Microstructure model optimization method and device based on diffusion prior and electronic equipment

Abstract

The invention relates to the technical field of medical image processing, in particular to a microstructure model optimization method and device based on diffusion prior and electronic equipment, and the method comprises the steps: encoding microstructure parameters of a microstructure model to a potential space through a vector quantization auto-encoder; a potential diffusion model is adopted to learn diffusion parameter distribution priori of microstructure parameters in a potential space; determining a noise potential feature of a previous time step according to diffusion parameter distribution prior, denoising the noise potential feature of the previous time step to obtain a noise potential feature of a current time step, and optimizing the noise potential feature of the current time step according to actually collected diffusion magnetic resonance data; and denoising the optimized noise potential features to obtain noiseless potential features, generating synthetic data by using the microstructure model and decoded noiseless potential feature parameters, and optimizing microstructure parameters of the microstructure model according to the synthetic data and the dispersion magnetic resonance data. Therefore, the problems of insufficient generalization ability, poor universality 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 optimizing a microstructure model based on diffusion prior. Background Art

[0002] Diffusion magnetic resonance imaging (MRI) is an important medical imaging technique widely used to model and map the microstructure of brain tissue. By measuring the diffusion of water molecules through tissue, this technique provides crucial information about tissue microstructure, which is of great value in neuroscience research and clinical diagnosis.

[0003] Related technologies employ deep learning-based techniques, particularly neural network-based approaches, and have demonstrated superior performance in generating high-quality microstructural model parameters. These methods typically train neural networks to directly learn the mapping from diffusion-weighted images from a specific acquisition configuration to microstructural parameters, reducing the tedious mathematical calculations required by traditional methods. These neural networks include multilayer perceptrons, transformer networks, and recurrent neural networks. To address the problem of data from different acquisition configurations not being generalizable, some algorithms unify data from specific acquisition configurations into a standardized data format before training the neural network.

[0004] However, the deep learning methods of related technologies are mainly trained based on specific acquisition configurations, and their generalization capabilities are insufficient. In addition, some methods have strong model specificity, a narrow scope of application, and are sensitive to noise, resulting in a significant decrease in performance under low signal-to-noise ratio conditions. Summary of the Invention

[0005] The present application provides a microstructure model optimization method, device and electronic device based on diffusion prior to solve the problems of insufficient generalization ability and poor versatility of related technologies.

[0006] The first aspect of the present application provides a microstructure model optimization method based on diffusion prior, comprising the following steps: encoding the microstructure parameters of the microstructure model into a latent space through a vector quantization autoencoder; using a latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space; determining the noise latent feature of the previous time step based on the diffusion parameter distribution prior, denoising the noise latent feature of the previous time step to obtain the noise latent feature of the current time step, and optimizing the noise latent feature of the current time step based on the actually collected diffusion magnetic resonance data; denoising the optimized noise latent feature to obtain a noise-free latent feature, decoding the noise-free latent feature through a decoder, generating synthetic data using the microstructure model and the decoded parameters, and optimizing the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data.

[0007] Optionally, the vector quantization autoencoder includes multiple downsampling layers and residual blocks, and the microstructure parameters of the microstructure model are encoded into the latent space through the vector quantization autoencoder, including: dividing the microstructure parameters into data blocks through multiple downsampling layers; and mapping the data blocks to the latent space through the residual blocks.

[0008] Optionally, the encoding formula of the vector quantization autoencoder is:

[0009]

[0010] z=En(x);

[0011] Where En(·) is the encoder neural network; x is the microstructure parameter; De(·) is the decoder neural network; E x is the expected operation; z is the potential feature representation.

[0012] Optionally, the potential diffusion model is a U-shaped network structure with multi-layer downsampling and self-attention modules, wherein the U-shaped network structure obtains a diffusion parameter distribution prior by approximating the potential distribution through denoising score matching. The formula of the U-shaped network structure is:

[0013]

[0014] Among them, s θ is the attention U-shaped network; z t is the potential feature of time step t; ∈ is the noise sampled from the standard normal distribution N(0,1); is the model parameter of the U-shaped network; E En(x),∈～N(0,1),t is the expectation operation.

[0015] Optionally, before using the latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space, it also includes: obtaining the minimized mean square error between the predicted noise and the actual added noise; using the mean square error to train the latent diffusion model, and during the training process, using multi-scale feature extraction and a jump connection mechanism.

[0016] Optionally, the noise latent feature at the current time step is generated as:

[0017]

[0018] Among them, z′ t is the potential feature of temporary noise at the current time step; z t+1 is the noise potential feature of the previous time step; is a predefined noise scheduling parameter; For z t+1 Noise-free estimate of s θ (z t+1,t+1) is the noise prediction calculated by the attention U-network; η is the step size parameter that controls the intensity of random sampling; δ t+1 is the time-step-dependent noise coefficient; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0019] Optionally, the optimization formula for the noise latent feature at the current time step is:

[0020]

[0021] Among them, z t is the noise potential feature of the current time step; z′ t is the potential feature of temporary noise at the current time step; ζ is the learning rate of gradient descent, which controls the optimization step size; For z′ t The gradient operator is ; S is the actual diffusion magnetic resonance data collected; De(·) is the decoder neural network of the potential diffusion model; Modeling(·) is the forward microstructure model used to generate synthetic data from the decoded parameters; is the mean square error.

[0022] Optionally, optimizing the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data includes: inputting the synthetic data and the diffusion magnetic resonance data into a loss function, and optimizing the microstructure parameters of the microstructure model using the loss function, wherein the loss function is:

[0023]

[0024] Where S is the actual diffusion magnetic resonance data collected; z0 is the Gaussian noise z T Initially, the noise-free latent features are obtained after T-step iterative denoising; Modeling(·) is the forward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by inputting the diffusion parameters into the microstructure model after mapping the noise-free latent features into the diffusion parameters through the decoder.

[0025] The second aspect of the present application provides a microstructure model optimization device based on diffusion prior, including: an encoding module for encoding the microstructure parameters of the microstructure model into a latent space through a vector quantization autoencoder; a learning module for learning the diffusion parameter distribution prior of the microstructure parameters in the latent space using a latent diffusion model; a determination module for determining the noise latent feature of the previous time step based on the diffusion parameter distribution prior, denoising the noise latent feature of the previous time step to obtain the noise latent feature of the current time step, and optimizing the noise latent feature of the current time step based on the actually collected diffusion magnetic resonance data; an optimization module for denoising the optimized noise latent feature to obtain a noise-free latent feature, decoding the noise-free latent feature through a decoder, generating synthetic data using the microstructure model and the decoded parameters, and optimizing the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data.

[0026] Optionally, the vector quantization autoencoder includes multiple downsampling layers and residual blocks, and the encoding module is further used to: divide the microstructure parameters into data blocks through the multiple downsampling layers; and map the data blocks to the latent space through the residual blocks.

[0027] Optionally, the encoding formula of the vector quantization autoencoder is:

[0028]

[0029] z=En(x);

[0030] Where En(·) is the encoder neural network; x is the microstructure parameter; De(·) is the decoder neural network; E x is the expected operation; z is the potential feature representation.

[0031] Optionally, the potential diffusion model is a U-shaped network structure with multi-layer downsampling and self-attention modules, wherein the U-shaped network structure obtains a diffusion parameter distribution prior by approximating the potential distribution through denoising score matching. The formula of the U-shaped network structure is:

[0032]

[0033] Among them, s θ is the attention U-shaped network; z t is the potential feature of time step t; ∈ is the noise sampled from the standard normal distribution N(0,1); is the model parameter of the U-shaped network; E En(x),∈～N(0,1),t is the expectation operation.

[0034] Optionally, it also includes: a training module for obtaining the minimized mean square error between the predicted noise and the actual added noise before using the potential diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space; using the mean square error to train the latent diffusion model, and using multi-scale feature extraction and jump connection mechanism during the training process.

[0035] Optionally, the noise latent feature at the current time step is generated as:

[0036]

[0037] Among them, z′ t is the potential feature of temporary noise at the current time step; z t+1 is the noise potential feature of the previous time step; is a predefined noise scheduling parameter; For z t+1 Noise-free estimate of s θ (z t+1 ,t+1) is the noise prediction calculated by the attention U-network; η is the step size parameter that controls the intensity of random sampling; δ t+1 is the time-step-dependent noise coefficient; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0038] Optionally, the optimization formula for the noise latent feature at the current time step is:

[0039]

[0040] Among them, z t is the noise potential feature of the current time step; z′ t is the potential feature of temporary noise at the current time step; ζ is the learning rate of gradient descent, which controls the optimization step size; For z′ t The gradient operator is ; S is the actual diffusion magnetic resonance data collected; De(·) is the decoder neural network of the potential diffusion model; Modeling(·) is the forward microstructure model used to generate synthetic data from the decoded parameters; is the mean square error.

[0041] Optionally, the optimization module is further configured to: input the synthetic data and the diffusion magnetic resonance data into a loss function, and optimize the microstructure parameters of the microstructure model using the loss function, wherein the loss function is:

[0042]

[0043] Where S is the actual diffusion magnetic resonance data collected; z0 is the Gaussian noise z TInitially, the noise-free latent features are obtained after T-step iterative denoising; Modeling(·) is the forward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by inputting the diffusion parameters into the microstructure model after mapping the noise-free latent features into the diffusion parameters through the decoder.

[0044] The 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 perform the microstructure model optimization method based on diffusion prior as described in the above embodiment.

[0045] Therefore, this application has at least the following beneficial effects:

[0046] The embodiment of the present application can encode the microstructure parameters of the microstructure model into the latent space, and adopt the latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space, and determine the noise latent feature of the previous time step based on the diffusion parameter distribution prior, denoise the noise latent feature of the previous time step to obtain the noise latent feature of the current time step, and optimize the noise latent feature of the current time step based on the diffusion magnetic resonance data actually collected. In the process of obtaining the noise latent feature and denoising, the prior knowledge learned by the latent diffusion model plays a role, so that the latent diffusion model can effectively identify and suppress noise, and can stably process data even under low signal-to-noise ratio conditions, reduce the interference of noise on the parameter estimation results, and then denoise the optimized noise latent feature to obtain the noise-free latent feature, and the noise-free latent feature is decoded by the decoder.

[0047] Decoding is performed, and synthetic data is generated using the microstructure model and decoded parameters. The microstructure parameters of the microstructure model are optimized based on the synthetic data and diffusion magnetic resonance data. The latent features are continuously adjusted through multiple denoising and optimization processes, making the final decoded synthetic data more realistic. This significantly improves the accuracy of parameter estimation for various microstructure models. Furthermore, because the latent diffusion model learns a universal diffusion parameter distribution prior and does not rely on a specific acquisition configuration, it can accurately determine the noise latent features and perform subsequent processing when faced with different data, maintaining high-quality parameter estimation and improving generalization and versatility. This solves technical problems such as insufficient generalization and poor versatility in related technologies.

[0048] 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

[0049] 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:

[0050] Figure 1 Flowchart of a microstructure model optimization method based on diffusion prior according to an embodiment of the present application;

[0051] Figure 2 A structural diagram of a microstructure model optimization method based on diffusion prior provided in an embodiment of the present application;

[0052] Figure 3 A schematic diagram showing a comparison of effects in tensor model parameter estimation according to an embodiment of the present application;

[0053] Figure 4 A schematic diagram showing a first effect comparison in kurtosis model parameter estimation according to an embodiment of the present application;

[0054] Figure 5 A schematic diagram showing a comparison of the second effect in kurtosis model parameter estimation according to an embodiment of the present application;

[0055] Figure 6 Schematic diagram showing the comparison of the effects of NODDI model parameter estimation according to the embodiments of the present application;

[0056] Figure 7 This is an example diagram of a microstructure model optimization device based on diffusion prior according to an embodiment of the present application;

[0057] Figure 8 A schematic diagram of the structure of an electronic device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0058] The following describes in detail embodiments of the present application, examples of which 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.

[0059] In order to assist in understanding the solution of this application, the related technology is first described.

[0060] Deep learning-based techniques, particularly those utilizing neural networks, have demonstrated superior performance in generating high-quality microstructural model parameters. These methods typically train neural networks to directly learn the mapping from diffusion-weighted images obtained from a specific acquisition configuration to microstructural parameters, reducing the tedious mathematical computations required in traditional methods. These neural networks include multilayer perceptrons, transformer networks, and recurrent neural networks. Some algorithms, to address the problem of data acquired from different acquisition configurations not being generalizable, unify data acquired from specific acquisition configurations into a standard data format before training the neural network. For example, deep tensor imaging networks use tensor models to convert arbitrary data into diffusion-weighted images with the same acquisition parameters. Furthermore, some studies have attempted to convert diffusion data into a spherical harmonic representation space to improve model applicability.

[0061] However, the related technology has the following problems:

[0062] Insufficient generalization: Most neural network-based methods are typically trained for a specific acquisition configuration (b-value and diffusion weighting direction) and cannot process data acquired with other configurations. This severely limits the applicability of these methods to data acquired with different configurations.

[0063] Model specificity: Although deep tensor imaging networks partially solve the generalization problem, their design is only applicable to tensor models and cannot be extended to other microstructure models.

[0064] Poor data sparsity handling: Related methods do not work well with sparsely sampled q-space data. Some methods that attempt to transform diffuse data into a common space often fail when faced with very sparse samples.

[0065] High retraining cost: Existing technologies require retraining for each new encoding scheme, which increases the time and computing cost in practical applications.

[0066] Noise sensitivity: Existing methods are sensitive to noise in diffusion MRI signals, and their performance degrades significantly under low signal-to-noise ratio conditions.

[0067] Limitations of clinical application: Due to the above-mentioned defects, the application of existing technologies in microstructure and connectivity mapping in clinical and neuroscience research is limited, making it difficult to fully realize the potential of diffusion MRI.

[0068] To this end, the present application provides a microstructure model optimization method based on diffusion prior, in which the microstructure parameters of the microstructure model can be encoded into a latent space, and the diffusion parameter distribution prior of the microstructure parameters in the latent space can be learned using a latent diffusion model, and the noise latent characteristics of the previous time step are determined based on the diffusion parameter distribution prior, and the noise latent characteristics of the previous time step are denoised to obtain the noise latent characteristics of the current time step, and the noise latent characteristics of the current time step are optimized based on the actually collected diffusion magnetic resonance data. In the process of obtaining the noise latent characteristics and denoising, the prior knowledge learned by the latent diffusion model plays a role, allowing the latent diffusion model to effectively identify and suppress noise, and can stably process data even under low signal-to-noise ratio conditions, reducing the impact of noise on parameters. The noise estimation results are interfered with, and the optimized noise latent features are denoised to obtain noise-free latent features. The noise-free latent features are decoded by a decoder, and synthetic data are generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized according to the synthetic data and diffusion magnetic resonance data. The latent features are continuously adjusted through multiple denoising and optimization processes, so that the synthetic data finally generated by decoding is more in line with the actual situation, which greatly improves the accuracy of parameter estimation of various microstructure models. Moreover, since the latent diffusion model learns a universal diffusion parameter distribution prior and does not rely on a specific acquisition configuration, it can accurately determine the noise latent features and perform subsequent processing when facing different data, maintain high-quality parameter estimation, and improve generalization and versatility.

[0069] Specifically, Figure 1 A flow chart of a microstructure model optimization method based on diffusion prior provided in an embodiment of the present application.

[0070] like Figure 1 As shown, the microstructure model optimization method based on diffusion prior includes the following steps:

[0071] In step S101 , the microstructure parameters of the microstructure model are encoded into a latent space through a vector quantization autoencoder.

[0072] Among them, the vector quantization autoencoder includes multiple downsampling layers and residual blocks.

[0073] In an embodiment of the present application, the microstructure parameters of the microstructure model are encoded into a latent space through a vector quantization autoencoder, including: dividing the microstructure parameters into data blocks through multiple downsampling layers; and mapping the data blocks to the latent space through residual blocks.

[0074] It can be understood that the embodiment of the present application can divide the microstructure parameters into data blocks through multiple downsampling layers, and map the data to the latent space through the residual block to complete the encoding of the microstructure parameters into the latent space, which can effectively extract key features.

[0075] In the embodiment of the present application, the encoding formula of the vector quantization autoencoder is:

[0076]

[0077] z=En(x);

[0078] Where En(·) is the encoder neural network; x is the microstructure parameter; De(·) is the decoder neural network; E x is the expected operation; z is the potential feature representation.

[0079] Specifically, the steps of encoding the microstructure model parameters into the latent space by a vector quantized autoencoder in this application include:

[0080] (1) Construct a three-dimensional vector quantization autoencoder with multiple downsampling layers and residual blocks;

[0081] (2) dividing the input microstructure model parameter volume data into data blocks of specific sizes;

[0082] (3) Mapping the data block to a discrete latent codebook space through the encoder;

[0083] (4) The latent representation is reconstructed back into the parameter space through the decoder.

[0084] In step S102 , a latent diffusion model is used to learn a diffusion parameter distribution prior of microstructure parameters in the latent space.

[0085] The potential diffusion model includes at least one of a tensor model, a kurtosis model and a NODDI model.

[0086] It is understandable that the embodiments of the present application may use a latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space so as to perform subsequent posterior sampling.

[0087] In the embodiment of the present application, the potential diffusion model is a U-shaped network structure with multi-layer downsampling and self-attention modules, wherein the U-shaped network structure obtains the diffusion parameter distribution prior by approximating the potential distribution through denoising score matching. The formula of the U-shaped network structure is:

[0088]

[0089] Among them, s θ is the attention U-shaped network; z t is the potential feature of time step t; ∈ is the noise sampled from the standard normal distribution N(0,1); is the model parameter of the U-shaped network; E En(x),∈～N(0,1),t is the expectation operation.

[0090] In an embodiment of the present application, before using the potential diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space, it also includes: obtaining the minimized mean square error between the predicted noise and the actual added noise; using the mean square error to train the latent diffusion model, and during the training process, using multi-scale feature extraction and jump connection mechanism.

[0091] It can be understood that, in the embodiment of the present application, before using the latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space, the latent diffusion model is trained. The specific training process is: training is performed by minimizing the mean square error between the predicted noise and the actual added noise, and multi-scale feature extraction and jump connection mechanism are adopted in the training process.

[0092] In step S103, the noise potential characteristics of the previous time step are determined based on the diffusion parameter distribution a priori, the noise potential characteristics of the previous time step are denoised to obtain the noise potential characteristics of the current time step, and the noise potential characteristics of the current time step are optimized based on the actually collected diffusion magnetic resonance data.

[0093] It can be understood that the embodiments of the present application can obtain the noise potential characteristics of the current time step based on the diffusion parameter distribution prior, and optimize the noise potential characteristics of the current time step based on the actual collected diffusion magnetic resonance data. By continuously optimizing the potential characteristics to make them close to the actual data, combined with the learning ability of the potential diffusion model for data characteristics, sparse sampling data can be effectively processed.

[0094] In the embodiment of the present application, the formula for generating the noise potential feature at the current time step is:

[0095]

[0096] Among them, z′ t is the potential feature of temporary noise at the current time step; z t+1 is the noise potential feature of the previous time step; is a predefined noise scheduling parameter; For z t+1 Noise-free estimate of s θ (z t+1 ,t+1) is the noise prediction calculated by the attention U-network; η is the step size parameter that controls the intensity of random sampling; δ t+1 is the time-step-dependent noise coefficient; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0097] In the embodiment of the present application, the optimization formula for the noise potential feature at the current time step is:

[0098]

[0099] Among them, zt is the noise potential feature of the current time step; z′ t is the potential feature of temporary noise at the current time step; ζ is the learning rate of gradient descent, which controls the optimization step size; For z′ t The gradient operator is ; S is the actual diffusion magnetic resonance data collected; De(·) is the decoder neural network of the potential diffusion model; Modeling(·) is the forward microstructure model used to generate synthetic data from the decoded parameters; is the mean square error.

[0100] Specifically, the present application performs an iteration from high to low for each time stamp t, and generates potential features based on conditional probability, that is, an iterative process is performed from 500 to 0 for the time stamp t. At each time step t, the system performs the following steps from the potential feature z of the previous time step t+1 Generate higher quality latent features z t′ .

[0101] In step S104, the optimized noise latent features are denoised to obtain noise-free latent features, the noise-free latent features are decoded by a decoder, synthetic data are generated using the microstructure model and the decoded parameters, and the microstructure parameters of the microstructure model are optimized based on the synthetic data and the diffusion magnetic resonance data.

[0102] It can be understood that the embodiments of the present application can denoise the optimized noise latent features to obtain noise-free latent features, and decode the noise-free latent features through a decoder, generate synthetic data using the microstructure model and the decoded parameters, optimize the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data, optimize the microstructure parameters of the microstructure model through high-quality data, and improve the robustness to noise processing by matching the potential distribution with the denoising score.

[0103] In an embodiment of the present application, optimizing the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data includes: inputting the synthetic data and the diffusion magnetic resonance data into a loss function, and optimizing the microstructure parameters of the microstructure model using the loss function, wherein the loss function is:

[0104]

[0105] Where S is the actual diffusion magnetic resonance data collected; z0 is the Gaussian noise z TInitially, the noise-free latent features are obtained after T-step iterative denoising; Modeling(·) is the forward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by inputting the diffusion parameters into the microstructure model after mapping the noise-free latent features into the diffusion parameters through the decoder.

[0106] The following describes the microstructure model optimization method based on diffusion prior of the present application through a specific embodiment, which mainly includes:

[0107] 1. Prior learning module.

[0108] In the training phase of this application, the system first trains the autoencoder neural network (encoder neural network En(·), decoder neural network De(·)) to encode the microstructure parameters x into the latent space:

[0109]

[0110] z=En(x);

[0111] Retraining Attention U-Networks θ , approximating the latent distribution by denoising score matching:

[0112]

[0113] Among them, z t represents the potential features at time step t; ∈ is the noise sampled from the standard normal distribution.

[0114] 2. Posterior sampling module.

[0115] In the inference phase of this application, an iterative process is performed on the timestamp t from 500 to 0. At each time step t, the latent feature z of the previous time step is obtained by the following steps: t+1 Generate higher quality latent features z t′ :

[0116] First, calculate the potential feature z′ according to the formula t :

[0117]

[0118] in, Scheduling parameters for predefined noises to control the progress of the denoising process; For z t+1 Noise-free estimate of s θ (z t+1 ,t+1) is the noise prediction calculated by the attention U-network; η is the step size parameter that controls the intensity of random sampling; δ t+1is the time-step-dependent noise coefficient; ∈ is the random noise sampled from the standard normal distribution N(0, 1).

[0119] To ensure the potential estimate z t ′ The consistency between the actual acquired diffusion magnetic resonance data S is optimized using the mean square error t ′, as shown in the following formula:

[0120]

[0121] Among them, ζ is the learning rate of gradient descent, which controls the optimization step size; Indicates about z t ′ is the gradient operator; S is the actual diffusion magnetic resonance data collected; De(·) is the decoder neural network of the potential diffusion model; Modeling(·) represents the forward microstructure model, which is used to generate synthetic data from the decoded parameters. Represents the square of the L2 norm, that is, the mean square error.

[0122] During the calculation process, it is necessary to estimate z′ t Noise-free version The calculation method is:

[0123]

[0124] The formula is based on the theory of diffusion model, from the noise potential feature z′ t Estimate the expected value of the noise-free latent feature z0.

[0125] 3. Physical tuning module.

[0126] To avoid domain shift, this module decodes the neural network by minimizing the difference between the collected and synthesized data:

[0127]

[0128] This application includes tensor models, kurtosis models and NODDI models.

[0129] 4. Microstructure parameter modeling module.

[0130] The tensor model is implemented as follows:

[0131] Tensor(S0,D)=S0e -bAD ;

[0132] Where A and b represent the diffusion tensor transformation matrix and b value, respectively, both of which are determined by the acquisition protocol. D and S0 represent the tensor parameters estimated by this application.

[0133] The kurtosis model is implemented as:

[0134]

[0135] Where P represents the kurtosis model parameters estimated by this application, including D jk , K jklm and S0;v i represents the diffusion coding direction when the i-th image is acquired; b i represents the b value when the i-th image is acquired. Both are determined by the acquisition protocol.

[0136] The implementation of the NODDI model is as follows:

[0137] 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 );

[0138] 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 iso ,f ic represent the isotropic volume fraction and intracellular volume fraction, respectively.

[0139] In general, the microstructure model optimization method based on diffusion prior of the present application includes the following steps:

[0140] 1. Encoding the microstructure model parameters into the latent space through a vector quantized autoencoder;

[0141] The steps of encoding the microstructure model parameters into the latent space through the vector quantization autoencoder include:

[0142] Construct a 3D vector quantized autoencoder with multiple downsampling layers and residual blocks;

[0143] Segmenting the input microstructure model parameter volume data into data blocks of a specific size;

[0144] The data block is mapped to a discrete latent codebook space through the encoder;

[0145] The latent representation is reconstructed back into the parameter space through the decoder.

[0146] 2. Train an attention U-network to approximate the latent distribution by matching denoised scores;

[0147] The steps of training the attention U-network to approximate the potential distribution by matching the denoised scores include:

[0148] Adopt a U-shaped network structure with multi-layer downsampling and self-attention modules;

[0149] Training is performed by minimizing the mean squared error between the predicted noise and the actual added noise;

[0150] Multi-scale feature extraction and skip connection mechanism are adopted during the training process.

[0151] 3. In the inference phase, perform iterations from high to low for each timestamp t to generate latent features based on conditional probabilities;

[0152] 4. Optimize latent features by minimizing the mean square error between the latent estimate and the collected data;

[0153] 5. Fine-tune the decoder of the latent diffusion model using a physical information parameter tuning mechanism;

[0154] 6. Output the final optimized microstructure model parameter map.

[0155] The following describes the microstructure model optimization method based on diffusion prior in this application in a systematic form, which mainly includes the following modules:

[0156] 1. Data preprocessing module, used to receive and preprocess diffusion magnetic resonance data;

[0157] 2. Biophysical prior learning module, used to learn the diffusion parameter distribution prior through the potential diffusion model;

[0158] 3. Subject-specific posterior sampling module, used to estimate parameter maps from collected data based on conditional probabilities;

[0159] Among them, the subject-specific posterior sampling module adopts a multi-step iterative process to optimize the latent features at each time step to reduce the difference with the actual collected data.

[0160] 4. Physical information parameter tuning module, used to fine-tune the model decoder to avoid domain shift;

[0161] 5. Result output module, used to generate and output the final microstructure parameter map.

[0162] 6. Parameter configuration module, used to set parameters such as diffusion steps, learning rate and noise coefficient.

[0163] In addition, it should be noted that all neural network models in this application are implemented on the PyTorch platform. As a dynamic neural network framework, PyTorch provides flexible computational graph construction capabilities and a rich library of deep learning tools, which is suitable for the development and training of complex network models in this application.

[0164] The autoencoder used in this application is a VQ (Vector Quantized) model with the following features: (1) Latent embedding dimension: 7 channels, used to encode key features of microstructure model parameters; (2) Codebook size: 16,000, providing sufficient expressive power to capture the complex distribution of parameter space; (3) Downsampling layers: 3 layers, gradually compressing input data into the latent space; (4) Residual block configuration: Contains 2 residual blocks at each scale to enhance feature extraction capability and training stability. This autoencoder is responsible for mapping the microstructure model parameters to the latent space, providing high-quality feature representation for the subsequent diffusion model.

[0165] The latent spatial denoiser used in this application θ The Attention U-Net is used for this purpose. Its structure includes: (1) Downsampling layer: 4 layers to achieve multi-scale feature extraction; (2) Self-attention module: a self-attention module is set between two residual blocks at each scale to enhance the model's ability to capture long-distance dependencies; (3) Skip connection: the skip connection in the U-Net ensures that low-level features can be directly passed to the decoding stage to retain detailed information.

[0166] The training process uses the following configurations: (1) Data partitioning and enhancement: Each input volume data is randomly divided into 75 blocks of size 64×64×64× channels; flipping along the anatomical left-right axis is used to achieve data enhancement and improve the generalization ability of the model. (2) Optimizer configuration: Adam optimizer is used; autoencoder learning rate: 5×10 -5 ; Latent space denoiser learning rate: 2.5×10 -5 ; Batch size: 12.

[0167] In addition, in order to test the solution of the present application, the present application adopts the following two embodiments for testing.

[0168] Example 1: Diffusion magnetic resonance imaging data of 83 subjects from the Human Connectome Project (HCP) were used as training and testing datasets. The data used had an isotropic spatial resolution of 1.25 mm and were preprocessed to ensure data quality and consistency. Specifically, the data for each subject included: (1) 18 b = 0 volume data, i.e., baseline images without diffusion weighting; (2) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (3) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (4) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (5) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (6) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (7) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (8) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (9) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (10) 18 b = 0 volume data distributed in three diffusion shells (b = 1000, 2000, and 3000 mm); (11) 18 b = 0 volume data distributed in -2 ) with 90 body data per shell.

[0169] To establish reference microstructural model parameter values during training and evaluation, the following method was used in this example to calculate the reference parameters of different microstructural models. Tensor model parameters: Ordinary least squares regression method implemented in FSL software package, based on all b = 0 volume data and b = 1000s mm -2 Single shell configuration calculation; Kurtosis model parameters: Ordinary least squares regression method implemented in MRtrix3 software package, based on all b = 0 volume data and b = 1000, 2000 s mm -2 Double shell configuration calculation; NODDI model parameters: using NODDI-Toolbox, based on all b = 0 volume data and b = 1000, 2000, 3000 smm -2 Calculations for three-shell configurations.

[0170] Example 2: The Chinese Human Connectome Project (CHCP) was used as an out-of-distribution test dataset. This dataset has the following characteristics: (1) Spatial resolution: 1.5 mm isotropic, which is different from the 1.25 mm resolution of the training dataset; (2) Number of subjects: 10 subjects, pre-processed diffusion MRI data; (3) Data composition: Each acquisition contains: 14 b = 0 volume data (reference images), 93 b = 1000 s mm-2 diffusion-weighted volume data; and 92 b = 2000 s mm-2 diffusion-weighted volume data.

[0171] To evaluate the performance of this application on this dataset, the reference parameters were calculated as follows: tensor model parameters: calculated using the FSL software package on all b = 0 volume data and single-shell (b = 1000s mm-2) configuration; kurtosis model parameters: calculated using the MRtrix3 software package on all b = 0 volume data and double-shell configuration (b = 1000 and 2000s mm-2); NODDI model parameters: calculated using the NODDI-Toolbox software package on all b = 0 volume data and double-shell configuration, which is significantly different from the training dataset using a three-shell configuration.

[0172] The evaluation indicators during the test are: tensor indicators include axial diffusion coefficient (AD), mean diffusion coefficient (MD), structural anisotropy (FA), radial diffusion coefficient and principal eigenvector (V1). Kurtosis model indicators include axial diffusion coefficient (AD), mean diffusion coefficient (MD, fractional anisotropy (FA), radial kurtosis coefficient (AK), mean kurtosis coefficient (MK) and radial kurtosis coefficient (RK). NODDI model indicators include isotropic volume fraction (f iso ), intracellular volume fraction (f ic ) and orientation dispersion index (ODI).

[0173] The implementation and effects of the diffusion prior-based microstructure model optimization method of the present application will be described below with reference to specific figures and tables.

[0174] Figure 2 This is a structural diagram of the microstructure model optimization method based on diffusion prior in this application, where (a) represents the training phase of the latent diffusion model (LDM), in which the diffusion model parameters are encoded into the latent space through a vector quantized autoencoder, and the attention U-shaped network is trained to learn the distribution prior of the diffusion model parameters. Figure 2 As shown in Figure 1, the autoencoder consists of an encoder and a decoder, mapping the high-dimensional parameter space to a low-dimensional latent space, while the U-shaped network approximates the latent distribution by matching the denoising scores. (b) represents the iterative optimization process of the inference phase. In this phase, the application performs iterations from high to low for each timestamp, and continuously reduces the difference between the synthetic data generated by the current estimate and the actual collected data through the posterior sampling strategy. Figure 2 As shown in Figure 1, the process uses conditional probability to generate latent features and optimizes them through the gradient descent method to make them more consistent with the actual collected diffusion magnetic resonance data. (c) represents the final decoder fine-tuning stage. In this stage, the application further reduces the difference between the estimated results and the actual data through the physical information parameter tuning mechanism. Figure 2 As shown in Figure 3, this process fine-tunes the decoder of the latent diffusion model by minimizing a specific loss function, effectively avoiding the domain shift problem and improving the accuracy of parameter estimation.

[0175] Figure 3 Comparison of the effects of tensor model parameter estimation for one subject from the Human Connectome Project dataset. The figures show the tensor model metric results (a, c, e, g) estimated using this application and other comparison methods on a representative subject, as well as the residual plots (b, d, f, h) comparing the metric results to the reference. The bottom right corner of the residual plots shows the mean absolute error of each plot compared to the reference.

[0176] Table 1 compares the mean absolute error (MAE) results for tensor model DTI parameter estimation for 10 subjects in the Human Connectome Project dataset. The mean ± standard deviation of the MAE between the tensor model metrics estimated using this application and other comparison methods and the reference values is shown, with the lowest MAE highlighted in the last column of Table 1.

[0177] Table 1

[0178] i ii iii method FSL software Deep Tensor Networks This application a FA 0.0975±0.0077 0.0362±0.0018 0.0355±0.0024 b <![CDATA[MD(μm 2 / ms)]]> 0.0658±0.004 0.0438±0.0038 0.0366±0.0019 c <![CDATA[AD(μm 2 / ms)]]> 0.128±0.0085 0.0599±0.0037 0.0542±0.0030 d V1(°) 26.09±1.15 14.52±0.86 14.31±0.71

[0179] Figure 4 Comparison of the effects of kurtosis model parameter estimation for one subject in the Human Connectome Project dataset. Shown are plots (a, c, e, g) of the kurtosis model metrics estimated using this application and other comparison methods, as well as plots of the residuals (b, d, f, h) compared to the reference, for a representative subject. The bottom right corner of the residual plots shows the mean absolute error of each plot compared to the reference.

[0180] Table 2 compares the mean absolute error (MAE) results for kurtosis model parameter estimation for 10 subjects in the Human Connectome Project dataset. The mean ± standard deviation of the MAE between the kurtosis model metrics estimated using this application and other comparison methods and the reference values is shown. The last column highlights the lowest MAE.

[0181] Table 2

[0182]

[0183] Figure 5 Comparison of the effects of kurtosis model parameter estimation on a representative subject from the Chinese Human Connectome Project dataset. The figures show the kurtosis model metric results (a, c) estimated using this application and other comparison methods, as well as the residuals (b, d) comparing the metric results to the reference. The lower right corner of the residual plots shows the mean absolute error of each plot compared to the reference.

[0184] Table 3 compares the mean absolute error (MAE) results for kurtosis model parameter estimation for 10 subjects in the Chinese Human Connectome Project dataset. The mean ± standard deviation of the MAE between the kurtosis model metrics estimated using this application and other comparison methods and the reference values is shown. The last column of Table 3 highlights the lowest MAE.

[0185] Table 3

[0186]

[0187] Figure 6A graphical comparison of the effects of NODDI model parameter estimation on a subject from the Chinese Human Connectome Project dataset. The figures show the NODDI model metric results (a, c, e) estimated using this application and other comparison methods on a representative subject, as well as the residual plots (b, d, f) comparing the metric results to the reference. The mean absolute error (MAE) of each plot compared to the reference is shown in the lower right corner of the residual plots.

[0188] Table 4 compares the mean absolute error (MAE) of NODDI model parameter estimation results for 10 subjects in the Chinese Human Connectome Project dataset. The mean ± standard deviation of the MAE between the NODDI model metrics estimated using this application and other comparison methods and the reference values is shown. The last column of Table 4 highlights the lowest MAE.

[0189] Table 4

[0190]

[0191] In summary, this application adopts a latent diffusion model to learn the diffusion parameter distribution prior, and uses posterior sampling to sample the parameter map from the conditional probability of given acquisition data, aiming to provide a microstructure model optimization method and system with high generalization and high-quality parameter estimation capabilities. The method can be directly deployed on any data set and is applicable to a variety of microstructure models, significantly improving the application value of diffusion magnetic resonance imaging in clinical and neuroscience research. The method can process diffusion magnetic resonance data of various spatial resolutions and adapt to different diffusion coding schemes. After loading the pre-trained weights, there is no need to retrain the network for data with a specific acquisition configuration. The method is a universal network architecture that supports parameter estimation of all microstructure models without the need for customized tuning for specific models.

[0192] The method of this application has the following effects:

[0193] Improved parameter estimation accuracy: Compared to traditional methods, this application demonstrates higher accuracy in parameter estimation for various microstructural models. In tensor models, the average error in fractional anisotropy (FA) and mean diffusivity (MD) estimates is reduced by 30-40%. In kurtosis models, the accuracy of mean kurtosis (MK) estimates is improved by over 45%.

[0194] Enhanced noise suppression: This application uses a latent diffusion model to learn the diffusion parameter distribution prior, demonstrating strong resistance to noise in the input data. Even under low signal-to-noise ratio conditions, it maintains stable parameter estimation performance, reducing the impact of noise on the results.

[0195] Advantages of sparsely sampled data processing: This application can obtain high-quality parameter estimates from sparsely sampled q-space data, achieving parameter estimation quality close to that of complete data, which is significantly better than traditional methods and existing deep learning methods.

[0196] Adaptability across diffusion encoding schemes: This application demonstrates excellent adaptability to data with varying b-value configurations, number of diffusion directions, and distribution. In testing on the Chinese Human Connectome Project (CHCP) dataset, this application maintains high-quality parameter estimates despite acquisition parameters significantly different from the training data.

[0197] Generalization across spatial resolutions: This application can adapt to diffusion magnetic resonance data of different spatial resolutions, from 1.25mm isotropic resolution in training data to 1.5mm resolution in test data, and even to the 2mm resolution commonly used in clinical practice, while maintaining excellent performance.

[0198] Cross-population stability: This application demonstrates good generalization capabilities across different populations, and model performance is not significantly affected by racial differences.

[0199] Improved clinical diagnostic efficiency: This application supports the acquisition of high-quality microstructural parameter maps from rapidly acquired sparse diffusion magnetic resonance data, which can shorten clinical scanning time from the traditional 15-20 minutes to 3-5 minutes while maintaining diagnostic quality, significantly improving clinical workflow efficiency.

[0200] Empowering neuroscience research: This application provides more reliable and consistent tools for mapping brain tissue microstructure and connectivity for neuroscience research, helping to gain a deeper understanding of microstructural changes during neural development, aging, and disease.

[0201] Multi-center research data integration: Due to its excellent generalization ability, this application can process diffusion magnetic resonance data from different centers, using different scanners and acquisition schemes, providing a unified parameter estimation standard for multi-center studies and promoting the comparability and reproducibility of research results.

[0202] According to the microstructure model optimization method based on diffusion prior proposed in the embodiment of the present application, the microstructure parameters of the microstructure model can be encoded into the latent space, and the diffusion parameter distribution prior of the microstructure parameters in the latent space can be learned by using the latent diffusion model, and the noise latent characteristics of the previous time step are determined based on the diffusion parameter distribution prior, and the noise latent characteristics of the previous time step are denoised to obtain the noise latent characteristics of the current time step, and the noise latent characteristics of the current time step are optimized based on the actually collected diffusion magnetic resonance data. In the process of obtaining the noise latent characteristics and denoising, the prior knowledge learned by the latent diffusion model plays a role, so that the latent diffusion model can effectively identify and suppress noise, and can stably process data even under low signal-to-noise ratio conditions, reducing the impact of noise on parameter estimation. The noise-free latent features are decoded by a decoder, and synthetic data are generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized according to the synthetic data and the diffusion magnetic resonance data. The latent features are continuously adjusted through multiple denoising and optimization processes, so that the synthetic data finally generated by decoding are more in line with the actual situation, which greatly improves the accuracy of parameter estimation of various microstructure models. Moreover, since the latent diffusion model learns a universal diffusion parameter distribution prior and does not rely on a specific acquisition configuration, it can accurately determine the noise latent features and perform subsequent processing when facing different data, maintain high-quality parameter estimation, and improve generalization and versatility.

[0203] Next, a microstructure model optimization device based on diffusion prior proposed in accordance with an embodiment of the present application will be described with reference to the accompanying drawings.

[0204] Figure 7 4 is a block diagram of a microstructure model optimization device based on diffusion prior according to an embodiment of the present application.

[0205] like Figure 7 As shown, the microstructure model optimization device 10 based on diffusion prior includes: an encoding module 100 , a learning module 200 , a determination module 300 and an optimization module 400 .

[0206] Among them, the encoding module 100 is used to encode the microstructure parameters of the microstructure model into the latent space through a vector quantization autoencoder; the learning module 200 is used to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space using a latent diffusion model; the determination module 300 is used to determine the noise latent characteristics of the previous time step based on the diffusion parameter distribution prior, denoise the noise latent characteristics of the previous time step to obtain the noise latent characteristics of the current time step, and optimize the noise latent characteristics of the current time step based on the actually collected diffusion magnetic resonance data; the optimization module 400 is used to denoise the optimized noise latent characteristics to obtain noise-free latent characteristics, decode the noise-free latent characteristics through a decoder, generate synthetic data using the microstructure model and the decoded parameters, and optimize the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data.

[0207] In an embodiment of the present application, the vector quantization autoencoder includes multiple downsampling layers and residual blocks.

[0208] In the embodiment of the present application, the encoding module 100 is further configured to: segment the microstructure parameters into data blocks through multiple downsampling layers; and map the data blocks to a latent space through residual blocks.

[0209] In the embodiment of the present application, the encoding formula of the vector quantization autoencoder is:

[0210]

[0211] z=En(x);

[0212] Where En(·) is the encoder neural network; x is the microstructure parameter; De(·) is the decoder neural network; E x is the expected operation; z is the potential feature representation.

[0213] In the embodiment of the present application, the potential diffusion model is a U-shaped network structure with multi-layer downsampling and self-attention modules, wherein the U-shaped network structure obtains the diffusion parameter distribution prior by approximating the potential distribution through denoising score matching. The formula of the U-shaped network structure is:

[0214]

[0215] Among them, s θ is the attention U-shaped network; z t is the potential feature of time step t; ∈ is the noise sampled from the standard normal distribution N(0,1); is the model parameter of the U-shaped network; E En(x),∈～N(0,1),t is the expectation operation.

[0216] In the embodiment of the present application, the device 10 of the embodiment of the present application further includes: a training module.

[0217] Among them, the training module is used to obtain the minimized mean square error between the predicted noise and the actual added noise before using the latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space; the latent diffusion model is trained using the mean square error, and multi-scale feature extraction and jump connection mechanism are adopted during the training process.

[0218] In the embodiment of the present application, the formula for generating the noise potential feature at the current time step is:

[0219]

[0220] Among them, z′ t is the potential feature of temporary noise at the current time step; z t+1 is the noise potential feature of the previous time step; is a predefined noise scheduling parameter; For z t+1 Noise-free estimate of s θ (z t+1 ,t+1) is the noise prediction calculated by the attention U-network; η is the step size parameter that controls the intensity of random sampling; δ t+1 is the time-step-dependent noise coefficient; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0221] In the embodiment of the present application, the optimization formula for the noise potential feature at the current time step is:

[0222]

[0223] Among them, z t is the noise potential feature of the current time step; z′ t is the potential feature of temporary noise at the current time step; ζ is the learning rate of gradient descent, which controls the optimization step size; For z t ′ is the gradient operator; S is the actual diffusion magnetic resonance data collected; De(·) is the decoder neural network of the potential diffusion model; Modeling(·) is the forward microstructure model used to generate synthetic data from the decoded parameters; is the mean square error.

[0224] In the embodiment of the present application, the optimization module 400 is further configured to input the synthetic data and the diffusion magnetic resonance data into a loss function, and optimize the microstructure parameters of the microstructure model using the loss function, wherein the loss function is:

[0225]

[0226] Where S is the actual diffusion magnetic resonance data collected; z0 is the Gaussian noise z TInitially, the noise-free latent features are obtained after T-step iterative denoising; Modeling(·) is the forward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by inputting the diffusion parameters into the microstructure model after mapping the noise-free latent features into the diffusion parameters through the decoder.

[0227] It should be noted that the above explanation of the embodiment of the microstructure model optimization method based on diffusion prior is also applicable to the microstructure model optimization device based on diffusion prior in this embodiment, and will not be repeated here.

[0228] According to the microstructure model optimization device based on diffusion prior proposed in the embodiment of the present application, the microstructure parameters of the microstructure model can be encoded into the latent space, and the diffusion parameter distribution prior of the microstructure parameters in the latent space can be learned by using the latent diffusion model, and the noise latent characteristics of the previous time step are determined based on the diffusion parameter distribution prior, and the noise latent characteristics of the previous time step are denoised to obtain the noise latent characteristics of the current time step. The noise latent characteristics of the current time step are optimized based on the actually collected diffusion magnetic resonance data. In the process of obtaining the noise latent characteristics and denoising, the prior knowledge learned by the latent diffusion model plays a role, so that the latent diffusion model can effectively identify and suppress noise. Even under low signal-to-noise ratio conditions, it can stably process data and reduce the impact of noise on parameter estimation. The noise-free latent features are decoded by a decoder, and synthetic data are generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized according to the synthetic data and the diffusion magnetic resonance data. The latent features are continuously adjusted through multiple denoising and optimization processes, so that the synthetic data finally generated by decoding are more in line with the actual situation, which greatly improves the accuracy of parameter estimation of various microstructure models. Moreover, since the latent diffusion model learns a universal diffusion parameter distribution prior and does not rely on a specific acquisition configuration, it can accurately determine the noise latent features and perform subsequent processing when facing different data, maintain high-quality parameter estimation, and improve generalization and versatility.

[0229] Figure 8 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:

[0230] A memory 801 , a processor 802 , and a computer program stored in the memory 801 and executable on the processor 802 .

[0231] When the processor 802 executes the program, the microstructure model optimization method based on diffusion prior provided in the above embodiment is implemented.

[0232] Furthermore, the electronic device further includes:

[0233] The communication interface 803 is used for communication between the memory 801 and the processor 802 .

[0234] The memory 801 is used to store computer programs that can be run on the processor 802.

[0235] The memory 801 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0236] If the memory 801, the processor 802, and the communication interface 803 are implemented independently, the communication interface 803, the memory 801, and the processor 802 can be connected to each other via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 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.

[0237] Optionally, in a specific implementation, if the memory 801, the processor 802 and the communication interface 803 are integrated on a chip, the memory 801, the processor 802 and the communication interface 803 can communicate with each other through an internal interface.

[0238] The processor 802 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0239] 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 microstructure model optimization method based on diffusion prior is implemented.

[0240] An embodiment of the present application further provides a computer program product, including a computer program or instructions, which, when executed, implements the above-mentioned microstructure model optimization method based on diffusion prior.

[0241] 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.

[0242] 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.

[0243] 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.

[0244] 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 embodiment, the N 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, it can be implemented using any one or a combination of the following technologies known in the art: 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.

[0245] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

Claims

1. A microstructure model optimization method based on diffusion prior, characterized in that: The following steps are involved: The microstructure parameters of the microstructure model are encoded into the latent space via a vector quantized autoencoder; A latent diffusion model is used to learn a diffusion parameter distribution prior of microstructure parameters in the latent space; determining a noise potential feature of a previous time step based on a diffusion parameter distribution prior, denoising the noise potential feature of the previous time step to obtain a noise potential feature of a current time step, and optimizing the noise potential feature of the current time step based on actually acquired diffusion magnetic resonance data; The optimized noise latent features are denoised to obtain noise-free latent features, the noise-free latent features are decoded by a decoder, synthetic data are generated using the microstructure model and the decoded parameters, and microstructure parameters of the microstructure model are optimized based on the synthetic data and the diffusion magnetic resonance data.

2. The microstructure model optimization method based on diffusion prior according to claim 1, characterized in that: The vector quantization autoencoder includes multiple downsampling layers and residual blocks, and encoding the microstructure parameters of the microstructure model into the latent space through the vector quantization autoencoder includes: dividing the microstructure parameters into data blocks through the multiple downsampling layers; The data block is mapped to the latent space via the residual block.

3. The microstructure model optimization method based on diffusion prior according to claim 1 or 2, characterized in that: The encoding formula of the vector quantization autoencoder is: z=En(x); Where En(·) is the encoder neural network; x is the microstructure parameter; De(·) is the decoder neural network; E x is the expected operation; z is the potential feature representation.

4. The microstructure model optimization method based on diffusion prior according to claim 1, characterized in that: The potential diffusion model is a U-shaped network structure with multi-layer downsampling and self-attention modules, wherein the U-shaped network structure obtains the diffusion parameter distribution prior by approximating the potential distribution through denoising score matching. The formula of the U-shaped network structure is: Among them, s θ is the attention U-shaped network; z t is the potential feature of time step t; ∈ is the noise sampled from the standard normal distribution N(0,1); is the model parameter of the U-shaped network; E En(x),∈～N(0,1),t is the expectation operation.

5. The microstructure model optimization method based on diffusion prior according to claim 1 or 4, characterized in that: Before adopting the latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space, the method further includes: Obtain the minimized mean square error between the predicted noise and the actual added noise; The latent diffusion model is trained using the mean square error, and during the training process, multi-scale feature extraction and a skip connection mechanism are adopted.

6. The microstructure model optimization method based on diffusion prior according to claim 1, characterized in that: The formula for generating the noise potential feature of the current time step is: Among them, z t ′ is the potential feature of temporary noise at the current time step; z t+1 is the noise potential feature of the previous time step; is a predefined noise scheduling parameter; For z t+1 Noise-free estimate of s θ (z t+1 ,t+1) is the noise prediction calculated by the attention U-network; η is the step size parameter that controls the intensity of random sampling; δ t+1 is the time-step-dependent noise coefficient; ∈ is the random noise sampled from the standard normal distribution N(0,1).

7. The microstructure model optimization method based on diffusion prior according to claim 6, characterized in that: The optimization formula for the noise potential feature of the current time step is: Among them, z t is the noise potential feature of the current time step; z t ′ is the potential feature of temporary noise at the current time step; ζ is the learning rate of gradient descent, which controls the optimization step size; For z′ t The gradient operator is ; S is the actual diffusion magnetic resonance data collected; De(·) is the decoder neural network of the potential diffusion model; Modeling(·) is the forward microstructure model used to generate synthetic data from the decoded parameters; is the mean square error.

8. The microstructure model optimization method based on diffusion prior according to claim 1, characterized in that: Optimizing the microstructure parameters of the microstructure model according to the synthetic data and the diffusion magnetic resonance data includes: The synthetic data and the diffusion magnetic resonance data are input into a loss function, and the microstructure parameters of the microstructure model are optimized using the loss function, wherein the loss function is: Where S is the actual diffusion magnetic resonance data collected; z0 is the Gaussian noise z T Initially, the noise-free latent features are obtained after T-step iterative denoising; Modeling(·) is the forward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by inputting the diffusion parameters into the microstructure model after mapping the noise-free latent features into the diffusion parameters through the decoder.

9. A microstructure model optimization device based on diffusion prior, characterized in that: include: An encoding module for encoding the microstructure parameters of the microstructure model into a latent space via a vector quantized autoencoder; A learning module for learning a diffusion parameter distribution prior of microstructure parameters in the latent space using a latent diffusion model; a determination module, configured to determine a noise potential feature of a previous time step based on a priori diffusion parameter distribution, denoise the noise potential feature of the previous time step to obtain a noise potential feature of a current time step, and optimize the noise potential feature of the current time step based on actually acquired diffusion magnetic resonance data; An optimization module is used to denoise the optimized noise latent features to obtain noise-free latent features, decode the noise-free latent features through a decoder, generate synthetic data using the microstructure model and the decoded parameters, and optimize the microstructure parameters of the microstructure model based on the synthetic data and the diffusion magnetic resonance data.

10. 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 microstructure model optimization method based on diffusion prior according to any one of claims 1 to 8.

Images