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

Abstract

The application 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, which comprises the following steps: encoding microstructure parameters of a microstructure model into a latent space through a vector quantization autoencoder; learning a diffusion parameter distribution prior of the microstructure parameters in the latent space by using a latent diffusion model; determining noise latent features of a previous time step according to the diffusion parameter distribution prior, denoising the noise latent features of the previous time step to obtain noise latent features of a current time step, and optimizing the noise latent features of the current time step according to actually collected diffusion magnetic resonance data; denoising the optimized noise latent features to obtain noise-free latent features, generating synthetic data by using a microstructure model and decoded noise-free latent feature parameters, and optimizing the microstructure parameters of the microstructure model according to the synthetic data and the diffusion magnetic resonance data. Thus, the problems of insufficient generalization ability and poor universality of related technologies are solved.

Description

TECHNICAL FIELD

[0001] The present application 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. BACKGROUND

[0002] Diffusion magnetic resonance imaging is an important medical imaging technology, widely used in modeling and mapping of brain tissue microstructure. This technology measures the diffusion motion of water molecules in tissue, providing important information about the microstructure of the tissue, and has important value in neuroscience research and clinical diagnosis.

[0003] Related technologies use deep learning-based techniques, especially neural network-based methods, which have shown superior performance in generating high-quality microstructure model parameters. These methods usually train neural networks to directly learn the mapping relationship from diffusion-weighted images acquired under specific acquisition configurations to microstructure parameters, reducing the cumbersome mathematical calculations in traditional methods. These neural networks include multilayer perceptrons, transformer networks, recurrent neural networks, etc. Some algorithms also unify the data obtained under specific acquisition configurations into a standard data format to train the neural network.

[0004] However, the deep learning methods of related technologies are mainly trained under specific acquisition configurations, with insufficient generalization ability, and some methods have strong model specificity, narrow application range, and are sensitive to noise, resulting in significant performance degradation under low signal-to-noise ratio conditions. SUMMARY

[0005] The present application provides a microstructure model optimization method and device based on diffusion prior to solve the problems of insufficient generalization ability and poor universality 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 the latent space by a vector quantization autoencoder; learning the diffusion parameter distribution prior of the microstructure parameters in the latent space using a latent diffusion model; determining the noise latent feature of the previous time step according to 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 according to the actually acquired 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 according to the synthetic data and the diffusion magnetic resonance data.

[0007] Optionally, the vector quantization autoencoder comprises a plurality of down-sampling layers and a residual block, and encoding the microstructure parameters of the microstructure model into the latent space by the vector quantization autoencoder comprises: segmenting the microstructure parameters into data blocks by the plurality of down-sampling layers; and mapping the data blocks to the latent space by the residual block.

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

[0009]

[0010] z=En(x);

[0011] wherein En(·) is an encoder neural network; x is the microstructure parameters; De(·) is a decoder neural network; E x is an expectation operation; and z is a latent feature representation.

[0012] Optionally, the latent diffusion model is a U-shaped network structure with a plurality of down-sampling layers and a self-attention module, wherein the U-shaped network structure obtains a diffusion parameter distribution prior by denoising score matching approximation of a latent distribution, and a formula of the U-shaped network structure is:

[0013]

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

[0015] Optionally, before learning the diffusion parameter distribution prior of the microstructure parameters in the latent space by the latent diffusion model, the method further comprises: obtaining a minimum mean square error between a predicted noise and an actual added noise; and training the latent diffusion model by using the mean square error, wherein in the training process, a multi-scale feature extraction and a skip connection mechanism are adopted.

[0016] Optionally, a generation formula of the noise latent feature at the current time step is:

[0017]

[0018] wherein z′ t is a temporary noise latent feature at the current time step; z t+1 is a noise latent feature at a previous time step; is a predefined noise scheduling parameter; is a noise-free estimation of z t+1 ; s θ (z t+1,t+1) represents the noise prediction computed using an attention U-shaped network; η is the step size parameter, controlling the intensity of random sampling; δ t+1 is the noise coefficient related to the time step; ∈ is the random noise sampled from the standard normal distribution N(0,1).

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

[0020]

[0021] Among them, z t The noise potential feature at the current time step; z′ t ζ represents the temporary noise potential features 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; S is the actual acquired diffuse magnetic resonance data; De(·) is the decoder neural network of the latent diffusion model; Modeling(·) is the feedforward microstructure model used to generate synthetic data from the decoded parameters; This represents the mean square error.

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

[0023]

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

[0025] A second aspect of this application provides a microstructure model optimization device based on diffusion prior, comprising: an encoding module for encoding microstructure parameters of a microstructure model into a latent space using 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 features of the previous time step based on the diffusion parameter distribution prior, denoising the noise latent features of the previous time step to obtain the noise latent features of the current time step, and optimizing the noise latent features of the current time step based on actually acquired diffuse magnetic resonance data; and an optimization module for denoising the optimized noise latent features to obtain noise-free latent features, decoding the noise-free latent features using 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 diffuse magnetic resonance data.

[0026] Optionally, the vector quantization autoencoder includes multiple downsampling layers and residual blocks, and the encoding module is further used to: segment the microstructure parameters into data blocks through 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 expectation operation; z is the latent feature representation.

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

[0032]

[0033] Among them, s θ For attention-based U-shaped networks; z t represents the latent feature at time step t; ∈ represents the noise sampled from the standard normal distribution N(0,1); E represents the model parameters of the U-shaped network. En(x),∈～N(0,1),t This is the expected operation.

[0034] Optionally, it also includes: a training module for obtaining the minimum mean square error between the predicted noise and the actual added noise before learning the prior distribution of diffusion parameters of microstructure parameters in the latent space using the latent diffusion model; training the latent diffusion model using the mean square error, employing multi-scale feature extraction and skip connection mechanisms during the training process.

[0035] Optionally, the formula for generating the noise latent features at the current time step is:

[0036]

[0037] Where, z′ t For the current time step, z represents the potential temporary noise feature; t+1 This represents the potential noise characteristics of the previous time step; For predefined noise scheduling parameters; For z t+1 The noise-free estimate; s θ (z t+1 ,t+1) represents the noise prediction computed using an attention U-shaped network; η is the step size parameter, controlling the intensity of random sampling; δ t+1 is the noise coefficient related to the time step; ∈ is the random noise sampled from the standard normal distribution N(0,1).

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

[0039]

[0040] Among them, z t The noise potential feature at the current time step; z′ t ζ represents the temporary noise potential features 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; S is the actual acquired diffuse magnetic resonance data; De(·) is the decoder neural network of the latent diffusion model; Modeling(·) is the feedforward microstructure model used to generate synthetic data from the decoded parameters; This represents the mean square error.

[0041] Optionally, the optimization module is further configured to: input the synthetic data and diffuse 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 represents the actual acquired diffuse magnetic resonance data; z0 represents the Gaussian noise z TInitially, noise-free latent features are obtained after T iterations of denoising; Modeling(·) is the feedforward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by mapping the noise-free latent features to diffusion parameters through the decoder and then inputting the diffusion parameters into the microstructure model.

[0044] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to perform the diffusion-prior-based microstructure model optimization method as described in the above embodiments.

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

[0046] This application embodiment encodes the microstructure parameters of a microstructure model into a latent space, and uses a latent diffusion model to learn the prior distribution of diffusion parameters of the microstructure parameters in the latent space. Based on the prior distribution of diffusion parameters, it determines the noise latent features of the previous time step, denoises the noise latent features of the previous time step to obtain the noise latent features of the current time step, and optimizes the noise latent features of the current time step based on the actual acquired diffusion magnetic resonance data. In the process of acquiring noise latent features and denoising, the prior knowledge learned by the latent diffusion model plays a role, enabling the latent diffusion model to effectively identify and suppress noise. Even under low signal-to-noise ratio conditions, it can stably process data and reduce the interference of noise on parameter estimation results. Then, the optimized noise latent features are denoised to obtain noise-free latent features, which are then decoded.

[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 diffuse magnetic resonance (DMR) data. Through multiple denoising and optimization processes, latent features are continuously adjusted, making the final decoded synthetic data more closely resemble reality. This significantly improves the accuracy of parameter estimation for various microstructure models. Furthermore, because the latent diffusion model learns a general prior distribution of diffusion parameters and does not depend on specific acquisition configurations, it can accurately identify and process latent noise features when faced with different data, maintaining high-quality parameter estimation and improving generalization and versatility. This solves the technical problems of insufficient generalization and poor versatility in related technologies.

[0048] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0049] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0050] Figure 1 This is a flowchart of a microstructure model optimization method based on diffusion priors provided in an embodiment of this application;

[0051] Figure 2 This is a structural diagram of the microstructure model optimization method based on diffusion prior provided in the embodiments of this application;

[0052] Figure 3 This is a schematic diagram comparing the effects of tensor model parameter estimation according to the embodiments of this application;

[0053] Figure 4 This is a schematic diagram comparing the first effect in kurtosis model parameter estimation according to the embodiments of this application;

[0054] Figure 5 This is a schematic diagram comparing the second effect in kurtosis model parameter estimation according to the embodiments of this application;

[0055] Figure 6 This is a schematic diagram comparing the performance of the NODDI model parameter estimation according to the embodiments of this application;

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

[0057] Figure 8 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0058] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0059] To aid in understanding the solution of this application, the relevant technologies will be described first.

[0060] Deep learning-based techniques, particularly those utilizing neural networks, have demonstrated superior performance in generating high-quality microstructure model parameters. These methods typically train neural networks to directly learn the mapping relationship between diffusion-weighted images acquired in a specific acquisition configuration and microstructure parameters, reducing the cumbersome mathematical calculations required in traditional methods. These neural networks include multilayer perceptrons, transformer networks, and recurrent neural networks. Some algorithms address the issue of data from different acquisition configurations not generalizing by standardizing the data in a specific acquisition configuration into a standard data format before training the neural network. For example, depth tensor imaging networks use tensor models to convert arbitrary data into diffusion-weighted images with the same acquisition parameters. Furthermore, some research attempts to transform diffusion data into a spherically harmonic representation space to improve the model's applicability.

[0061] However, the relevant technologies have the following problems:

[0062] Insufficient generalization ability: Most neural network-based methods are typically trained for specific acquisition configurations (b-value and diffusion weighting direction), and cannot handle data acquired using other configurations. This severely limits the application of these methods on data with different acquisition configurations.

[0063] Model specificity: Although the depth tensor imaging network solves some of the generalization problems, its design is only applicable to tensor models and cannot be extended to other microstructure models.

[0064] Poor handling of data sparsity: Related methods perform poorly when processing sparsely sampled q-space data. Some methods that attempt to transform diffused data into a common space often fail when faced with very sparse samples.

[0065] High retraining costs: Existing technologies require retraining for each new encoding scheme, increasing time and computational costs in practical applications.

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

[0067] Limitations in clinical applications: Due to the aforementioned limitations, the application of existing technologies in microstructural and connectivity mapping in clinical and neuroscience research is restricted, making it difficult to fully realize the potential of diffusion MRI.

[0068] To address this, this application provides a microstructure model optimization method based on diffusion prior. In this method, the microstructure parameters of the microstructure model are encoded into a latent space. A latent diffusion model is then used to learn the prior distribution of diffusion parameters within the latent space. Based on this prior distribution, the noise latent features of the previous time step are determined. These features are then denoised to obtain the noise latent features of the current time step. Finally, the noise latent features of the current time step are optimized based on the actual acquired diffuse magnetic resonance (DMR) data. During the acquisition and denoising processes, the prior knowledge learned by the latent diffusion model plays a crucial role, enabling the model to effectively identify and suppress noise. Even under low signal-to-noise ratio (SNR) conditions, the model can stably process data and reduce the impact of noise on the parameters. The noise estimation results are distorted, and the optimized noise latent features are then denoised to obtain noise-free latent features. These noise-free latent features are decoded using a decoder, and synthetic data is generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized based on the synthetic data and diffuse magnetic resonance data. Through multiple denoising and optimization processes, the latent features are continuously adjusted so that the final decoded synthetic data more closely reflects the real situation. This significantly improves the accuracy of parameter estimation for various microstructure models. Furthermore, since the latent diffusion model learns a general prior distribution of diffusion parameters and does not depend on a specific acquisition configuration, it can accurately determine noise latent features and perform subsequent processing when faced with different data, maintaining high-quality parameter estimation and improving generalization and versatility.

[0069] Specifically, Figure 1 This is a flowchart illustrating a microstructure model optimization method based on diffusion prior, provided in an embodiment of this 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 the latent space using a vector quantization autoencoder.

[0072] The vector quantization autoencoder includes multiple downsampling layers and residual blocks.

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

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

[0075] In this embodiment of the 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 expectation operation; z is the latent feature representation.

[0079] Specifically, this application includes the following steps for encoding microstructure model parameters into a latent space using a vector quantization autoencoder:

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

[0081] (2) Divide the input microstructure model parameter volume data into data blocks of a specific size;

[0082] (3) The data blocks are mapped to discrete latent codebook spaces by the encoder;

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

[0084] In step S102, a priori distribution of diffusion parameters of microstructure parameters in the potential space is learned using a potential diffusion model.

[0085] Among them, the potential diffusion model includes at least one of the tensor model, kurtosis model and NODDI model.

[0086] It is understood that the embodiments of this application may employ a latent diffusion model to learn the prior distribution of diffusion parameters of microstructure parameters in the latent space, so as to perform subsequent posterior sampling.

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

[0088]

[0089] Among them, s θ For attention-based U-shaped networks; z t represents the latent feature at time step t; ∈ represents the noise sampled from the standard normal distribution N(0,1); E represents the model parameters of the U-shaped network. En(x),∈～N(0,1),t This is the expected operation.

[0090] In this embodiment of the application, before learning the prior distribution of diffusion parameters of microstructure parameters in the latent space using the latent diffusion model, the method further includes: obtaining the mean square error between the predicted noise and the actual added noise that is minimized; training the latent diffusion model using the mean square error, and employing multi-scale feature extraction and skip connection mechanisms during the training process.

[0091] It is understood that, in the embodiments of this application, before learning the prior distribution of diffusion parameters of microstructure parameters in the latent space using the latent diffusion model, the latent diffusion model is trained. The specific training process is as follows: training is performed by minimizing the mean square error between the predicted noise and the actual added noise, and multi-scale feature extraction and skip connection mechanisms are used during the training process.

[0092] In step S103, the noise potential features of the previous time step are determined a priori based on the diffusion parameter distribution. The noise potential features of the previous time step are then denoised to obtain the noise potential features of the current time step. Finally, the noise potential features of the current time step are optimized based on the actual collected diffuse magnetic resonance data.

[0093] It is understood that the embodiments of this application can obtain the potential noise features of the current time step based on the prior distribution of diffusion parameters, and optimize the potential noise features of the current time step based on the actual collected diffuse magnetic resonance data. By continuously optimizing the potential features to make them closer to the actual data, and combining the learning ability of the potential diffusion model on data features, sparse sampled data can be effectively processed.

[0094] In this embodiment, the formula for generating the noise latent features at the current time step is:

[0095]

[0096] Where, z′ t For the current time step, z represents the potential temporary noise feature; t+1 This represents the potential noise characteristics of the previous time step; For predefined noise scheduling parameters; For z t+1 The noise-free estimate; s θ (z t+1 ,t+1) represents the noise prediction computed using an attention U-shaped network; η is the step size parameter, controlling the intensity of random sampling; δ t+1 is the noise coefficient related to the time step; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0097] In this embodiment, the optimization formula for the noise latent features at the current time step is:

[0098]

[0099] Among them, zt The noise potential feature at the current time step; z′ t ζ represents the temporary noise potential features 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; S is the actual acquired diffuse magnetic resonance data; De(·) is the decoder neural network of the latent diffusion model; Modeling(·) is the feedforward microstructure model used to generate synthetic data from the decoded parameters; This represents the mean square error.

[0100] Specifically, this application performs an iterative process from high to low for each timestamp t, generating latent features based on conditional probability, i.e., an iterative process decreasing from 500 to 0 for timestamp t. At each time step t, the system generates latent features z from the previous time step through the following steps. 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 is generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized based on the synthetic data and the diffuse magnetic resonance data.

[0102] It is understood that the embodiments of this application can denoise the optimized noise latent features to obtain noise-free latent features, and decode the noise-free latent features through a decoder. Synthetic data is generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized based on the synthetic data and diffuse magnetic resonance data. The optimization of the microstructure parameters of the microstructure model is achieved through high-quality data, and the robustness of noise processing is improved by matching the latent distribution with the denoising score.

[0103] In this embodiment of the application, optimizing the microstructure parameters of a microstructure model based on synthetic data and diffuse magnetic resonance (DMR) data includes: inputting the synthetic data and DMR data into a loss function, and using the loss function to optimize the microstructure parameters of the microstructure model, wherein the loss function is:

[0104]

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

[0106] The following specific embodiment describes the diffusion-prior-based microstructure model optimization method of this application, which mainly includes:

[0107] I. Prior Learning Module.

[0108] In the training phase of this application, the system first trains an 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 The latent distribution is approximated by matching the denoised scores:

[0112]

[0113] Among them, z t Represents the latent characteristics of time step t; ∈ represents noise sampled from a standard normal distribution.

[0114] II. Posterior sampling module.

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

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

[0117]

[0118] in, The progress of the denoising process is controlled by predefined noise scheduling parameters; For z t+1 The noise-free estimate; s θ (z t+1 ,t+1) represents the noise prediction computed using an attention U-shaped network; η is the step size parameter, controlling the intensity of random sampling; δ t+1is the noise coefficient related to the time step; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0119] To ensure the potential estimate z t ′ Consistency between the data and the actual acquired diffusion magnetic resonance (DMR) data S was optimized using the mean square error. t ′, as shown in the following formula:

[0120]

[0121] Where ζ is the learning rate of gradient descent, which controls the optimization step size; Indicates about z t The gradient operator of ′; S is the actual acquired diffuse magnetic resonance data; De(·) is the decoder neural network of the latent diffusion model; Modeling(·) represents the forward microstructure model, used to generate synthetic data from the decoded parameters. This represents the square of the L2 norm, i.e., the mean square error.

[0122] During the calculation process, it is necessary to estimate z′. t Noiseless version The calculation method is as follows:

[0123]

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

[0125] III. Physical optimization module.

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

[0127]

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

[0129] IV. 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 the value of b, both of which are determined by the acquisition protocol. D and S0 represent the tensor parameters estimated in this application.

[0133] The kurtosis model is implemented as follows:

[0134]

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

[0136] The NODDI model is implemented 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 These represent the diffusion signals of the intracellular compartment, extracellular compartment, and isotropic compartment, respectively; f iso ,f ic These represent isotropic volume fraction and intracellular volume fraction, respectively.

[0139] In summary, the microstructure model optimization method based on diffusion priors in this application includes the following steps:

[0140] 1. Encode the parameters of the microstructure model into the latent space using a vector quantization autoencoder;

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

[0142] Construct a three-dimensional vector quantization autoencoder with multiple downsampling layers and residual blocks;

[0143] The input microstructure model parameter volume data is divided into data blocks of a specific size;

[0144] The encoder maps data blocks to a discrete potential codebook space.

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

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

[0147] The steps involved in training the attention U-shaped network to approximate the latent distribution by matching the denoised scores include:

[0148] A U-shaped network structure with multi-layer downsampling and self-attention modules is adopted;

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

[0150] Multi-scale feature extraction and skip connection mechanisms are employed during training.

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

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

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

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

[0155] The microstructure model optimization method based on diffusion priors proposed in this application is described in a systematic manner below, mainly including 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 prior distributions of diffusion parameters through a latent diffusion model;

[0158] 3. A subject-specific posterior sampling module is used to estimate the parametric map from the collected data based on conditional probability;

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

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

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

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

[0163] Furthermore, it should be noted that all neural network models in this application are implemented based on the PyTorch platform. As a dynamic neural network framework, PyTorch provides flexible computational graph construction capabilities and a rich deep learning toolkit, making it 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, which has the following characteristics: (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 the parameter space; (3) Downsampling layers: 3 layers, progressively compressing the input data into the latent space; (4) Residual block configuration: containing 2 residual blocks at each scale, enhancing feature extraction capability and training stability. This autoencoder is responsible for mapping the microstructure model parameters to the latent space, providing high-quality feature representations for subsequent diffusion models.

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

[0166] The training process adopted the following configuration: (1) Data partitioning and augmentation: Each input data was randomly partitioned into 75 blocks of size 64×64×64×channel; the blocks were flipped along the left and right anatomical axes to achieve data augmentation and improve the generalization ability of the model. (2) Optimizer configuration: Adam optimizer was used; autoencoder learning rate: 5×10 -5 Latent spatial denoiser learning rate: 2.5 × 10⁻⁶ -5 Batch size: 12.

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

[0168] Example 1: Diffusion magnetic resonance imaging (DMRI) data from 83 subjects in the Human Connectome Project (HCP) was used as the training and testing dataset. The data used had an isotropic spatial resolution of 1.25 mm and underwent preprocessing 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) data distributed across three diffusion shells (b=1000, 2000, and 3000 s mm). -2 Data for 90 individuals in each shell layer.

[0169] To establish reference microstructure model parameter values ​​during training and evaluation, this embodiment uses the following method to calculate reference parameters for different microstructure models. Tensor model parameters: Ordinary least squares regression implemented using the FSL software package, based on all b=0 volume data and b=1000 s mm. -2 Single-shell configuration calculation; kurtosis model parameters: Ordinary least squares regression method implemented using the MRtrix3 software package, based on all volume data with b=0 and b=1000, 2000 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 Three-shell configuration calculation.

[0170] Example 2: The CHCP (Chinese Human Connectome Project) was used as the out-of-distribution test dataset. This dataset has the following characteristics: (1) Spatial resolution: 1.5 mm isotropic, which differs from the 1.25 mm resolution of the training dataset; (2) Number of subjects: 10 subjects, preprocessed diffusion magnetic resonance imaging data; (3) Data composition: Each acquisition includes: 14 b=0 volume data (baseline image), 93 b=1000s mm-2 diffusion-weighted volume data; 92 b=2000s mm-2 diffusion-weighted volume data.

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

[0172] The evaluation metrics used in the testing process are as follows: Tensor metrics include axial dispersion coefficient (AD), mean dispersion coefficient (MD), structural anisotropy (FA), radial diffusion coefficient, and principal eigenvector (V1). Kurtosis model metrics include axial dispersion coefficient (AD), mean dispersion coefficient (MD), fractional anisotropy (FA), radial kurtosis coefficient (AK), mean kurtosis coefficient (MK), and radial kurtosis coefficient (RK). NODDI model metrics 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 this application will be described below with reference to specific figures and tables.

[0174] Figure 2 This is a structural diagram of the diffusion prior-based microstructure model optimization method of this application, where (a) represents the training phase of the latent diffusion model (LDM). In this phase, the diffusion model parameters are encoded into the latent space using a vector quantization autoencoder, and an attention U-shaped network is trained to learn the prior distribution of the diffusion model parameters. Figure 2 As shown, the autoencoder consists of an encoder and a decoder, mapping the high-dimensional parameter space to the low-dimensional latent space, while the U-shaped network approximates the latent distribution through denoising score matching. (b) represents the iterative optimization process during the inference stage. In this stage, this application performs iterations from high to low for each timestamp, continuously reducing the difference between the currently estimated synthetic data and the actual collected data through a posterior sampling strategy. Figure 2 As shown, this process uses conditional probability to generate latent features and optimizes them using gradient descent to better match the actual acquired diffuse magnetic resonance (DMR) data. (c) represents the final decoder fine-tuning stage, where this application further reduces the difference between the estimation results and the actual data through a physical information parameter tuning mechanism. Figure 2 As shown, this process effectively avoids the domain shift problem and improves the accuracy of parameter estimation by fine-tuning the decoder of the potential diffusion model by minimizing a specific loss function.

[0175] Figure 3 A graphical comparison of the performance of a single subject in tensor model parameter estimation using the Human Connectome Project dataset is presented. The graphs showing the tensor model metric results (a, c, e, g) estimated using this application and other comparative methods on a representative subject, along with the residual graphs (b, d, f, h) compared to the reference graph, are displayed. The mean absolute error of each graph compared to the reference graph is shown in the lower right corner of the residual graph.

[0176] Table 1 compares the mean absolute errors in tensor model DTI parameter estimation for 10 subjects in the Human Connectome Project dataset. It shows the mean ± standard deviation of the mean absolute errors of the tensor model metrics estimated using this application and other comparative methods compared to reference values. The last column of Table 1 highlights the lowest mean absolute errors.

[0177] Table 1

[0178] i ii iii Method FSL software Deep tensor network The present application a FA 0.0975±0.0077 0.0362±0.0018 0.0355±0.0024 b MD (pm 2 / ms)]]> 0.0658±0.004 0.0438±0.0038 0.0366±0.0019 c AD (pm 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 A graphical comparison of the performance of a subject in kurtosis model parameter estimation using a human connectome project dataset. The graphs showing the kurtosis model metric results (a, c, e, g) estimated using this application and other comparative methods on a representative subject, along with the residual graphs (b, d, f, h) compared to the reference graph, are presented. The mean absolute error of each graph compared to the reference graph is shown in the lower right corner of the residual graph.

[0180] Table 2 compares the mean absolute errors in kurtosis model parameter estimation for 10 subjects in the Human Connectome Project dataset. It shows the mean ± standard deviation of the mean absolute errors of the kurtosis model metrics estimated using this application and other comparative methods compared to reference values. The last column highlights the lowest mean absolute errors.

[0181] Table 2

[0182]

[0183] Figure 5 A graphical comparison of the results of kurtosis model parameter estimation for one participant in the Chinese Human Connectome Project dataset is presented. The results (a, c) show the kurtosis model metric estimated using this application and other comparative methods on a representative participant, along with residual plots (b, d) comparing the metric plots to the reference. The mean absolute error of each plot compared to the reference plot is shown in the lower right corner of the residual plots.

[0184] Table 3 compares the mean absolute errors in kurtosis model parameter estimation for 10 subjects in the Chinese Human Connectome Project dataset. It shows the mean ± standard deviation of the mean absolute errors of the kurtosis model metrics estimated using this application and other comparative methods compared to reference values. The last column of Table 3 highlights the lowest mean absolute errors.

[0185] Table 3

[0186]

[0187] Figure 6A graphical comparison of the performance of a Chinese participant in the Human Connectome Project dataset in NODDI model parameter estimation is presented. The graphs (a, c, e) showing the NODDI model metric results estimated using this application and other comparative methods on a representative participant, along with residual graphs (b, d, f) comparing the metric results to the reference graph, are shown. The mean absolute error of each graph compared to the reference graph is displayed in the lower right corner of the residual graphs.

[0188] Table 4 compares the mean absolute errors in NODDI model parameter estimation for 10 subjects in the Chinese Human Connectome Project dataset. It shows the mean ± standard deviation of the mean absolute errors of the NODDI model metrics estimated using this application and other comparative methods compared to the reference values. The last column of Table 4 highlights the lowest mean absolute errors.

[0189] Table 4

[0190]

[0191] In summary, this application aims to provide a microstructure model optimization method and system with high generalization and high-quality parameter estimation capabilities by employing a latent diffusion model to learn the prior distribution of diffusion parameters and using posterior sampling to sample parameter maps from the conditional probabilities of given acquisition data. This method can be directly deployed on any dataset, is applicable to various microstructure models, significantly improves the application value of diffusion magnetic resonance imaging (DMRI) in clinical and neuroscience research, can handle DMRI data with multiple spatial resolutions, and adapts to different diffusion coding schemes. After loading pre-trained weights, the network does not need to be retrained for data with specific acquisition configurations, making it a general network architecture that supports parameter estimation for all microstructure models without the need for customized tuning for specific models.

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

[0193] Improved Parameter Estimation Accuracy: Compared with traditional methods, this application demonstrates higher accuracy in parameter estimation for various microstructure models. In the tensor model, the average error of anisotropy fraction (FA) and mean diffusivity (MD) estimation is reduced by 30-40%; in the kurtosis model, the accuracy of mean kurtosis (MK) estimation is improved by more than 45%.

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

[0195] Advantages of sparse sampling 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 coding schemes: This application demonstrates excellent adaptability to data with different b-value configurations, the number of diffusion directions, and their distribution patterns. In tests on the Chinese Human Connectome Project (CHCP) dataset, despite significant differences in collected parameters compared to the training data, this application maintained high-quality parameter estimates.

[0197] Generalization across spatial resolutions: This application can adapt to diffusion magnetic resonance data with different spatial resolutions, from the 1.25mm iso-directional resolution of training data to the 1.5mm resolution of test data, and even extended to the clinically commonly used 2mm resolution, while maintaining excellent performance.

[0198] Cross-population stability: This application demonstrates good generalization ability on data from different population groups, and the 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 diffuse magnetic resonance data, which can reduce clinical scan time from the traditional 15-20 minutes to 3-5 minutes, while maintaining diagnostic quality and significantly improving the efficiency of clinical workflow.

[0200] Empowering Neuroscience Research: This application provides neuroscience research with a more reliable and consistent tool for mapping the microstructure and connectivity of brain tissue, which helps to deepen the understanding of microstructural changes in neural development, aging and disease processes.

[0201] Multicenter study 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 standard for parameter estimation for multicenter studies and promoting the comparability and reproducibility of research results.

[0202] According to the microstructure model optimization method based on diffusion prior proposed in this application, the microstructure parameters of the microstructure model can be encoded into a latent space. A latent diffusion model is then used to learn the diffusion parameter distribution prior of the microstructure parameters within the latent space. Based on the diffusion parameter distribution prior, the noise latent features of the previous time step are determined. These noise latent features are then denoised to obtain the noise latent features of the current time step. Finally, the noise latent features of the current time step are optimized based on the actually acquired diffuse magnetic resonance (DMR) data. In the process of acquiring and denoising the noise latent features, the prior knowledge learned by the latent diffusion model plays a crucial role, enabling the latent diffusion model to 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 interference of the results is addressed by denoising the optimized noise latent features to obtain noise-free latent features. The noise-free latent features are then decoded using a decoder. Synthetic data is generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized based on the synthetic data and diffusion magnetic resonance data. Through multiple denoising and optimization processes, the latent features are continuously adjusted so that the final decoded synthetic data more closely matches the real situation. This significantly improves the accuracy of parameter estimation for various microstructure models. Furthermore, since the latent diffusion model learns a general prior distribution of diffusion parameters and does not depend on a specific acquisition configuration, it can accurately determine noise latent features and perform subsequent processing when faced with different data, maintaining high-quality parameter estimation and improving generalization and versatility.

[0203] Next, referring to the accompanying drawings, a microstructure model optimization device based on diffusion priors proposed according to an embodiment of this application is described.

[0204] Figure 7 This is a block diagram of a microstructure model optimization device based on diffusion priors according to an embodiment of this application.

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

[0206] The microstructure module 100 encodes the microstructure parameters of the microstructure model into the latent space using a vector quantization autoencoder; the learning module 200 learns the prior distribution of diffusion parameters of the microstructure parameters in the latent space using a latent diffusion model; the determination module 300 determines the noise latent features of the previous time step based on the prior distribution of diffusion parameters, removes noise from the noise latent features of the previous time step to obtain the noise latent features of the current time step, and optimizes the noise latent features of the current time step based on the actual acquired diffuse magnetic resonance data; the optimization module 400 removes noise from the optimized noise latent features to obtain noise-free latent features, decodes the noise-free latent features using a decoder, generates synthetic data using the microstructure model and the decoded parameters, and optimizes the microstructure parameters of the microstructure model based on the synthetic data and the diffuse magnetic resonance data.

[0207] In this embodiment, the vector quantization autoencoder includes multiple downsampling layers and residual blocks.

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

[0209] In this embodiment of the 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 expectation operation; z is the latent feature representation.

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

[0214]

[0215] Among them, s θ For attention-based U-shaped networks; z t represents the latent feature at time step t; ∈ represents the noise sampled from the standard normal distribution N(0,1); E represents the model parameters of the U-shaped network. En(x),∈～N(0,1),t This is the expected operation.

[0216] In this embodiment of the application, the apparatus 10 further includes a training module.

[0217] The training module is used to obtain the mean square error between the predicted noise and the actual added noise before learning the prior distribution of diffusion parameters of microstructure parameters in the latent space using the latent diffusion model. The latent diffusion model is trained using the mean square error, and multi-scale feature extraction and skip connection mechanisms are used during the training process.

[0218] In this embodiment, the formula for generating the noise latent features at the current time step is:

[0219]

[0220] Where, z′ t For the current time step, z represents the potential temporary noise feature; t+1 This represents the potential noise characteristics of the previous time step; For predefined noise scheduling parameters; For z t+1 The noise-free estimate; s θ (z t+1 ,t+1) represents the noise prediction computed using an attention U-shaped network; η is the step size parameter, controlling the intensity of random sampling; δ t+1 is the noise coefficient related to the time step; ∈ is the random noise sampled from the standard normal distribution N(0,1).

[0221] In this embodiment, the optimization formula for the noise latent features at the current time step is:

[0222]

[0223] Among them, z t The noise potential feature at the current time step; z′ t ζ represents the temporary noise potential features 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 of ′; S is the actual acquired diffuse magnetic resonance data; De(·) is the decoder neural network of the latent diffusion model; Modeling(·) is the feedforward microstructure model used to generate synthetic data from the decoded parameters; This represents the mean square error.

[0224] In this embodiment, the optimization module 400 is further configured to: input the synthetic data and diffuse 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 represents the actual acquired diffuse magnetic resonance data; z0 represents the Gaussian noise z TInitially, noise-free latent features are obtained after T iterations of denoising; Modeling(·) is the feedforward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by mapping the noise-free latent features to diffusion parameters through the decoder and then inputting the diffusion parameters into the microstructure model.

[0227] It should be noted that the foregoing explanation of the embodiment of the diffusion prior-based microstructure model optimization method also applies to the diffusion prior-based microstructure model optimization device of this embodiment, and will not be repeated here.

[0228] The microstructure model optimization device based on diffusion prior proposed in this application can encode the microstructure parameters of the microstructure model into a latent space, and use a latent diffusion model to learn the diffusion parameter distribution prior of the microstructure parameters in the latent space. Based on the diffusion parameter distribution prior, the noise latent features of the previous time step are determined, and the noise latent features of the previous time step are denoised to obtain the noise latent features of the current time step. The noise latent features of the current time step are optimized based on the actually acquired diffuse magnetic resonance data. In the process of acquiring noise latent features and denoising, the prior knowledge learned by the latent diffusion model plays a role, enabling the latent diffusion model to 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 interference of the results is addressed by denoising the optimized noise latent features to obtain noise-free latent features. The noise-free latent features are then decoded using a decoder. Synthetic data is generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are optimized based on the synthetic data and diffusion magnetic resonance data. Through multiple denoising and optimization processes, the latent features are continuously adjusted so that the final decoded synthetic data more closely matches the real situation. This significantly improves the accuracy of parameter estimation for various microstructure models. Furthermore, since the latent diffusion model learns a general prior distribution of diffusion parameters and does not depend on a specific acquisition configuration, it can accurately determine noise latent features and perform subsequent processing when faced with different data, maintaining high-quality parameter estimation and improving generalization and versatility.

[0229] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0230] The memory 801, the processor 802, and the computer program stored on the memory 801 and capable of running on the processor 802.

[0231] When the processor 802 executes the program, it implements the microstructure model optimization method based on diffusion prior provided in the above embodiments.

[0232] Furthermore, electronic devices also include:

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

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

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

[0236] If the memory 801, processor 802, and communication interface 803 are implemented independently, then the communication interface 803, memory 801, and processor 802 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 8 The bus is represented by a single thick line, 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, processor 802, and communication interface 803 are integrated on a single chip, then the memory 801, processor 802, and 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 this application.

[0239] This application also provides a computer-readable storage medium storing a computer program or instructions thereon, which, when executed by a processor, implements the above-described microstructure model optimization method based on diffusion priors.

[0240] This application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described diffusion-prior-based microstructure model optimization method.

[0241] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0242] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0243] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0244] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0245] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A microstructure model optimization method based on diffusion prior, characterized in that, Includes the following steps: The microstructure parameters of the microstructure model are encoded into the latent space using a vector quantization autoencoder. The prior distribution of diffusion parameters of microstructure parameters in the potential space is learned using a latent diffusion model; The noise potential features of the previous time step are determined a priori based on the diffusion parameter distribution. The noise potential features of the previous time step are then denoised to obtain the noise potential features of the current time step. The noise potential features of the current time step are then optimized based on the actual collected diffusion magnetic resonance data. The optimized noise latent features are de-noised to obtain noise-free latent features. The noise-free latent features are then decoded using a decoder. Synthetic data is generated using the microstructure model and the decoded parameters. The microstructure parameters of the microstructure model are then optimized based on the synthetic data and the diffuse magnetic resonance data.

2. The microstructure model optimization method based on diffusion prior as described in claim 1, characterized in that, The vector quantization autoencoder includes multiple downsampling layers and residual blocks. Encoding the microstructure parameters of the microstructure model into the latent space using the vector quantization autoencoder includes: The microstructure parameters are divided into data blocks using the multiple downsampling layers; The data block is mapped to the potential space through the residual block.

3. The microstructure model optimization method based on diffusion prior as described in 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 expectation operation; z is the latent feature representation.

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

5. The microstructure model optimization method based on diffusion prior as described in claim 1 or 4, characterized in that, Before learning the prior distribution of diffusion parameters of microstructure parameters within the latent space using a latent diffusion model, the following steps are also included: To obtain the minimum mean square error between the predicted noise and the actual added noise; The latent diffusion model is trained using the mean squared error, employing multi-scale feature extraction and skip connection mechanisms during the training process.

6. The microstructure model optimization method based on diffusion prior as described in claim 1, characterized in that, The formula for generating the noise latent features at the current time step is: Among them, z t ′ represents the potential temporary noise feature at the current time step; z t+1 This represents the potential noise characteristics of the previous time step; For predefined noise scheduling parameters; For z t+1 Noise-free estimation; s θ (z t+1 ,t+1) represents the noise prediction computed using an attention U-shaped network; η is the step size parameter, controlling the intensity of random sampling; δ t+1 is the noise coefficient related to the time step; ∈ is the random noise sampled from the standard normal distribution N(0,1).

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

8. The microstructure model optimization method based on diffusion prior as described in claim 1, characterized in that, The optimization of the microstructure parameters of the microstructure model based on the synthesized data and the diffuse magnetic resonance data includes: The synthesized data and the diffuse 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 represents the actual acquired diffuse magnetic resonance data; z0 represents the Gaussian noise z T Initially, noise-free latent features are obtained after T iterations of denoising; Modeling(·) is the feedforward microstructure model; De(·) is the decoder neural network of the latent diffusion model; Modeling(De(z0)) is the estimated signal value obtained by mapping the noise-free latent features to diffusion parameters through the decoder and then inputting the diffusion parameters into the microstructure model.

9. A microstructure model optimization device based on diffusion prior, characterized in that, include: The encoding module is used to encode the microstructure parameters of the microstructure model into the latent space using a vector quantization autoencoder; The learning module is used to learn the prior distribution of diffusion parameters of microstructure parameters in the potential space using a potential diffusion model; The determination module is used to determine the potential noise features of the previous time step based on the prior distribution of diffusion parameters, remove noise from the potential noise features of the previous time step to obtain the potential noise features of the current time step, and optimize the potential noise features of the current time step based on the actual collected diffusion magnetic resonance data. An optimization module is used to remove noise from the optimized noise latent features to obtain noise-free latent features, decode the noise-free latent features using 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 diffuse 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, the processor executing the program to implement the diffusion prior-based microstructure model optimization method as described in any one of claims 1-8.

Images