the embodiment of this application specifically adopts 3DBottleneck-Conv-KAN to extend the bottleneck design to three-dimensional space. Each Bottleneck-Conv-KAN layer consists of three different steps. First, linear compression is performed by 1×1×1 convolution to convert the channel from C in Reduce to the bottleneck dimension C b :

[0102]

[0103] Then, the compressed feature Z b,d,h,w Perform pixel-wise nonlinear spline activation:

[0104]

[0105] Finally, the bottleneck feature is expanded to the required number of output channels C by linear expansion through another 1×1×1 convolution out :

[0106]

[0107] Among them, W b,c ,V c,b are the learnable weights of the linear convolution. By using these Bottleneck-Conv-KANs in a supervised learning architecture, the number of parameters and computational load can be further reduced while still maintaining strong nonlinear approximation capabilities.

[0108] In an embodiment of the present application, before inputting the diffusion magnetic resonance imaging data into the microstructure model, the method further includes: acquiring training data, wherein the training data includes a plurality of diffusion nuclear magnetic resonance imaging data; using the training data to train the microstructure model, and during the training process, optimizing the network parameters of the microstructure model based on the loss function established by the forward diffusion model function.

[0109] Among them, the loss function is:

[0110]

[0111] Among them, E is the mathematical expectation; Y is the true value; is the predicted value; For the model according to the parameters The process of processing input data and obtaining prediction results.

[0112] It can be understood that, in the embodiment of the present application, before the diffusion magnetic resonance imaging data is input into the microstructure model, the microstructure model can be trained to improve the accuracy of the microstructure model output.

[0113] Specifically, for each subject, Connect-KAN maps the input volume Y to the diffusion model parameters and minimizes the difference between the input volume Y and the network output through the forward diffusion model. The resulting composite volume The difference between them is used to optimize the network parameters:

[0114]

[0115] In step S103 , the microstructure parameters of the imaging target are estimated according to the diffusion index.

[0116] In an embodiment of the present application, estimating the microstructure parameters of the imaging target according to the diffusion index includes: using the diffusion index to input a forward diffusion model function, and the forward diffusion model function outputs the microstructure parameters of the imaging target.

[0117] It is understandable that, in the embodiment of the present application, the diffusion index can be input into the forward diffusion model function, and the forward diffusion model function outputs the microstructure parameters of the imaging target.

[0118] In the embodiment of the present application, the forward diffusion model function includes a tensor model and a neurite directional diffusion and density imaging model, wherein,

[0119] The formula of the tensor model is:

[0120] Tensor(S0,D)=S0e -bAD ;

[0121] Where S0 is the magnetic resonance signal intensity without diffusion weighting; D is the diffusion coefficient; b is the diffusion sensitivity factor, both D and b are determined by the acquisition protocol; A is the diffusion tensor transformation vector, which is determined by the diffusion gradient encoding direction; Tensor is the tensor model;

[0122] The formula for the neurite directional diffusion and density imaging model is:

[0123] NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso );

[0124] Where S0 is the magnetic resonance signal intensity without diffusion weighting; f iso is the volume fraction of the isotropic diffusion component; f ic is the volume fraction of the intracellular diffusive component in the non-isotropic part; κ is the dispersion parameter of neurite orientation; μ is the average orientation of neurites; A ic is the signal attenuation function of intracellular diffusion; A ec is the signal attenuation function of extracellular diffusion; A iso is the signal attenuation function of isotropic diffusion; NODDI is the neurite directional diffusion and density imaging model.

[0125] Specifically, the forward diffusion modeling function Modeling(·) can be derived from any existing or new magnetic resonance imaging diffusion modeling method, including the tensor model and the NODDI model in this application.

[0126] The formula of the tensor model is as follows:

[0127] Tensor(S0,D)=S0e -bAD ;

[0128] where A and b represent the diffusion tensor transformation matrix and b value, respectively, both determined by the acquisition protocol. D and S0 represent the diffusion tensor and non-diffusion weighted signal estimated by DIMOND-KAN.

[0129] The implementation of the NODDI model is as follows:

[0130] NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso );

[0131] Among them, A ic (μ,κ),A ec (μ,f ic ),A iso Represent the diffusion signals of the intracellular compartment, extracellular compartment and isotropic compartment respectively; f ic ,1-f ic ,f lso represent the volume fractions of the intracellular compartment, extracellular compartment and isotropic compartment, respectively.

[0132] The following describes the diffusion magnetic resonance imaging microstructure parameter estimation method of the present application through a specific embodiment.

[0133] This application proposes a deep learning framework, MicroKAN, based on Kolmogorov-Arnold networks and adaptive spline activation to address the problem of efficiently and accurately extracting brain microstructural parameters in diffusion magnetic resonance imaging. This framework significantly accelerates the acquisition process and reduces the error in microstructural index estimation. Specifically, it includes:

[0134] 1. Supervised learning architecture, such as Figure 2 Figure 1 shows an overview of the supervised learning pipeline, illustrating the workflow from dMRI images to microstructural metrics. (b) The mathematical formula for the convo-kan operation details the nonlinear convolution mechanism based on spline activation. (c) A detailed description of the network architecture compares the Plain-KAN architecture with the Connect-KAN architecture, which uses skip connections to effectively enhance feature propagation and preserve spatial information.

[0135] The supervised learning architecture Connect-KAN proposed in this application uses Conv-KANs to capture the complex nonlinearities inherent in dMRI data. Unlike traditional CNNs, which rely on linear convolution operations and fixed activation functions, Conv-KANs implement convolution operations based on the Kolmogorov-Arnold representation theorem. According to this theorem, any continuous multivariate function can be approximated using a finite combination and superposition of univariate functions. Leveraging this theorem, Conv-KANs replace traditional linear convolutions with activation functions based on adaptive splines, significantly enhancing the representational power of the convolution operation.

[0136] Formally, the 3D Conv-KAN operation for input volumetric data is Among them C in Denotes the input channel and represents the spatial dimension, D, H, W represent the spatial dimension:

[0137]

[0138] in, It is for each spatial offset (a, b, e) and channel pair in the convolution kernel Separately learned nonlinear spline activation functions. Each spline activation function φ(z) is explicitly parameterized as a linear combination of spline basis functions B k (z):

[0139]

[0140] α k As a trainable spline parameter, K represents the number of spline nodes.

[0141] In order to effectively reduce the computational complexity introduced by numerous spline parameters, this application specifically adopts 3DBottleneck-Conv-KAN to extend the bottleneck design to three-dimensional space. Each Bottleneck-Conv-KAN layer consists of three different steps. First, linear compression is performed by 1×1×1 convolution to convert the channel from C in Reduce to the bottleneck dimension C b :

[0142]

[0143] Then, the compressed feature Z b,d,h,w Perform pixel-wise nonlinear spline activation:

[0144]

[0145] Finally, the bottleneck feature is expanded to the required number of output channels C by linear expansion through another 1×1×1 convolution out :

[0146]

[0147] Among them, W b,c ,V c,b are the learnable weights of the linear convolution. By using these Bottleneck-Conv-KANs in a supervised learning architecture, the number of parameters and computational load can be further reduced while still maintaining strong nonlinear approximation capabilities.

[0148] In addition, the supervised Connect-KAN adopts skip connections between multiple stacked Bottleneck-Conv-KAN layers, which promotes efficient gradient propagation and detailed spatial information preservation, ultimately improving prediction accuracy and training stability. For ablation comparison, the Plain-KAN architecture without skip connections is also used.

[0149] Second, self-supervised learning architecture, such as Figure 3 As shown in

[15] , the Connect-KAN architecture is extended to the setting without external data and training reference data. The model exploits inherent dMRI signal relationships, avoids explicit dependence on reference labels, and promotes accurate microstructure estimation through self-supervision.

[0150] The self-supervised learning framework DIMOND-KAN in this application uses diffusion MRI data of a single subject to train the network, that is, subject-specific training. For each subject, Connect-KAN maps the input volume Y to the diffusion model parameters and minimizes the difference between the input volume Y and the network output through the forward diffusion model input. The resulting composite volume The difference between them is used to optimize the network parameters:

[0151]

[0152] The forward diffusion modeling function Modeling(·) can be derived from any existing or new MRI diffusion modeling method. This application includes the tensor model and the NODDI model, which follows the implementation method in our previous DIMOND work.

[0153] The formula of the tensor model is as follows:

[0154] Tensor(S0,D)=S0e -bAD ;

[0155] where A and b represent the diffusion tensor transformation matrix and b value, respectively, both determined by the acquisition protocol. D and S0 represent the diffusion tensor and non-diffusion weighted signal estimated by DIMOND-KAN.

[0156] The implementation of the NODDI model is as follows:

[0157] NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso );

[0158] Among them, A ic (μ,κ),A ec (μ,f ic ),A iso Represent the diffusion signals of the intracellular compartment, extracellular compartment and isotropic compartment respectively; f ic ,1-f ic ,f iso represent the volume fractions of the intracellular compartment, extracellular compartment and isotropic compartment, respectively.

[0159] To evaluate the effectiveness of our diffusion MRI microstructural parameter estimation method, we also tested it using preprocessed dMRI data from the Human Connectome Project Consortium public database. We selected 93 unrelated healthy subjects, 83 of whom were assigned for training and 10 were reserved for evaluation of the supervised learning framework. The diffusion data from the same 10 subjects were also used in self-supervised learning experiments for subject-specific training and evaluation.

[0160] The dMRI (Diffusion Magnetic Resonance Imaging) acquisition protocol included multiple b values (b = 0, 1000, 2000, and 3000 s / mm²), with 18 b = 0 volumes and 90 DWIs collected per nonzero b shell. These b = 0 volumes were interspersed with 15 DWI volumes each. Notably, the sampling order of all DWIs was optimized to ensure that the sequentially acquired diffusion encoding directions remained uniformly distributed across the sphere, even if the scan was terminated prematurely. For comprehensive artifact correction, each DWI was acquired with both left-right and right-left phase-encoding directions. During preprocessing, distortions induced by magnetic susceptibility and eddy currents were corrected, all images were coregistered, and DWI volumes with opposite phase-encoding directions were combined using FSL software.

[0161] In the DTI experiment, a subset of dMRI data was selected, which included 1 dMRI data with b = 0 and 6 dMRI data with b = 1000s / mm 2 DWI volumes along the optimized direction for each subject were used as highly accelerated DTI data. Furthermore, in the NODDI experiment, a subset of dMRI data was selected, including three b = 0 dMRI data and three 15 DWI volumes each at b = 1000, 2000, and 3000 s / mm2 shells for each subject, as highly accelerated NODDI data. All 18 b = 0 dMRI data and 90 b = 1000 s / mm2 DWI volumes were used to derive reference DTI indices. All 288 b = 0 dMRI data and DWI volumes at different b values were used to derive the reference NODDI indices.

[0162] For DTI, the diffusion tensor was fitted to the highly accelerated and fully sampled dMRI data using the "dtifit" function of FSL, and the results were used as baseline and reference. DTI indicators include axial diffusion coefficient (AD), mean diffusion coefficient (MD), fractional anisotropy (FA), radial diffusion coefficient (RD), and principal eigenvector (V1). For NODDI, the NODDI three-compartment model was fitted to the highly accelerated, fully sampled dMRI data using the MATLAB toolbox of NODDIinventors, and the results were used as baseline and reference. NODDI indicators include isotropic volume fraction (f iso ), intracellular volume fraction (f ic ) and orientation dispersion index (ODI).

[0163] Specifically, this application proposes MicroKAN, a novel deep learning framework specifically tailored for dMRI microstructural modeling. Compared to traditional optimization-based fitting and conventional neural networks, MicroKAN significantly improves the accuracy and robustness of DTI and NODDI parameter estimation. MicroKAN utilizes adaptive spline-based nonlinear convolution operations to effectively exploit the inherent nonlinear relationships in dMRI data, significantly improving modeling accuracy while reducing computational complexity.

[0164] In a supervised learning scenario, we propose Connect-KAN, which integrates the Kolmogorov-Arnold theorem to facilitate highly adaptive nonlinear spline activations. Unlike standard CNNs that use linear convolutions and fixed activation functions, Micro-KAN utilizes spline-based activations within a skip-connection architecture to effectively capture the complex nonlinear dependencies inherent in dMRI signals. Quantitative evaluations demonstrate significantly lower mean average evolutionary error (MAEs) for both scalar and directed diffusion metrics in both DTI and NODDI frameworks, highlighting MicroKAN's superior performance in accurately capturing changes in brain microstructure.

[0165] In addition to supervised learning scenarios, this application also proposes DIMOND-KAN, a self-supervised extension of Connect-KAN, to address scenarios with limited labeled data availability. DIMOND-KAN uniquely exploits the internal consistency of diffusion signals with respect to established biophysical models, eliminating the reliance on reference labels during training. This approach significantly improves parameter estimation quality, producing results nearly equivalent to fully supervised methods. Notably, when applied to severely undersampled dMRI data, DIMOND-KAN significantly reduces error compared to traditional direct fitting methods, highlighting its robustness and clinical applicability in limited data settings.

[0166] The introduction of Bottleneck-KAN-Convs in this application significantly reduces computational complexity while retaining the ability to model highly nonlinear phenomena, a key advantage over traditional CNN methods. Furthermore, the inclusion of jump connections enhances the stability of gradient flow and preserves critical spatial information, further improving estimation accuracy. Therefore, compared to traditional methods, Connect-KAN and DIMOND-KAN not only provide superior quantitative performance but also reconstruct brain microstructures more visually clearer and anatomically more faithfully.

[0167] Furthermore, it is important to note that the MicroKAN model, implemented using the PyTorchDL framework, is specifically designed to directly process full 3D diffusion MRI data, eliminating the need for traditional slice-based preprocessing while maintaining information continuity between slices. The model utilizes a kernel-based attention mechanism to effectively capture the complex spatial features inherent in diffusion MRI data, facilitating accurate prediction of diffusion parameters.

[0168] For the supervised task, data from 83 subjects were used for training, while data from another 10 independent subjects were used exclusively for testing. In the supervised DTI task, training was performed in two stages, initially using 1×10 -4 The learning rate is then reduced to 1×10 -6 The supervised NODDI task also adopts two-stage training, starting from the learning rate 1×10 -4 Started and then dropped to 1×10 -5 For the DTI and NODDI tasks, each stage lasts 100 epochs and takes about 5 hours.

[0169] For the self-supervised task, the data from the same 10 evaluation subjects were used for training and validation. Specifically, the two-stage learning rates for the self-supervised DTI task were 1×10 -4 and 1×10 -5 The two-stage learning rates of the self-supervised NODDI task are 1×10 -3 and 1×10 -4 . For the DTI and NODDI tasks, each stage lasted 200 epochs and took approximately 2 hours. Due to limited GPU memory, training was performed using 64×64×64 voxel blocks, and 12 blocks were extracted for each subject. Both training and validation were performed using an A800 GPU (NVIDIA, Santa Clara, CA), and validation error was closely monitored throughout the process to identify and save the model that achieved the lowest validation error. After training was completed, the optimized network parameters were applied to the entire brain volume of each evaluation subject for performance evaluation.

[0170] The quality of the diffusion model metrics was quantified by calculating the mean absolute errors (MAEs) between the DTI and NODDI estimated metrics and the reference metrics. Specifically, the DTI metrics included AD, MD, FA, RD, and V1, which were evaluated within the brain atlas (i.e., excluding cerebrospinal fluid). The error in V1 was evaluated as the angular deviation between the estimated direction vector and the reference direction vector. For NODDI, the metrics included f iso ,f ic and ODI, which were assessed within the brain mask. iso In voxels close to 1, fic and ODI values become arbitrary. Therefore, f iso Voxels with values exceeding 0.8 are excluded from the f ic and ODI MAE calculations.

[0171] The DTI metric from the supervised Connect-KAN in this application is as follows Figure 4 As shown, Figure 4 Representative coronal images (a, c, e, g) of DTI metrics (FA, AD, MD, RD) estimated using supervised Connect-KAN and other compared methods and their residual maps (b, d, f, h) are compared with the reference map (i column). The lower right corner of the residual map shows the MAE of each map compared to the reference map. The accuracy of the supervised Connect-KAN DTI metrics is shown in Table 1. The MAEs between the results and the reference DTI metrics (FA, AD, MD, RD) are used for comparison of Connect-KAN and other methods. The last column of Table 1 highlights the lowest MAEs.

[0172] Table 1

[0173]

[0174] The NODDI metric from the supervised learning Connect-KAN in this application is as follows Figure 5 As shown, Figure 5 The NODDI parameters (f) estimated using supervised Connect-KAN and other comparison methods are shown. iso ,f ic ,ODI) and their residual maps (b, d, f) compared with the reference map (i column). The lower right corner of the residual map shows the MAE of each map compared with the reference map. The accuracy of the NODDI metric from the supervised learning Connect-KAN is shown in Table 2, and the results are comparable to the reference NODDI metric (f iso , f ic , ODI) are used for comparison between Connect-KAN and other methods, and the last column of Table 2 highlights the lowest MAEs.

[0175] Table 2

[0176]

[0177] The DTI metric based on self-supervised learning DIMOND-KAN in this application is as follows Figure 6As shown, representative coronal images of DTI metrics (FA, AD, MD, RD) estimated using self-supervised DIMOND-KAN and other compared methods are shown (a, c, e, g), as well as residual maps (b, d, f, h) compared to the reference map (column i). The lower right corner of the residual map shows the MAE of each map compared to the reference map. The accuracy of the self-supervised learning DIMOND-KAN DTI metrics is shown in Table 3. The MAEs of the DTI metrics (FA, AD, MD, V1) obtained by DIMOND-KAN and other methods are compared with the reference metrics. The last column of Table 3 highlights the lowest MAEs.

[0178] Table 3

[0179]

[0180] The NODDI metric mapping based on self-supervised learning DIMOND-KAN in this application is as follows Figure 7 As shown, Figure 7 The NODDI parameters (f) estimated using supervised DIMOND-KAN and other comparison methods are shown. iso ,f ic ,ODI) and their residual maps (b, d, f) compared with the reference map (i column). The lower right corner of the residual map shows the MAE of each map compared with the reference map. The accuracy of the NODDI metric based on self-supervised learning DIMOND-KAN is shown in Table 4. The obtained NODDI metric is comparable to the reference NODDI metric (f iso , f ic , ODI) for the comparison between DIMOND-KAN and other methods, and the last column of Table 4 highlights the lowest MAEs.

[0181] Table 4

[0182]

[0183] In summary, the diffusion magnetic resonance imaging microstructure parameter estimation method of the present application has the following main features:

[0184] 1. KAN-based core network architecture innovation. We propose MicroKAN, a 3D convolutional framework that integrates the principles of the Kolmogorov-Arnold network. This framework replaces the fixed activation and linear convolution of traditional CNNs with adaptive spline activation functions to dynamically capture complex nonlinear relationships (such as fiber crossings and axonal heterogeneity) in dMRI signals. We also introduce a skip connection architecture (Connect-KAN) to enhance gradient propagation stability while preserving spatial details, improving the accuracy of microstructural parameter estimation and addressing the limitations of traditional CNN nonlinear modeling capabilities.

[0185] 2. Optimize the bottleneck layer design of the supervised learning module. The Bottleneck-Conv-KAN layer is designed to reduce computational complexity through a three-step process of "linear compression-spline activation-linear expansion," reducing the number of parameters while maintaining nonlinear modeling capabilities. A phased training strategy (such as two-stage learning rate adjustment) is employed to adapt to parameter estimation tasks for different models, such as DTI and NODDI, optimizing network convergence and improving training efficiency and accuracy.

[0186] 3. Physical model consistency mechanism for self-supervised learning modules. We propose a self-supervised framework, DIMOND-KAN, which leverages the internal consistency of diffusion signals and biophysical models (such as tensors and NODDI models) to design a loss function. This framework can be trained on single-subject data without external reference labels. By minimizing the error between the input image and the model-synthesized image, it addresses the scarcity of labeled data and enhances robustness in undersampling scenarios. This framework is suitable for real-time clinical processing and large-scale unlabeled data scenarios.

[0187] 4. Compatibility and generalization of multiple diffusion models. Design a general network architecture to support parameter estimation of DTI, NODDI and other diffusion models (such as diffusion kurtosis imaging) without the need for customized tuning for specific models. Through a unified nonlinear convolution and activation mechanism, the mapping relationship between signals and parameters under different models is automatically learned to achieve FA, AD, f iso , ODI and other high-precision estimates, with both model independence and task scalability.

[0188] According to the diffusion magnetic resonance imaging microstructure parameter estimation method proposed in an embodiment of the present application, diffusion magnetic resonance imaging data can be input into a microstructure model, the microstructure model outputs a diffusion index of the imaging target, and the microstructure parameters of the imaging target are estimated based on the diffusion index, wherein the microstructure model includes multiple stacked hybrid neural network layers, and jump connections are used between the multiple stacked hybrid neural network layers. The hybrid neural network layer performs linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data. Through the nonlinear convolution operation of the adaptive spline, the inherent nonlinear relationship in the diffusion magnetic resonance imaging data is effectively utilized, which significantly improves the modeling accuracy, reduces the computational complexity, and improves the computational efficiency. In addition, the jump connection method can promote efficient gradient propagation and detailed spatial information retention, thereby improving the prediction accuracy and training stability.

[0189] Next, a device for estimating microstructure parameters by diffusion magnetic resonance imaging according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0190] Figure 8 4 is a block diagram of a device for estimating microstructure parameters by diffusion magnetic resonance imaging according to an embodiment of the present application.

[0191] like Figure 8 As shown, the diffusion magnetic resonance imaging microstructure parameter estimation device 10 includes: an acquisition module 100 , an input module 200 and an evaluation module 300 .

[0192] Among them, the acquisition module 100 is used to acquire the diffusion magnetic resonance imaging data of the imaging target; the input module 200 is used to input the diffusion magnetic resonance imaging data into the microstructure model, and the microstructure model outputs the diffusion index of the imaging target, wherein the microstructure model includes multiple stacked hybrid neural network layers, and the multiple stacked hybrid neural network layers are connected by jump connections. The hybrid neural network layers perform linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data; the estimation module 300 is used to estimate the microstructure parameters of the imaging target based on the diffusion index.

[0193] In an embodiment of the present application, the hybrid neural network layer includes a bottleneck Kolmogorov-Arnold convolution layer and a jump connection structure, wherein the bottleneck Kolmogorov-Arnold convolution layer includes a first convolution, linear compression, a second convolution and a linear expansion, the first convolution and the second convolution are different, the second convolution is a Kolmogorov-Arnold convolution, the bottleneck Kolmogorov-Arnold convolution layer performs linear compression on the input data to obtain compressed features, the second convolution performs nonlinear spline activation on the compressed features according to pixels to obtain nonlinear features, and the bottleneck Kolmogorov-Arnold convolution layer performs linear expansion on the nonlinear features.

[0194] In the embodiment of the present application, the linear compression formula in the bottleneck Kolmogorov-Arnold convolutional layer is:

[0195]

[0196] Among them, Z b,d,h,w is the eigenvalue after linear compression, with channel b and spatial position (d, h, w); d is the depth dimension; h is the height dimension; w is the width dimension; C in is the number of channels of input data; W b,c is the learnable weight of the 1×1×1 convolution during the linear compression process; Y c,d,h,w The input to the bottleneck structure layer is channel c, and the eigenvalue in the three-dimensional space position is (d, h, w);

[0197] The nonlinear spline activation formula for the second convolution in the bottleneck Kolmogorov-Arnold convolution layer is:

[0198]

[0199] in, For Z b,d,h,w Eigenvalue after nonlinear spline activation; α b,k is the learnable coefficient in the nonlinear spline activation function; B k represents the kth spline basis function; φ b is a nonlinear spline activation function;

[0200] The linear expansion formula of the bottleneck Kolmogorov-Arnold convolutional layer is:

[0201]

[0202] in, is the output eigenvalue after linear expansion, with channel c and spatial position (d, h, w); C b is the bottleneck dimension; is the learnable weight of the 1×1×1 convolution during the linear expansion process; For Z b,d,g,w Eigenvalues after nonlinear spline activation.

[0203] In the embodiment of the present application, the device 10 of the embodiment of the present application further includes: a training module.

[0204] Among them, the training module is used to obtain training data before inputting the diffusion magnetic resonance imaging data into the microstructure model, wherein the training data includes multiple diffusion nuclear magnetic resonance imaging data; use the training data to train the microstructure model, and during the training process, optimize the network parameters of the microstructure model based on the loss function established by the forward diffusion model function.

[0205] In the embodiment of the present application, the loss function is:

[0206]

[0207] Among them, E is the mathematical expectation; Y is the true value; is the predicted value; For the model according to the parameters The process of processing input data and obtaining prediction results.

[0208] In the embodiment of the present application, the estimation module 300 is further configured to: utilize the diffusion index to input a forward diffusion model function, and the forward diffusion model function outputs microstructure parameters of the imaging target.

[0209] In the embodiment of the present application, the forward diffusion model function includes a tensor model and a neurite directional diffusion and density imaging model, wherein,

[0210] The formula of the tensor model is:

[0211] Tensor(S0,D)=S0e -bAD ;

[0212] Where S0 is the magnetic resonance signal intensity without diffusion weighting; D is the diffusion coefficient; b is the diffusion sensitivity factor, both D and b are determined by the acquisition protocol; A is the diffusion tensor transformation vector, which is determined by the diffusion gradient encoding direction; Tensor is the tensor model;

[0213] The formula for the neurite directional diffusion and density imaging model is:

[0214] NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso );

[0215] Where S0 is the magnetic resonance signal intensity without diffusion weighting; f iso is the volume fraction of the isotropic diffusion component; f ic is the volume fraction of the intracellular diffusive component in the non-isotropic part; κ is the dispersion parameter of neurite orientation; μ is the average orientation of neurites; A ic is the signal attenuation function of intracellular diffusion; A ec is the signal attenuation function of extracellular diffusion; A iso is the signal attenuation function of isotropic diffusion; NODDI is the neurite directional diffusion and density imaging model.

[0216] It should be noted that the above explanations of the embodiment of the diffusion magnetic resonance imaging microstructure parameter estimation method are also applicable to the diffusion magnetic resonance imaging microstructure parameter estimation device of this embodiment, and will not be repeated here.

[0217] According to the diffusion magnetic resonance imaging microstructure parameter estimation device proposed in the embodiment of the present application, diffusion magnetic resonance imaging data can be input into a microstructure model, the microstructure model outputs a diffusion index of the imaging target, and the microstructure parameters of the imaging target are estimated based on the diffusion index, wherein the microstructure model includes multiple stacked hybrid neural network layers, and jump connections are used between the multiple stacked hybrid neural network layers. The hybrid neural network layer performs linear compression, nonlinear spline activation and linear expansion on the diffusion magnetic resonance imaging data. Through the nonlinear convolution operation of the adaptive spline, the inherent nonlinear relationship in the diffusion magnetic resonance imaging data is effectively utilized, which significantly improves the modeling accuracy, reduces the computational complexity, and improves the computational efficiency. In addition, the jump connection method can promote efficient gradient propagation and detailed spatial information retention, thereby improving the prediction accuracy and training stability.

[0218] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include:

[0219] A memory 901 , a processor 902 , and a computer program stored in the memory 901 and executable on the processor 902 .

[0220] When the processor 902 executes the program, the diffusion magnetic resonance imaging microstructure parameter estimation method provided in the above embodiment is implemented.

[0221] Furthermore, the electronic device further includes:

[0222] The communication interface 903 is used for communication between the memory 901 and the processor 902 .

[0223] The memory 901 is used to store computer programs that can be run on the processor 902 .

[0224] The memory 901 may include a high-speed RAM (Random Access Memory) memory, and may also include a non-volatile memory, such as at least one disk memory.

[0225] If the memory 901, the processor 902, and the communication interface 903 are implemented independently, the communication interface 903, the memory 901, and the processor 902 can be connected to each other via a bus and communicate with each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0226] Optionally, in a specific implementation, if the memory 901, the processor 902 and the communication interface 903 are integrated on a chip, the memory 901, the processor 902 and the communication interface 903 can communicate with each other through an internal interface.

[0227] The processor 902 may be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0228] An embodiment of the present application further provides a computer-readable storage medium having a computer program or instruction stored thereon. When the computer program or instruction is executed by a processor, the above-mentioned diffusion magnetic resonance imaging microstructure parameter estimation method is implemented.

[0229] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0230] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0231] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing a custom logical function or process step, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed in a different order than shown or discussed, including performing functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application pertain.

[0232] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, the steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement the method: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0233] A person skilled in the art may understand that all or part of the steps carried out in the method for implementing the above-mentioned embodiment may be completed by instructing the relevant hardware through a program, and the above-mentioned program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0234] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for estimating microstructure parameters using diffusion magnetic resonance imaging, characterized in that: The following steps are involved: acquiring diffusion magnetic resonance imaging data of an imaging target; Inputting the diffusion magnetic resonance imaging data into a microstructure model, the microstructure model outputting a diffusion index of the imaging target, wherein the microstructure model comprises a plurality of stacked hybrid neural network layers, wherein the plurality of stacked hybrid neural network layers are connected by jump connections, and the hybrid neural network layers perform linear compression, nonlinear spline activation, and linear expansion on the diffusion magnetic resonance imaging data; The microstructure parameters of the imaging target are estimated according to the diffusion index.

2. The method for estimating microstructure parameters using diffusion magnetic resonance imaging according to claim 1, wherein: The hybrid neural network layer includes a bottleneck Kolmogorov-Arnold convolution layer and a jump connection structure, wherein the bottleneck Kolmogorov-Arnold convolution layer includes a first convolution, linear compression, a second convolution and a linear expansion, the first convolution and the second convolution are different, the second convolution is a Kolmogorov-Arnold convolution, the bottleneck Kolmogorov-Arnold convolution layer performs linear compression on the input data to obtain a compressed feature, the second convolution performs nonlinear spline activation on the compressed feature according to pixels to obtain a nonlinear feature, and the bottleneck Kolmogorov-Arnold convolution layer performs linear expansion on the nonlinear feature.

3. The method for estimating microstructure parameters using diffusion magnetic resonance imaging according to claim 2, wherein: The linear compression formula in the bottleneck Kolmogorov-Arnold convolutional layer is: Among them, Z b,d,h,w is the eigenvalue after linear compression, with channel b and spatial position (d, h, w); d is the depth dimension; h is the height dimension; w is the width dimension; C in is the number of channels of input data; W b,c is the learnable weight of the 1×1×1 convolution during the linear compression process; Y c,d,h,w The input to the bottleneck structure layer is channel c, and the eigenvalue in the three-dimensional space position is (d, h, w); The nonlinear spline activation formula for the second convolution in the bottleneck Kolmogorov-Arnold convolutional layer is: in, For Z b,d,h,w Eigenvalue after nonlinear spline activation; α b,k is the learnable coefficient in the nonlinear spline activation function; B k represents the kth spline basis function; φ b is a nonlinear spline activation function; The linear expansion formula of the bottleneck Kolmogorov-Arnold convolutional layer is: in, is the output eigenvalue after linear expansion, with channel c and spatial position (d, h, w); C b is the bottleneck dimension; is the learnable weight of the 1×1×1 convolution during the linear expansion process; For Z b,d,h,w Eigenvalues after nonlinear spline activation.

4. The method for estimating microstructure parameters using diffusion magnetic resonance imaging according to claim 1, wherein: Before inputting the diffusion magnetic resonance imaging data into the microstructure model, the method further includes: Acquiring training data, wherein the training data includes a plurality of diffusion magnetic resonance imaging data; The microstructure model is trained using the training data. During the training process, the network parameters of the microstructure model are optimized based on a loss function established based on a forward diffusion model function.

5. The method for estimating microstructure parameters using diffusion magnetic resonance imaging according to claim 4, wherein: The loss function is: Among them, E is the mathematical expectation; Y is the true value; is the predicted value; For the model according to the parameters The process of processing input data and obtaining prediction results.

6. The method for estimating microstructure parameters using diffusion magnetic resonance imaging according to claim 1, wherein: The estimating the microstructure parameters of the imaging target according to the diffusion index includes: The diffusion index is used to input a forward diffusion model function, and the forward diffusion model function outputs microstructure parameters of the imaging target.

7. The method for estimating microstructure parameters using diffusion magnetic resonance imaging according to claim 4 or 6, wherein: The forward diffusion model function includes a tensor model and a neurite directional diffusion and density imaging model, wherein, The formula of the tensor model is: Tensor(S0,D)=S0e -bAD ; Where S0 is the magnetic resonance signal intensity without diffusion weighting; D is the diffusion coefficient; b is the diffusion sensitivity factor, both D and b are determined by the acquisition protocol; A is the diffusion tensor transformation vector, which is determined by the diffusion gradient encoding direction; Tensor is the tensor model; The formula for the neurite directional diffusion and density imaging model is: NODDI(S0,f iso ,f ic ,κ,μ)=S0((1-f iso )(f ic A ic +(1-f ic )A ec )+f iso A iso ); Where S0 is the magnetic resonance signal intensity without diffusion weighting; f iso is the volume fraction of the isotropic diffusion component; f ic is the volume fraction of the intracellular diffusive component in the non-isotropic part; κ is the dispersion parameter of neurite orientation; μ is the average orientation of neurites; A ic is the signal attenuation function of intracellular diffusion; A ec is the signal attenuation function of extracellular diffusion; A iso is the signal attenuation function of isotropic diffusion; NODDI is the neurite directional diffusion and density imaging model.

8. A device for estimating microstructure parameters using diffusion magnetic resonance imaging, characterized in that: include: an acquisition module, configured to acquire diffusion magnetic resonance imaging data of an imaging target; an input module, configured to input the diffusion MRI data into a microstructure model, wherein the microstructure model outputs a diffusion index of the imaging target, wherein the microstructure model comprises a plurality of stacked hybrid neural network layers, wherein the plurality of stacked hybrid neural network layers are connected by jump connections, and the hybrid neural network layers perform linear compression, nonlinear spline activation, and linear expansion on the diffusion MRI data; An estimation module is used to estimate the microstructure parameters of the imaging target according to the diffusion index.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for estimating microstructure parameters by diffusion magnetic resonance imaging according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed, the method for estimating microstructure parameters by diffusion magnetic resonance imaging according to any one of claims 1 to 7 is implemented.

Images