Diffusion magnetic resonance imaging method and system for high b value synthesis

Abstract

The invention provides a diffusion magnetic resonance imaging method and system for high-b-value synthesis. The method comprises the following steps: acquiring a plurality of low-b-value diffusion weighted images; gradient direction consistency resampling is carried out by using a spherical harmonic function; establishing a mathematical model to carry out parameterized representation on the low b value DWI signal; establishing and training a de-noising diffusion probability model, and synthesizing ultra-high b value data by taking the low b value DWI data as input; the finally generated ultrahigh b-value diffusion weighted image can be used for tasks such as nerve tissue microenvironment analysis, nerve disease detection, brain connection group research and the like. According to the method, a high-quality ultrahigh-b-value DWI image can be obtained without high-gradient magnetic resonance scanning equipment, the data acquisition cost is reduced, the accuracy of neural image analysis is improved, and important technical support is provided for neuroscience research and clinical application.

Description

Technical Field

[0001] The present invention belongs to the technical field of magnetic resonance imaging, and particularly relates to a diffusion magnetic resonance imaging method and system for high b-value synthesis. Background Art

[0002] Diffusion magnetic resonance imaging (dMRI) is an advanced neuroimaging technique that can detect the microstructural features of human brain tissue and provides an important tool for studying the microstructure of brain tissue and its changes in diseases. Among the dMRI acquisition parameters, a higher b-value can enhance the sensitivity to microstructural changes and pathological changes. In particular, ultra-high b-value acquisition allows for a more detailed characterization of brain tissue properties and is of great value for advanced research and clinical applications. However, obtaining high b-value dMRI data faces many challenges, including long scanning times, the need for high-performance MRI scanners, and the complexity of the data acquisition process, which significantly limit its feasibility in clinical applications.

[0003] Currently, the main challenge in generating high b-value DWI images lies in the significant difference between low b-value and high b-value diffusion signals. Low b-value images retain a higher signal intensity but have lower sensitivity to microstructures; while high b-value imaging, although improving the sensitivity to microstructural details, results in a significant decrease in signal intensity and reduces the signal-to-noise ratio (SNR). Achieving sufficient diffusion weighting requires a high-performance gradient system with extremely high gradient strength and high magnetic field switching rate, which is crucial for reducing the echo time (TE) and achieving shorter diffusion pulses. These requirements can usually only be met by the most advanced MRI scanners (such as Connectome scanners), whose gradient strength can reach 300 mT / m. In addition, flexibly and continuously predicting high b-value dMRI from various low b-value data is also a challenge. Existing DWI synthesis methods mainly include generative adversarial networks (GANs) and diffusion models. Although these methods have achieved high results in DWI synthesis quality, their scalability to new protocols is limited because changes in b-value acquisition settings in clinical acquisitions usually require retraining the model. In addition, existing methods cannot generate ultra-high b-value DWI images and are difficult to capture subtle microstructural characteristics. Therefore, developing an ultra-high b-value dMRI synthesis method that can adapt to different b-value acquisition protocols and utilize the attenuation pattern of low b-value data is of great significance for promoting the development of this field. Summary of the Invention

[0004] The present invention aims to provide a diffusion magnetic resonance imaging method and system for high b-value synthesis to solve the difficulty of obtaining ultra-high b-value dMRI data in the prior art, improve the characterization ability of the microstructure of neurons and cell body tissues, and enhance the depth and breadth of clinical applications and neuroscience research.

[0005] The technical solution for achieving the object of the present invention is as follows:

[0006] A diffusion magnetic resonance imaging method for high b-value synthesis, comprising the following steps:

[0007] Step 1, the subject undergoes diffusion-weighted imaging scanning in a magnetic resonance device, samples according to the setting of low b-values, and obtains a plurality of diffusion magnetic resonance imaging data with low b-values;

[0008] Step 2, process the plurality of diffusion magnetic resonance imaging data with low b-values based on spherical harmonics, obtain a plurality of spherical harmonic representations with low b-values, and resample them in accordance with a unified gradient direction to obtain a plurality of diffusion-weighted images with consistent gradient directions.

[0009] Step 3, establish an exponential model to fit the plurality of diffusion-weighted images with low b-values to obtain a continuous representation of the data, and resample the data using the continuous representation coefficients to obtain diffusion-weighted images with b = 1000, 3000, 5000 s / mm 2 Finally, normalize the data using the b = 0 image;

[0010] Step 4, establish and train a model for synthesizing diffusion-weighted images with b = 10000 s / mm 2 from diffusion-weighted images with b = 1000, 3000, 5000 s / mm 2 ;

[0011] Step 5, perform spherical harmonic fitting, gradient direction resampling, continuous representation, b-value input consistency resampling, and normalization on the diffusion-weighted images with different b-values to be tested, and then input them into the pre-trained network to obtain diffusion-weighted images with ultra-high b-values. The generated ultra-high b-value DWI data can be used for applications such as nerve tissue microenvironment analysis and nerve disease detection

[0012] Further, in the step 1, the number of b-values in the multi-b-value diffusion-weighted imaging data is greater than or equal to 3, and the higher the b-value, the better the quality of the finally synthesized ultra-high b-value diffusion-weighted image.

[0013] Further, the spherical harmonic fitting and gradient direction resampling in the step 2 are respectively performed using the amp2sh and sh2amp instructions in the MRtrix3 tool library to obtain diffusion-weighted images with consistent gradient directions.

[0014] Further, the exponential model in the step 3 is:

[0015] S b = α·exp(-β·b)+γ, α, β, γ > 0,

[0016] where α, β, and γ are the three coefficients of the exponential model, i.e., the continuous representation, and S b represents the multi-b-value diffusion weighted image signal of a single voxel in the same gradient direction; the exponential model is used to fit the multi-b-value diffusion weighted image signals corresponding to each voxel, encoding its attenuation characteristics so that it can adapt to data of different scanning protocols. The exponential fitting process uses the non-linear least squares method to provide stable fitting results.

[0017] Further, the value range of the coefficient β in the exponential model is [0, 0.007], and this value range is determined by analyzing diffusion magnetic resonance imaging data of different acquisition protocols, making the parameterization process more robust and the obtained information more accurately represent the attenuation law.

[0018] Further, resampling is performed using the continuous representations α, β, γ obtained in step 3 to obtain diffusion weighted images with b = 1000, 3000, 5000 s / mm 2 and the images with each b value are normalized by dividing by the b = 0 image.

[0019] Further, the model established in step 4 uses the Denoising Diffusion Probabilistic Model (DDPM) framework to generate high-b-value DWI data from low-b-value DWI data through an iterative denoising process: in the forward diffusion process, Gaussian noise is gradually added to the low-b-value DWI data until it becomes a pure Gaussian noise image; in the reverse diffusion process, the signal attenuation pattern is learned through a deep learning network (U-Net) and denoised step by step to finally synthesize an ultra-high-b-value DWI image. This method combines the microstructural information of low-b-value DWI and utilizes the high-quality generation ability of the diffusion model to improve the structural consistency and biological interpretability of ultra-high-b-value DWI.

[0020] Further, the neural network established by the diffusion model in step 4 is based on the U-Net architecture, with a basic number of channels of 64 and a Dropout rate of 0.2. The encoder part consists of four stages (the channel multiplication coefficients are 1, 2, 4, 8 respectively), and each stage consists of two residual blocks and a downsampling operation. To enhance the feature expression ability, a self-attention layer is introduced at the 16×16 resolution to capture global information.

[0021] Furthermore, the decoder part of the U-Net architecture is symmetric to the encoder, using upsampling and skip connections to restore spatial details and maintain the consistency of structural information. During the decoding process, the model performs self-attention optimization on the intermediate feature maps and combines positional encoding to enhance the feature learning ability. In addition, the model introduces noise level embeddings through feature channel affine transformation to adapt to input data with different noise distributions and improve the generation effect. The model finally uses a 3×3 convolutional layer for reconstruction to ensure the structural fidelity of ultra-high b-value DWI and achieve high-quality image synthesis without adding excessive computational overhead.

[0022] A diffusion magnetic resonance imaging system for high b-value synthesis, comprising a data acquisition module, a data processing module, a data fitting and normalization module, a neural network training module, and an ultra-high b-value data acquisition module.

[0023] The data acquisition module samples according to the settings of different b-values to obtain multiple diffusion magnetic resonance imaging data with low b-values.

[0024] The data processing module fits the multiple diffusion magnetic resonance imaging data with low b-values based on spherical harmonics and resamples the gradient directions to obtain a diffusion-weighted image with consistent gradient directions.

[0025] The data fitting and normalization module fits the multi-b-value diffusion-weighted image using an exponential model to obtain a continuous representation of the data, and uses the continuous representation coefficients to resample the data for b-value consistency and normalize the data.

[0026] The neural network training module constructs and trains a denoising diffusion probability model based on the U-Net architecture. This model uses low b-value data as the condition of the diffusion model to perform reverse denoising from a pure Gaussian noise image according to the low b-value conditional input to gradually restore the high b-value image; ultimately, it realizes the synthesis of diffusion-weighted images with b = 10000 s / mm 2 of diffusion-weighted images.

[0027] The ultra-high b-value data acquisition module inputs the diffusion-weighted images with different b-values to be tested after being processed by steps 2-3 into the trained denoising diffusion probability model network to obtain ultra-high b-value diffusion-weighted images.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The method of the present invention uses spherical harmonic function fitting to resample multiple low b-value diffusion-weighted images in the gradient direction, ensuring the consistency of the input data, enabling the model to be applicable to data collected under different gradient direction acquisition protocols, and enhancing its cross-dataset applicability; in the method of the present invention, for the first time, it is proposed to continuously characterize the diffusion signal through an exponential model, effectively encoding the signal attenuation pattern, enabling the model to adapt to different b-value scanning parameters and enhancing the generalization ability.

[0030] 2. The present invention is based on the denoising diffusion probability model (DDPM), and generates high b-value DWI data by gradually denoising. Compared with generation methods such as GAN, it has higher stability and image quality; the present invention uses a U-Net neural network for feature extraction and reconstruction, combines residual connections (ResNet) and attention mechanisms to improve the adaptability of the model to data of different subjects, while reducing the training and inference time and improving the computing efficiency.

[0031] 3. The present invention performs denoising generation through the DDPM diffusion model to gradually reconstruct high-quality ultra-high b-value DWI data, making the synthetic image have a high signal-to-noise ratio (SNR) and microstructural consistency, and solving the limitation that existing methods usually have difficulty in reliably synthesizing ultra-high b-value DWI with b = 10000 s / mm 2 ; through the parametric modeling of the exponential model, the present invention extracts the attenuation pattern of the low b-value DWI data signal, improving the accuracy and interpretability of the synthetic image. By applying the synthetic ultra-high b-value DWI data in the analysis of the microstructure of nerve tissue, the accuracy of nerve fiber tracking, cell body density measurement, free water imaging, etc. can be improved, providing more accurate image data support for neuroscience research and disease diagnosis.

[0032] 4. The present invention proposes a method for synthesizing ultra-high b-value DWI based on a diffusion model, which has strong cross-protocol adaptability, higher-quality image generation effect, lower equipment cost, more efficient computing performance and broader clinical application prospects, and has important technical value and application potential in the fields of neuroimaging and medical image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flowchart of the diffusion magnetic resonance imaging method for high b-value synthesis of the present invention;

[0034] Figure 2 The overall workflow block diagram;

[0035] Figure 3 is a performance comparison diagram of the method of the present invention with the DDPM and SR3 methods in generating images with b = 10000 s / mm 2 ;

[0036] Figure 4 The downstream microstructure analysis index derived from the data generated by using the method of the present invention with b = 10000 s / mm 2 is compared with the microstructure analysis index (benchmark) derived only from low b-value data and the microstructure index (reference image) derived from the data with the true b = 10000 s / mm 2 data;

[0037] Figure 5 After pre-training with the method of the present invention, b = 10000 s / mm data is exported on datasets with different acquisition protocols. The figure shows the comparison of the downstream microstructure analysis index derived from the high b-value data generated by using the method of the present invention with the microstructure analysis index (benchmark) derived only from low b-value data and the microstructure index (reference image) derived from the data with the true b = 9850 s / mm 2 data. 2 is compared with the microstructure index (reference image) derived from the data with the true b = 9850 s / mm Detailed implementation manners

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] In conjunction with Figure 1 and Figure 2 , a diffusion magnetic resonance imaging method for high b-value synthesis specifically includes:

[0040] First, a subject undergoes diffusion-weighted imaging scanning in a magnetic resonance device to acquire multiple diffusion magnetic resonance imaging data (DWI) with low b-values (such as b = 1000, 3000, 5000 s / mm 2 ). Since the b-value settings of different datasets are different, in order to make the data applicable to a unified analysis framework, the present invention uses the Spherical Harmonics (SH) method to process the low b-value DWI data to make its gradient directions consistent. Specifically, the amp2sh instruction in the MRtrix3 tool library is used to convert multiple low b-value DWIs into spherical harmonic representations, and the sh2amp instruction is used to resample to a unified gradient direction, thereby obtaining low b-value diffusion-weighted images with consistent gradient directions. This preprocessing step can reduce the errors caused by gradient direction differences and improve the consistency and generalization ability of the data.

[0041] The present invention uses an exponential model to parametrically model the diffusion signals at different b-values to characterize the signal attenuation characteristics. The exponential model is defined as follows:

[0042] S b = α·exp(-β·b) + γ

[0043] S b represents the signal of a single voxel in the same gradient direction for multi-b value diffusion weighted images; the exponential model is used to fit the multi-b value diffusion weighted image signals corresponding to each voxel, encoding its attenuation characteristics to adapt to data from different scanning protocols. The present invention uses the non-linear least squares method to fit the multi-b value diffusion weighted image signals to obtain stable model parameters. To make the fitting process more robust, the present invention refers to healthy subjects in the training set and restricts β between 0 and 0.007. By resampling using the obtained continuous representations α, β, γ, diffusion weighted images with b = 1000, 3000, 5000 s / mm 2 can be resampled and calculated, and the images with each b value are normalized by dividing by the b = 0 image:

[0044]

[0045] where S0 is the DWI image with b = 0 to ensure that the signal intensity distributions of images with the same b value between different datasets are closer, improving the stability and comparability of the synthetic data.

[0046] The present invention uses the Denoising Diffusion Probabilistic Model (DDPM) to synthesize ultra-high b value DWI data with b = 10000 s / mm 2 . DDPM consists of two processes: the forward diffusion process and the reverse denoising process. In the forward process, Gaussian noise is gradually added to the low b value DWI data, and finally pure Gaussian noise data is generated. This process can be described as:

[0047]

[0048] where α t is the preset noise attenuation coefficient, and x t represents the noise data at step t.

[0049] The reverse process then learns the denoising process and gradually recovers the ultra-high b value DWI image from the pure noise data. This process can be expressed as:

[0050] p θ (x t-1 |x t ,y) = N(x t-1 ; μ θ (x t ,y,t),∑ θ (x t ,y,t))

[0051] where y is the low b-value data input, t represents the time step information, and μ θ and ∑ θ represent the mean and variance estimated by the neural network U-Net. Through conditional diffusion modeling, the present invention can make full use of the information of low b-value DWI to improve the structural integrity and biological interpretability of ultra-high b-value DWI.

[0052] The present invention designs a neural network based on the U-Net architecture, with a basic number of channels of 64 and a Dropout rate of 0.2. The encoder part contains four stages (the channel multiplication coefficients are 1, 2, 4, and 8 respectively), and each stage consists of two residual blocks and a downsampling operation. To enhance the feature expression ability, a self-attention layer is introduced at the 16×16 resolution to capture global information. The decoder part is symmetric with the encoder, using upsampling and skip connections to restore spatial details and maintain the consistency of structural information. During the decoding process, the model performs self-attention optimization on the intermediate feature maps and combines positional encoding to enhance the feature learning ability. In addition, the model introduces noise level embedding through feature channel affine transformation to adapt to input data with different noise distributions and improve the generation effect. The model finally uses a 3×3 convolutional layer for reconstruction to ensure the structural fidelity of ultra-high b-value DWI and achieve high-quality image synthesis without adding too much computational overhead.

[0053] Finally, in the inference stage, for DWI data with different b-values, first, spherical harmonic function fitting and gradient direction resampling are performed to make the gradient direction distribution uniform under different scanning protocols. Subsequently, the exponential model is used to continuously characterize multi-b-value DWI, and b-value consistency adjustment is performed based on the fitted α, β, γ parameters, that is, resampling is performed according to b = 1000, 3000, 5000 s / mm 2 so that the input data with different b-values can adapt to the pre-training distribution of the model. Then, the b = 0 image is used for normalization operation to ensure the consistency of the network input distribution. The normalized low b-value DWI data is used as conditional information to input the trained diffusion model, and an ultra-high b-value DWI image with b = 10000 s / mm 2 is generated through the conditional denoising process. Since the DDPM model fully learns the corresponding structural information between low b-value DWI and ultra-high b-value DWI during the training process, the generated ultra-high b-value DWI has high consistency in terms of signal-to-noise ratio, microstructural consistency, and biological interpretability. The finally obtained ultra-high b-value DWI data can be widely applied to tasks such as the analysis of the microenvironment of nerve tissue (such as the density of neuron cell bodies and axons), the detection of nerve diseases, and the study of brain network connectivity, providing a low-cost and high-quality alternative solution for high b-value diffusion magnetic resonance imaging.

[0054] A diffusion magnetic resonance imaging system for high b-value synthesis, comprising a data acquisition module, a data processing module, a data fitting and normalization module, a neural network training module, and an ultra-high b-value data acquisition module. The data acquisition module samples according to the setting of different b-values to obtain a plurality of diffusion magnetic resonance imaging data with low b-values; the data processing module fits the plurality of diffusion magnetic resonance imaging data with low b-values based on spherical harmonic functions and resamples the gradient directions to obtain a diffusion-weighted image with consistent gradient directions; the data fitting and normalization module fits the multi-b-value diffusion-weighted image using an exponential model to obtain a continuous representation of the data, and resamples based on this representation to obtain diffusion-weighted images with b = 1000, 3000, 5000 s / mm 2 At the same time, the b = 0 image is used for normalization processing; the neural network training module establishes and trains a deep learning network for synthesizing an ultra-high b-value diffusion-weighted image with b = 10000 s / mm from diffusion-weighted images with b = 1000, 3000, 5000 s / mm based on the denoising diffusion probability model (DDPM) framework; the ultra-high b-value data acquisition module inputs the diffusion-weighted image to be tested, which has been subjected to spherical harmonic function fitting, gradient direction resampling, exponential model parameterization, and normalization, into the trained diffusion model, and finally generates an ultra-high b-value diffusion-weighted image for applications such as neurotissue microenvironment analysis and neurological disease detection. 2 Diffusion-weighted image synthesis of b = 10000 s / mm 2 The ultra-high b-value data acquisition module inputs the diffusion-weighted image to be tested, which has been subjected to spherical harmonic function fitting, gradient direction resampling, exponential model parameterization, and normalization, into the trained diffusion model, and finally generates an ultra-high b-value diffusion-weighted image for applications such as neurotissue microenvironment analysis and neurological disease detection.

[0055] Figure 3 This is a performance comparison chart of the method of the present invention with the DDPM and SR3 methods in generating an image with b = 10000 s / mm 2 The results show that the synthetic image generated by the method of the present invention is most similar to the reference image, effectively retaining fine details and structural integrity. The error heat map shows that the method of the present invention achieves the lowest and most uniform error distribution, highlighting its accuracy and reliability.

[0056] Figure 4 For downstream microstructural analysis metrics derived from data with b = 10000 s / mm generated using the method of the present invention, compared with microstructural analysis metrics (benchmark) derived from only low b-value data and microstructural metrics (reference image) derived from using real b = 10000 s / mm 2 Data; the results show that the SANDI metric obtained from the synthetic DWI by the method of the present invention is very close to the SANDI metric obtained from the real DWI, and is better than the comparison metrics derived from only low b-value data. This consistency indicates that the synthetic method of the present invention effectively captures the key attributes required for accurate microstructural imaging, highlighting the accuracy and reliability of the method of the present invention in capturing microstructural attributes. 2 Data; the results show that the SANDI metric obtained from the synthetic DWI by the method of the present invention is very close to the SANDI metric obtained from the real DWI, and is better than the comparison metrics derived from only low b-value data. This consistency indicates that the synthetic method of the present invention effectively captures the key attributes required for accurate microstructural imaging, highlighting the accuracy and reliability of the method of the present invention in capturing microstructural attributes.

[0057] Figure 5 After pre-training the method of the present invention, b = 10000 s / mm is exported on datasets with different acquisition protocols 2 data. The figure shows the downstream microstructural analysis metrics derived from the high-b value data generated by the method of the present invention, compared with the microstructural analysis metrics (benchmark) derived only from low-b value data and the microstructural metrics (reference image) derived from data with a true b = 9850 s / mm 2 data. The results show that the metrics obtained by the method of the present invention are more similar to the reference metrics, confirming the effectiveness of the method of the present invention in retaining microstructural details and generating high-quality DWI images suitable for accurate microstructural analysis when using only low-b value diffusion MRI data.

[0058] The present invention can obtain high-quality ultra-high b-value DWI images without high-gradient magnetic resonance scanning equipment, reduce the data acquisition cost, improve the accuracy of neuroimaging analysis, and provide important technical support for neuroscience research and clinical applications.

[0059] It should be understood that the above description of the preferred embodiments is relatively detailed, and thus it should not be considered as limiting the protection scope of the patent of the present invention. Under the inspiration of the present invention, those of ordinary skill in the art can also make substitutions or deformations without departing from the protection scope defined by the claims of the present invention, and all fall within the protection scope of the present invention. The scope of protection claimed by the present invention shall be subject to the appended claims.

Claims

1. A diffusion magnetic resonance imaging method for high b-value synthesis, characterized in that, Including: Step 1: Perform diffusion-weighted imaging scanning using a magnetic resonance device, sample according to the setting of low b-values, and obtain a plurality of diffusion magnetic resonance imaging data with low b-values; Step 2: Process the plurality of diffusion magnetic resonance imaging data with low b-values based on spherical harmonic functions, obtain a plurality of spherical harmonic function representations with low b-values, and resample them in accordance with a unified gradient direction to obtain a plurality of diffusion-weighted images with low b-values and consistent gradient directions; Step 3: Establish an exponential model to fit the plurality of diffusion-weighted images with low b-values to obtain a continuous representation of the data, and use the continuous representation coefficients to resample the data for b-value consistency and normalize the data; Step 4: Construct and train a denoising diffusion probabilistic model based on the U-Net architecture. This model uses low b-value data as the condition for the diffusion model to perform reverse denoising and gradually recover high b-value images from pure Gaussian noise images according to the input of low b-value conditions. Finally, it realizes the synthesis of diffusion-weighted images with b = 10000 s / mm 2 from diffusion-weighted images with different b-values; Step 5: After processing the diffusion-weighted images with different b-values to be tested through Steps 2-3, input them into the trained denoising diffusion probabilistic model network to obtain diffusion-weighted images with ultra-high b-values.

2. The diffusion magnetic resonance imaging method for high b-value synthesis according to claim 1, wherein Resample using the exponential model parameters obtained in step 3, and calculate the diffusion-weighted images with b = 1000, 3000, 5000 s / mm 2 and normalize each b-value image by dividing it by the b = 0 image.

3. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 1, characterized in that, The denoising diffusion probabilistic model in Step 4 generates high b-value DWI data from low b-value DWI data through an iterative denoising process, including a forward diffusion process and a reverse denoising process. In the forward diffusion process, Gaussian noise is gradually added to the low b-value DWI data until it becomes a pure Gaussian noise image; The reverse denoising process learns the signal attenuation pattern through a trained deep learning network and gradually denoises it, finally synthesizing an ultra-high b-value DWI image.

4. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 3, characterized in that, The forward diffusion process is: Among them, α t is a preset noise attenuation coefficient, and x t represents the noise data at the t-th step.

5. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 3, characterized in that The reverse denoising process is: p θ (x t-1 |x t ,y) = N(x t-1 ; μ θ (x t ,y,t), Σ θ (x t ,y,t) where y is the low b-value data input, t represents the time step information, μ θ and ∑ θ represent the mean and variance estimated by U-Net.

6. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 3, characterized in that The basic number of channels of the denoising diffusion probabilistic model is 64, and the Dropout rate is 0.

2.

7. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 6, characterized in that, The encoder in the U-Net architecture includes four stages, and the channel multiplication coefficients are 1, 2, 4, and 8 respectively. Each stage consists of two residual blocks and a downsampling operation; a self-attention layer is introduced at a resolution of 16×16.

8. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 7, characterized in that The decoder in the U-Net architecture is symmetric with the encoder, and uses upsampling and skip connections. During the decoding process, the model performs self-attention optimization on the intermediate feature maps and combines positional encoding to enhance the feature learning ability.

9. A diffusion magnetic resonance imaging method for high b-value synthesis according to claim 8, characterized in that, The denoising diffusion probabilistic model introduces noise level embeddings through feature channel affine transformations to adapt to input data with different noise distributions, and finally uses a 3×3 convolutional layer for reconstruction to ensure the structural fidelity of ultra-high b-value DWI.

10. A diffusion magnetic resonance imaging system for high b-value synthesis implementing the method according to any one of claims 1-9, characterized in that, Including a data acquisition module, a data processing module, a data fitting and normalization module, a neural network training module, and an ultra-high b-value data acquisition module, The data acquisition module samples according to the setting of different b-values to obtain a plurality of diffusion magnetic resonance imaging data with low b-values; The data processing module fits the plurality of diffusion magnetic resonance imaging data with low b-values based on spherical harmonic functions and resamples the gradient direction to obtain a diffusion-weighted image with a consistent gradient direction; The data fitting and normalization module uses an exponential model to fit the multi-b-value diffusion-weighted images to obtain a continuous representation of the data, and uses the continuous representation coefficients to resample the data for b-value consistency and normalize the data; The neural network training module constructs and trains a denoising diffusion probability model based on the U-Net architecture. This model uses low b-value data as the condition for the diffusion model and is used to perform reverse denoising step by step to recover high b-value images from pure Gaussian noise images according to the input of low b-value conditions. Finally, it realizes the synthesis of diffusion-weighted images with b = 10000 s / mm 2 from diffusion-weighted images with different b-values; The ultra-high b-value data acquisition module inputs the diffusion-weighted images with different b-values to be tested after being processed through Steps 2-3 into the trained denoising diffusion probabilistic model network to obtain diffusion-weighted images with ultra-high b-values.

Images