Diffusion magnetic resonance imaging method and system for multi-scale brain tissue microenvironment measurement

Abstract

The invention provides a diffusion magnetic resonance imaging method and system for multi-scale brain tissue microenvironment measurement, and the method comprises the steps: obtaining a spherical average signal of multi-b-value diffusion magnetic resonance data, carrying out the parameterization and normalization, and constructing a data training set; a microstructure dictionary comprising a feature extraction module and a multi-scale aggregation module is constructed, the feature extraction module is based on a visual Mangbar-convolution hybrid architecture and is used for extracting multi-scale structure information, and the multi-scale aggregation module is used for aggregating the extracted multi-scale structure information to obtain compact feature representation; and training the microstructure dictionary through the data training set, and carrying out joint characterization on the plurality of microstructure indexes by using the microstructure dictionary obtained by training to obtain a microenvironment measurement index. According to the method, the microenvironment parameters of multiple human brain tissues can be rapidly and quantitatively described in a cross-data acquisition protocol mode, and clinical application of dMRI microstructure imaging is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical fields of magnetic resonance imaging and medical image processing, and particularly relates to a diffusion magnetic resonance imaging method and system for multi-scale measurement of brain tissue microenvironment. Background Art

[0002] Diffusion magnetic resonance imaging (dMRI) is an important imaging modality for detecting the microstructure of tissues, which can non-invasively provide information such as nerve fiber structure, cell body density, and free water content, and has wide applications in neuroscience research and clinical diagnosis. The biophysical diffusion model based on diffusion magnetic resonance imaging plays an important role in describing the complex changes of tissue microstructures such as dendrites, axons, and glial cells, providing valuable information for understanding the structural basis of the human brain. The progress of magnetic resonance imaging hardware, the improvement of computing power, and the development of neuroscience have promoted the development of more complex dMRI biophysical models, such as mean signal diffusion kurtosis imaging (MSDKI). Advanced biophysical models can relate diffusion-weighted signals to potential tissue microstructures, thereby capturing more realistic tissue characteristics, such as cell size, membrane permeability, and the geometry of the cellular microenvironment. However, the microenvironment measurement technology based on dMRI models still faces a series of challenges, including dependence on precise acquisition protocols, long scan times, intensive data processing procedures, and limitations in specificity for specific neurological diseases. These factors have largely limited the clinical application of these microstructure imaging technologies.

[0003] Currently, microstructural imaging methods based on biophysical model fitting rely on predefined biophysical assumptions to fit the dMRI signal and thus provide reliable microstructural indices. Advanced biophysical modeling techniques, such as diffusion-based spectrum imaging, restricted spectrum imaging, and spherical average spectrum imaging, provide sophisticated means to probe tissue microstructure. These methods can provide advanced microstructural imaging modalities but require extensive computational resources and densely sampled q-space data. To address these limitations, various deep learning methods have been proposed and can be divided into three main categories. The first category focuses on improving the estimation accuracy of a single microstructural model using advanced architectures and optimization strategies; the second category focuses on reducing the reliance on densely sampled q-space data and accelerating data processing; and the third category aims to jointly estimate multiple microstructural indices from multiple biophysical models. However, current microstructural imaging techniques still face several interrelated challenges: 1) High-quality microstructural estimation requires dense sampling of multiple b-values, which results in long scan times and demanding equipment. 2) Existing deep learning-based methods generalize poorly across different datasets, making them difficult to apply clinically. 3) Existing deep learning methods lack scalability and can only predict microstructural metrics included in training. 4) Some dMRI models (such as SANDI) rely on ultra-high b-value acquisitions, making it difficult to calculate reliable SANDI metrics from low-b-value data. These interrelated adjustments highlight the need for more flexible and versatile imaging methods. Summary of the Invention

[0004] In order to overcome the challenges of existing technologies, the present invention provides a diffusion magnetic resonance imaging method and system for multi-scale brain tissue microenvironment measurement, which is used to perform high-quality, rapid and interpretable multi-parameter microstructure imaging, thereby improving the clinical applicability and scientific research applicability of human brain tissue microstructure measurement technology.

[0005] The technical solutions for achieving the purpose of the present invention are:

[0006] A diffusion magnetic resonance imaging method for multi-scale brain tissue microenvironment measurement comprises the following steps:

[0007] Step 1: Obtain the spherical average signal of multi-b-value diffusion magnetic resonance data, perform parameterization and normalization, and construct a data training set;

[0008] Step 2: Construct a microstructure dictionary including a feature extraction module and a multi-scale aggregation module. The feature extraction module is based on a visual mamba-convolution hybrid architecture and is used to extract multi-scale structural information. The multi-scale aggregation module aggregates the extracted multi-scale structural information to obtain a compact feature representation.

[0009] Step 3: Train the microstructure dictionary using the data training set, and use the trained microstructure dictionary to jointly characterize multiple microstructure indicators to obtain microenvironment measurement indicators.

[0010] Furthermore, the number of b values in the diffusion MRI data with multiple b values in step 1 is greater than or equal to 3 (including b=0s / mm 2 ), and the more gradients there are for each b-value, the better the final tissue microenvironment measurement estimation effect.

[0011] Furthermore, the spherical average signal in step 1 is the average of signals in different gradient directions, eliminating the influence of fiber orientation distribution; the spherical average signal is defined as:

[0012] Furthermore, in step 1, an exponential model is used to parameterize the multi-b-value spherical average signal. The exponential model is:

[0013]

[0014] Among them, α, β and γ are the three coefficients of the exponential model, that is, the parameterized representation. Represents the spherical average signal of different b-values of a single voxel This exponential model is used to fit the multi-b-value spherical average signal corresponding to each voxel, encoding its attenuation characteristics, and thus obtaining a parametric representation (α, β, γ) that is independent of a specific b-value acquisition protocol. This allows networks trained based on this parametric representation of the multi-b-value spherical average signal to generalize well across data acquisition protocols, reduce input dimensionality, and accelerate data processing. The exponential fitting process is performed using a nonlinear least squares method, and a parallel approach is used to accelerate the fitting of whole-brain voxels. α, β, and γ are then z-normalized to reduce inter-sample variability, resulting in a mean of 0 and a variance of 1 for each parameter map. This reduces inter-individual variability in signal amplitude, improves cross-subject consistency, and enhances network generalization.

[0015] Furthermore, the feature extraction module in step 2 is composed of a 1×1 convolutional layer and four cascaded Vim-CNN units in series, wherein the convolutional layer uses a 1×1 convolution kernel to adjust the number of channels of the input image from 3 to 32; the number of input channels of each Vim-CNN module is C, and the number of output channels is 2C.

[0016] Furthermore, the specific process of the feature extraction module to extract multi-scale structural information is as follows:

[0017] Assume that the input diffusion magnetic resonance parameter map is H and W are the height and width of the image respectively, C0 is the number of channels, and the 1×1 convolution layer uses the convolution kernel and bias b for X 0 is mapped, and combined with the ReLU activation function to obtain X 0 ∈R H×W×32 ;

[0018] The input of the i-th Vim-CNN unit is The output is C i are 32, 64, 128, 256 in sequence, 1 ≤ i ≤ 4, representing the number of input channels of the i-th Vim-CNN unit; X i is obtained through the following steps:

[0019] (5) Dual 3×3 convolution branch: Through two 3×3 convolutional kernels and bias p i , and combined with ReLU to extract local features;

[0020] (6) Vision Mamba branch: Capture global dependencies through the Vision Mamba module to obtain At the channel dimension, is concatenated with to obtain

[0021] (7) Single convolution branch: Through a 1×1 convolutional kernel and bias b i , and combined with ReLU to map X i to obtain

[0022] (8) Residual fusion:

[0023] Furthermore, the specific implementation process of the Vision Mamba module is as follows:

[0024] First, normalize the input feature sequence T l-1 , and then project it into x and z, where x is the main input feature after projection of the input sequence, and z is another set of gating control information extracted from the input, which is used to determine how information is fused in the subsequent gating fusion mechanism. For the forward (f, that is, forward) and backward (b, that is, backward) d ∈ {f, b}, discrete parameters:

[0025]

[0026] Among them, Δ d is the state step size, which controls the discrete time interval of state update, A d is the state transition matrix, which describes the change trend of the state, exp is the exponential matrix operation, and is used to simulate the evolution of continuous states in the state space, Bd is the input control matrix that adjusts the influence of the input on the current state. Calculate the hidden state and update the state iteratively through the following formula

[0027]

[0028] where represents the feature in a specific direction enhanced by one-dimensional convolution, is the hidden state at the previous time step, is the output in direction d at time step t, C d is the output projection matrix used to map the updated state to the output space, is the updated hidden state at the current time step;

[0029] Finally, the bidirectional output Y is calculated through the state h after iterative update d ∈{h f ,h b}, f = C f h f and Y b = C b h b , and an adaptive fusion is performed through z using a gating mechanism to form the final feature sequence T l :

[0030] T l = SiLU(z)⊙(Y f + Y b ) + T l-1

[0031] where SiLU is the Sigmoid gated linear unit, z is the gating information used to control the fusion, and ⊙ is the element-wise multiplication operation.

[0032] Furthermore, the multi-scale aggregation module in step 2 is composed of a Vim module, a spatial aggregation module, and a 1×1 convolutional layer connected in series. The Vim module uses Vision MambaBlock to perform bidirectional state space modeling on the input feature map, capture global spatio-temporal dependencies, and keep the number of output channels the same as the input; the spatial aggregation module is based on a multi-order gating aggregation structure to adaptively fuse multi-scale context information in the spatial dimension with the number of channels unchanged; the 1×1 convolutional layer uses a convolutional kernel size of 1×1 to compress the feature dimension and enhance the information integration ability.

[0033] Furthermore, the spatial aggregation module includes a feature decomposition module and a multi-order gated aggregation module. The feature decomposition module uses 1×1 convolution operation to decompose the input feature X, and then uses a gated mechanism to re-weight the complementary interaction components. The expression is:

[0034] Y = Conv 1×1 (X), Z = GELU(Y + γ s ⊙(Y - GAP(Y)))

[0035] where γ s is a learnable scaling factor, GAP(·) represents the global average pooling operation, and GELU(·) represents the activation function;

[0036] The multi-order gated aggregation module adopts three parallel deep convolutional layers to encode low-order, middle-order, and high-order features at different dilation rates, and then concatenates the outputs of these layers into a unified multi-order context Y C = Concat(Y l , Y m , Y h ), and uses the gated mechanism of SiLU to fuse the context and the input feature:

[0037] Z = SiLU(Conv 1×1 (X)) ⊙ SiLU(Conv 1×1 (Y C ))

[0038] Furthermore, the microenvironment measurement indexes include: isotropic volume fraction, neurite density, neurite direction dispersion, free water content, axonal volume fraction, intrinsic diffusion coefficient, mean signal kurtosis, mean signal diffusion, fractional anisotropy, perfusion fraction, axonal water content, hindered axial diffusion coefficient, hindered radial diffusion coefficient, axonal diffusion coefficient, axial kurtosis, mean kurtosis, radial kurtosis, axial diffusion coefficient, mean diffusion coefficient, radial diffusion coefficient, volume fraction within neurites, intrinsic diffusion coefficient, mean square displacement, and return-to-origin probability.

[0039] Furthermore, it also includes: when other parameters need to be estimated, the microstructure dictionary is retrained using the training data, and the estimation of other microstructure parameters is migrated and characterized after the training adjustment.

[0040] A diffusion magnetic resonance imaging system for multi-scale measurement of the brain tissue microenvironment includes:

[0041] A data processing module, which acquires the spherical average signal of multi-b value diffusion magnetic resonance data, performs parameterization and normalization, and constructs a data training set;

[0042] The microstructural dictionary training module constructs a microstructural dictionary including a feature extraction module and a multi-scale aggregation module. The feature extraction module is based on the Vision Mamba-Convolution hybrid architecture and is used to extract multi-scale structural information. The multi-scale aggregation module aggregates the extracted multi-scale structural information to obtain a compact feature representation;

[0043] The joint representation module trains the microstructural dictionary through a data training set, and uses the trained microstructural dictionary to jointly represent multiple microstructural metrics to obtain a microenvironment measurement metric.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The method of the present invention can estimate up to 24 microstructural metrics from undersampled dMRI data, including: isotropic volume fraction, neurite density, neurite orientation dispersion, free water content, axonal volume fraction, intrinsic diffusion coefficient, mean signal kurtosis, mean signal diffusivity, fractional anisotropy, perfusion fraction, axonal water content, hindered axial diffusivity, hindered radial diffusivity, axonal diffusivity, axial kurtosis, mean kurtosis, radial kurtosis, axial diffusivity, mean diffusivity, radial diffusivity, intra-neurite volume fraction, intra-axonal diffusivity, mean squared displacement, and return-to-origin probability; covering 8 biophysical diffusion models widely used in clinical research, significantly reducing the data acquisition requirements while maintaining high accuracy; compared with traditional methods that require densely sampled high-b-value data for fitting, the present invention can still achieve high-quality microstructural parameter estimation with less data input through microstructural dictionary learning, reducing data acquisition requirements and improving imaging efficiency;

[0046] 2. The method of the present invention does not depend on a specific b-value acquisition protocol and can be applied to diffusion-weighted imaging data with different b-values and gradient directions, significantly enhancing the generalization ability of the model on different data sets; compared with existing methods that can only be applied to data within a specific b-value range, the present invention adopts spherical mean technology (SMT) combined with a novel exponential parameterization modeling method to ensure that the model can consistently estimate multiple high-quality microstructural metrics under different acquisition protocols, improving the generalization performance of the model under different MRI scanning conditions;

[0047] 3. The method of the present invention provides a structured and interpretable microstructural imaging framework that can effectively represent different dMRI biophysical models; compared with existing deep learning methods that directly learn the mapping from dMRI signals to microstructural parameters, the present invention constructs a unified representation system by proposing a learnable tissue compartment kernel, that is, a microstructural dictionary, enabling the model not only to support existing microstructural metrics but also to adapt to new microstructural models through efficient fine-tuning, greatly improving the scalability of the method;

[0048] 4. The present invention proposes a novel Vision Mamba-Convolution hybrid network structure for accurately estimating brain tissue microenvironment parameters. Based on the traditional convolutional neural network, this network integrates VisionMamba (a global modeler based on a bidirectional state space model) with local 3×3 convolutions, and parallelly extracts local spatial features and global context information through Vim-CNN units. Then, through channel projection and residual fusion, it realizes multi-scale and efficient feature expression. Subsequently, a spatial aggregation module is introduced to adaptively fuse multi-order contexts, and 1×1 convolutions are used to compress channels, obtaining a compact and discriminative feature representation, improving the estimation accuracy of 24 microstructural indexes such as anisotropic volume fraction and diffusion kurtosis. It can also be quickly adapted to new microstructural models through a small amount of fine-tuning, and has excellent scalability and robustness.

[0049] 5. The method of the present invention can accurately estimate microstructural indexes that usually require high b-values (b≥5000 s / mm 2 ) in the case of only having low b-value DWI data (b = 0, 1000, 3000 s / mm 2 ), such as the extracellular volume fraction of the SANDI model, etc. Compared with traditional methods that rely on expensive high-gradient field strength MRI scanners to obtain high b-value data, the method of the present invention reduces the dependence on high-end MRI equipment, making it more practical in different clinical environments and research scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flowchart of the diffusion magnetic resonance imaging method for multi-scale brain tissue microenvironment measurement of the present invention;

[0051] Figure 2 The overall work flow block diagram;

[0052] Figure 3 is a performance comparison diagram of the method of the present invention with methods such as multi-layer perceptron (MLP) and convolutional neural network (CNN) under different undersampling patterns (60 diffusion-weighted images);

[0053] Figure 4 is a result diagram of the method of the present invention first training on the dMRI data of dataset 1 (b = 0, 1000, 3000 s / mm 2 ) and then testing on the dMRI data of dataset 2 (b = 0, 7000, 1200, 2800 s / mm 2 ) using the obtained pre-trained model;

[0054] Figure 5The method of the present invention is first pre-trained on healthy samples in the dataset of dataset one, and then the test result graph on brain tumor patients in dataset two;

[0055] Figure 6 The SANDI index obtained after fine-tuning using the microstructure dictionary pre-trained by the method of the present invention is compared with the index (benchmark) directly obtained by fitting with a low b-value model and the reference high-quality SANDI index. Detailed implementation manners

[0056] 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.

[0057] Combined with Figure 1 and Figure 2 , a diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment specifically includes:

[0058] Step 1, data processing: The subject undergoes diffusion-weighted imaging scanning in a magnetic resonance device, samples according to different b-value settings, and obtains multi-b-value diffusion magnetic resonance data; processes the multi-b-value diffusion magnetic resonance data based on the spherical averaging technique to obtain the spherical average signal of the multi-b-value, so as to eliminate the influence of the fiber orientation distribution; uses an exponential model to parameterize the multi-b-value spherical average signal, obtains the parameterized representation of the data, and normalizes the parameterized representation to improve the adaptability across datasets; specifically includes:

[0059] First, the subject undergoes diffusion-weighted imaging (DWI) scanning in a magnetic resonance device to collect diffusion magnetic resonance imaging data with multiple b-values (the number of b-values is greater than or equal to 3, for example, b = 0, 1000, 3000 s / mm 2 ). Since the t-gradient direction settings of different datasets may be different, in order to ensure the robustness of the method for data with different gradient directions, the present invention uses the spherical averaging technique to calculate the spherical average signal, thereby eliminating the influence of fiber orientation on the signal and improving the generalization ability of the data. The spherical averaging technique focuses on direction-averaging the signal to eliminate the influence of the fiber direction distribution. Given a b-value, the spherical average signal is defined as:

[0060]

[0061] In a multi-compartment model, the spherical average signal can be expressed by the following formula:

[0062]

[0063] where S0 is the diffusion-weighted signal with b = 0, and λ ∥ is the parallel diffusivity, and λ ⊥ is the perpendicular diffusivity, and p(λ ∥ , λ ⊥ ) is the probability distribution; the probability distribution p(λ ∥ , λ ⊥ ) is used to encode the fraction of the spin packet, and then the perpendicular diffusivity λ ⊥ and the parallel diffusivity λ ∥ are used to parameterize as follows:

[0064]

[0065] where erf represents the error function.

[0066] Some studies have suggested that the signal measurement results of each voxel are the combined results of homogeneous spin packets from different positions within the voxel. Therefore, the spherical mean signal (SMS) can be expressed as a weighted combination of the spherical mean signals of the spin packets. The probability p(λ ∥ , λ ⊥ ) weights the diffusivity of the specified tissue compartment and allows the use of the diffusion scale spectrum [λ ∥ , λ ⊥ to probe the tissue microstructure. The method of the present invention introduces a learnable microstructure dictionary, which contains a series of tissue compartment kernels. By relating the spherical mean signal to multiple microstructure indices, the microstructure indices can be effectively estimated. Compared with the traditional model, the method of the present invention supports the use of a small amount of data for fine-tuning to estimate accurate dMRI indices, or migration to other microstructure indices, ensuring robustness, adaptability, and scalability under different acquisition protocols.

[0067] To reduce the dependence on a specific b-value acquisition protocol, the present invention proposes a novel method for continuously representing discrete multi-b-value SMS, which can enhance the generalization ability of the model across different data acquisition protocols. The present invention uses an exponential model to parameterize the spherical mean signal of each voxel to obtain a parameterized representation that can reflect the attenuation pattern. The exponential model is defined as follows:

[0068]

[0069] Among them, α, β, and γ are the three coefficients of each voxel, all of which are greater than 0. The present invention uses the non-linear least squares method to complete the above fitting process. And for the fitting of all brain voxels, a parallel method is adopted to accelerate the fitting. Subsequently, α, β, and γ are respectively normalized by Z-score to reduce the differences between samples, making the mean of each parameter map 0 and the variance 1, so as to reduce the differences in signal amplitudes between individuals, improve the cross-subject consistency of the data, and enhance the generalization ability of the network.

[0070] To learn a set of high-quality microstructural dictionaries from the mapping between spherical mean signals and multiple microstructural metrics, the present invention adopts eight typical microstructural models widely used in clinical research, including: (1) Neurite Orientation Dispersion and Density Imaging (NODDI) can detect the microstructures of dendrites and axons and reasonably estimate neurite density and orientation dispersion; (2) Free Water Diffusion Tensor Imaging (FwDTI) extends the DTI model, considers additional compartments to eliminate the influence of free water, and then infers the macroscopic geometry of tissues based on the displacement of water; (3) Mean Signal Diffusion Kurtosis Imaging (MSDKI) describes kurtosis through the mean signal in all directions. It does not depend on the fiber distribution function and overcomes the limitations of traditional DKI; (4) Intravoxel Incoherent Motion Imaging (IVIM) measures the diffusion characteristics of water molecules in static tissues and blood flow and can simultaneously measure microvascular and parenchymal abnormalities in the brain. (5) White Matter Tract Integrity (WMTI) can estimate the contributions of hindered and restricted diffusion of orderly arranged fibers and can estimate axonal volume fraction and diffusional extracellular curvature to distinguish axonal loss processes and myelin degeneration processes; (6) 3D Simple Harmonic Oscillator-based Reconstruction and Estimation (3D-SHORE) is an extension of the one-dimensional SHORE basis. It models the signal as a sum of continuous functions, enabling the analysis of the average diffusion rate within a voxel, the probability that water molecules in diffusion return to their original positions after a specified time interval, etc.; (7) Diffusional Kurtosis Imaging (DKI) provides a more complex characterization of the restricted water diffusion process and helps explore non-Gaussian diffusion metrics in tissues composed of complex components and structures; (8) Multi-Compartment Microscopic Diffusion Imaging (MCMDI) models non-Gaussian diffusion in brain tissues and uses spherical mean technology to eliminate the influence of orientation dispersion, thereby providing estimates of neuronal density indices in intra- and extra-cerebral neuronal regions. Through simple modifications, the method of the present invention can flexibly support more dMRI microstructural models.

[0071] Step 2, establish and train a hybrid Mamba-CNN network architecture for high-precision microstructure metric estimation. This network introduces the Vim-CNN module, which fuses convolutional operations and the Vision Mamba module to extract multi-scale structural information and learn a set of microstructure dictionaries for characterizing new microstructure model parameters.

[0072] The present invention proposes a Mamba-CNN hybrid network architecture for high-quality microstructure metric estimation. Specifically, the present invention designs a feature extraction module based on a visual mamba-convolution hybrid architecture, which integrates convolutional operations with the Vision Mamba module (Vim module) to extract multi-scale structural information, thereby achieving more accurate microstructure analysis. The feature extraction module consists of a 1×1 convolutional layer and four cascaded Vim-CNN units connected in series. Among them, the convolutional layer uses a 1×1 convolutional kernel to adjust the number of channels of the input image from 3 to 32; for the four cascaded Vim-CNN modules, the number of input channels of each Vim-CNN module is C, and the number of output channels is 2C, which are 32, 64, 128, 256, and 512 in sequence. The specific process of the feature extraction module is as follows:

[0073] Let the input diffusion magnetic resonance parameter map be H and W are the height and width of the image respectively, C0 is the number of channels, and the 1×1 convolutional unit uses the convolutional kernel and bias b to map X0, and combines with the ReLU activation function to obtain X 0 ∈R H×W×32 , that is:

[0074]

[0075] where represents the convolutional operation, and ReLU(·) represents the activation function.

[0076] For the four Vim-CNN units, according to the input channel number C i which are 32, 64, 128, 256 in sequence, the input of its i-th unit (1 ≤ i ≤ 4) is and the output is X i is obtained through the following steps:

[0077] (1) Dual 3×3 convolution branches: Through two 3×3 convolutional kernels and bias p i , combined with ReLU to extract

[0078] local features:

[0079]

[0080] (2) Vision Mamba Branch: Capture global dependencies through the VisionMamba module to obtain

[0081] Concatenate along the channel dimension with to obtain

[0082] (3) Single Convolution Branch: Through a 1×1 convolution kernel and bias b i , combined with the ReLU mapping X i , to obtain

[0083] (4) Residual Fusion:

[0084]

[0085] First, normalize the input feature sequence T l-1 , and then project it into x and z. Here, x is the main input feature after projection of the input sequence, and z is another set of gated control information extracted from the input, which is used to determine how information is fused in the subsequent gated fusion mechanism. For the forward (f, i.e., forward) and backward (b, i.e., backward) d ∈ {f, b}, the module uses discrete parameters:

[0086]

[0087] where, Δ d is the state step size, controlling the discrete time interval of state update, A d is the state transition matrix, describing the change trend of the state, exp is the exponential matrix operation, used to simulate the evolution of continuous states in the state space, B d is the input control matrix, adjusting the influence of the input on the current state. Calculate the hidden state and update the state iteratively through the following formula

[0088]

[0089] where represents the specific direction feature enhanced by one-dimensional convolution, is the hidden state of the previous time step, is the output at time step t in direction d, C d is the output projection matrix, used to project the updated state into the output space, is the updated hidden state at the current time step;

[0090] Finally, the bidirectional output Y is calculated through the state h after the iterative update d ∈{h f ,h b}, and the gating mechanism is used for adaptive fusion through z to form the final feature sequence T f =C f h f and Y b =C b h b , and the gating mechanism is used for adaptive fusion through z to form the final feature sequence T l :

[0091] T l =SiLU(z)⊙(Y f +Y b )+T l-1

[0092] where SiLU is the Sigmoid gated linear unit, z is the gating information for controlling the fusion, and ⊙ is the element-wise multiplication operation. By utilizing the efficient state-space modeling of Mamba, the network can effectively integrate spatial dependencies while ensuring computational efficiency, improve the feature expression ability, and enhance the accuracy of microstructure estimation.

[0093] Meanwhile, the present invention designs a multi-scale aggregation module, which is composed of a Vim module, a spatial aggregation module, and a 1×1 convolutional layer connected in series. Among them, the Vim module uses the VisionMamba Block to perform bidirectional state-space modeling on the input feature map, capture global spatio-temporal dependencies, and keep the output channel number consistent with the input (512); the spatial aggregation module is based on a multi-order gating aggregation structure to adaptively fuse multi-scale context information in the spatial dimension with the channel number unchanged (512); the 1×1 convolutional layer uses a convolutional kernel size of 1×1 to map the channel number from 512 to 128, which is used to compress the feature dimension and enhance the information integration ability. The spatial aggregation module includes a feature decomposition module and a multi-order gating aggregation module. The feature decomposition module can decompose the input feature X using 1×1 convolutional operations, and then use the gating mechanism to re-weight the complementary interaction components, and its formula is:

[0094] Y=Conv 1×1 (X),Z=GELU(Y+γ s ⊙(Y-GAP(Y)))

[0095] where γ s is a learnable scaling factor, GAP(·) represents the global average pooling operation, and GELU(·) represents the activation function. The multi-order gating aggregation module uses three parallel depth convolutional layers to encode low-order, middle-order, and high-order features at different dilation rates (d∈{1,2,3}), and then outputs the low-order, middle-order, and high-order of these layers (Yl , Y m , Y h ) The features are connected into a unified multi-order context Y C = Concat(Y l , Y m , Y h ). To adaptively aggregate the extracted features, it also adopts a SiLU gating mechanism for the context and input features, defined as:

[0096] Z = SiLU(Conv 1×1 (X)) ⊙ SiLU(Conv 1×1 (Y C ))

[0097] This design effectively extracts discriminative multi-order representations while maintaining computational efficiency and parameter compactness, improving the accuracy of microstructure estimation while maintaining efficiency and compactness.

[0098] The network structure and channel change strategy of the present invention are designed based on spectral imaging technology, and a learnable microstructure dictionary including a feature extraction module and a multi-scale aggregation module is introduced. The network first uses a feature extraction module to expand the number of channels from 3 to 512 to extract rich multi-level microstructure features. Subsequently, the Vim block and the spatial aggregation module are used to further optimize the feature representation, and the features are integrated into 128 overcomplete tissue compartment nuclei through a 1×1 convolutional layer to optimize the representation ability of the tissue compartment nuclei, that is, the microstructure dictionary, and reduce computational redundancy. Finally, the network adjusts the number of output channels according to the requirements of the target task through a flexible 1×1 convolutional output layer (microstructure joint representation layer) to adapt to different numbers and types of microstructure metrics, improving the scalability of the network, so that it can jointly represent different biophysical diffusion models according to the pre-trained microstructure dictionary.

[0099] Step 3, input the normalized parametric representation of the diffusion magnetic resonance imaging data with multiple b-values to be tested into the trained neural network, and using the pre-trained microstructure dictionary, joint estimates of multiple tissue microenvironment parameters can be obtained.

[0100] The method of the present invention will obtain an interpretable and over-complete microstructure dictionary to achieve the joint characterization of multiple microstructure index maps. Diffusion MRI (dMRI) data applicable to different data acquisition protocols are first processed by spherical averaging technology to calculate the spherical average signal, and then parameterized by an exponential model to obtain the parameterized characterization of the spherical average signal. After normalization using the Z-SCORE method, it is finally input into the trained model, and the output includes the joint estimation of 24 brain tissue microenvironment parameters, such as isotropic volume fraction, neurite density, neurite direction dispersion, free water content, axon volume fraction, intrinsic diffusion coefficient, mean signal kurtosis, mean signal diffusion, fractional anisotropy, perfusion fraction, axon water content, hindered axial diffusion coefficient, hindered radial diffusion coefficient, axon diffusion coefficient, axial kurtosis, mean kurtosis, radial kurtosis, axial diffusion coefficient, mean diffusion coefficient, radial diffusion coefficient, intra-neurite volume fraction, intra-axonal diffusion coefficient, mean square displacement, and return-to-origin probability.

[0101] In addition, the network architecture of the present invention has high flexibility and scalability, and supports adapting to new microstructure models by modifying the output layer. Specifically, after the microstructure dictionary of the network is pre-trained, it can be migrated between different biophysical diffusion models. Only by adjusting the final 1×1 convolutional layer and fine-tuning with a small amount of data can the new microstructure index be accurately estimated. The present invention can also estimate microstructure parameters that usually require high b-value data, such as the SANDI model-derived index, with only low b-value DWI data through minor fine-tuning. This design reduces the dependence on high b-value scan data and high-end MRI equipment, making the estimation of microstructure parameters more efficient and economical, and further enhancing the applicability of this method in various clinical applications and scientific research scenarios.

[0102] The present invention also proposes a diffusion magnetic resonance imaging system for multi-scale measurement of the brain tissue microenvironment, including a data processing module, a feature extraction module, a multi-scale fusion module, and a microstructure joint characterization module. The data processing module processes multi-b-value diffusion magnetic resonance imaging data based on spherical averaging technology to obtain the spherical average signal of multi-b values, and models it using an exponential model to obtain the parameterized characterization, and finally normalizes it; the feature extraction module obtains multi-scale structure information through a feature extraction network based on a visual mamba-convolution hybrid architecture; the multi-scale aggregation module is used to aggregate the extracted features to obtain a compact feature representation and enhance the accuracy of microstructure estimation; the microstructure joint characterization module uses the trained over-complete microstructure dictionary to jointly characterize multiple microstructure indexes to obtain 24 microenvironment measurement indexes;

[0103] The system also includes a migration characterization module, which obtains other microstructure parameters, such as the estimation of SANDI indexes, after fine-tuning with a small amount of data.

[0104] Figure 3 This is a performance comparison chart of the method of the present invention with methods such as multi-layer perceptron (MLP) and convolutional neural network (CNN) under different undersampling modes (60 diffusion-weighted images). The absolute error heat map shows that the method of the present invention has the smallest estimation error. The results show that the method of the present invention has good estimation accuracy;

[0105] Figure 4 This is a result chart of the method of the present invention, which is first trained on the dMRI data of dataset 1 (b = 0, 1000, 3000 s / mm2), and then the obtained pre-trained model is used to test the dMRI data of dataset 2 (b = 0, 7000, 1200, 2800 s / mm2). The results show that the estimation accuracy of the method of the present invention on cross-data acquisition protocols is significantly higher than that of the comparative method, indicating that the present invention has the ability to quickly estimate microstructural indices across data acquisition protocols.

[0106] Figure 5 This is a result chart of the method of the present invention, which is first pre-trained on healthy samples of dataset 1, and then tested on brain tumor patients of dataset 2. The results show that the present invention can give high-accuracy microstructural estimations inside the tumor regions of brain tumor patients only using the training of healthy samples.

[0107] Figure 6 This is a comparison of the SANDI index obtained by fine-tuning using the microstructural dictionary pre-trained by the method of the present invention, the index directly obtained by fitting with a low b-value model (benchmark), and the reference high-quality SANDI index. The results show that the method of the present invention supports the accurate estimation of the SANDI microstructural index that usually requires high b-values in the case of only having low b-value DWI data.

[0108] In summary, the present invention can quickly and across data acquisition protocols quantitatively describe the microenvironmental parameters of multiple human brain tissues, promoting the clinical application of dMRI microstructural imaging.

[0109] It should be understood that the above description of the preferred embodiment is relatively detailed, and it should not be considered as a limitation to 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 requested by the present invention shall be subject to the appended claims.

Claims

1. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment, characterized in that, It includes the following steps: Step 1: Obtain the spherical mean signal of multi-b value diffusion magnetic resonance data, perform parameterization and normalization, and construct a data training set; Step 2: Construct a microstructure dictionary including a feature extraction module and a multi-scale aggregation module. The feature extraction module is based on a Vision Mamba-convolution hybrid architecture and is used to extract multi-scale structural information. The multi-scale aggregation module aggregates the extracted multi-scale structural information to obtain a compact feature representation; Step 3: Train the microstructure dictionary with the data training set, and use the trained microstructure dictionary to jointly characterize multiple microstructure metrics to obtain a microenvironment measurement metric.

2. The diffusion magnetic resonance imaging method for multi-scale brain tissue microenvironment measurement according to claim 1, wherein The feature extraction module in Step 2 is composed of a 1×1 convolutional layer and four cascaded Vim-CNN units in series. The convolutional layer uses a 1×1 convolutional kernel to adjust the number of channels of the input image from 3 to 32; the number of input channels of each Vim-CNN module is C, and the number of output channels is 2C.

3. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment according to claim 2, characterized in that, The specific process of the feature extraction module extracting multi-scale structural information is as follows: Let the input diffusion magnetic resonance parameter map be Let H and W be the height and width of the image respectively, and C0 be the number of channels. The 1×1 convolutional layer uses a convolutional kernel and a bias b to map X0, and combines with the ReLU activation function to obtain X 0 ∈R H×W×32 ; The input of the $i$-th Vim-CNN unit is The output is $C$ i are 32, 64, 128, 256 in sequence, where $1\leq i\leq4$, representing the number of input channels of the $i$-th Vim-CNN unit; $X$ i is obtained through the following steps: (1) Dual 3×3 Convolution Branches: Using two 3×3 convolutional kernels and bias p i , combined with ReLU to extract local features; (2) Vision Mamba Branch: Capture global dependencies through the VisionMamba module to obtain Concatenate along the channel dimension with to obtain (3) Single - volume integral branch: Through a 1×1 convolutional kernel and bias b i , combined with the ReLU mapping X i , to obtain (4) Residual fusion:

4. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment according to claim 3, characterized in that, The specific implementation process of the Vision Mamba module is as follows: First, normalize the input feature sequence T l-1 and then project it into x and z, where x is the main input feature after projection of the input sequence, and z is another set of gating control information extracted from the secondary input, which is used to determine how information is fused in the subsequent gating fusion mechanism. For the forward (f, i.e., forward) and backward (b, i.e., backward) d ∈ {f, b}, discretize the parameters by the following formula: where, Δ d is the state step size, controlling the discrete time interval for state update, A d is the state transition matrix, describing the change trend of the state, exp is the exponential matrix operation, used to simulate the evolution of continuous states in the state space, B d is the input control matrix, adjusting the influence of the input on the current state, calculating the hidden state, and iteratively updating the state through the following formula Among them represents the specific directional feature enhanced by one-dimensional convolution is the hidden state at the previous time step is the output of direction d at time step t, C d is the output projection matrix used to project the updated state onto the output space is the updated hidden state at the current time step; Finally, through the state h after the iterative update d the bidirectional output Y is calculated f and T b and adaptively fused through the gating mechanism using z to form the final feature sequence T l : T l = SiLU(z) ⊙ (Y f + Y b ) + T l-1 where SiLU is the Swish activation function, z is the gating information used to control the fusion, and ⊙ is the element-wise multiplication operation.

5. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment according to claim 1, characterized in that, The multi-scale aggregation module in Step 2 is composed of a Vim module, a spatial aggregation module, and a 1×1 convolutional layer in series. The Vim module uses the Vision MambaBlock to perform bidirectional state space modeling on the input feature map, capture global spatio-temporal dependencies, and keep the number of output channels the same as the input; the spatial aggregation module is based on a multi-order gating aggregation structure and adaptively fuses multi-scale context information in the spatial dimension with the number of channels unchanged; The 1×1 convolutional layer uses a convolutional kernel size of 1×1 and is used to compress the feature dimension and enhance the information integration ability.

6. The diffusion magnetic resonance imaging method for multi-scale brain tissue microenvironment measurement according to claim 5, characterized in that, The spatial aggregation module includes a feature decomposition module and a multi-order gating aggregation module. The feature decomposition module uses a 1×1 convolutional operation to decompose the input feature X, and then uses a gating mechanism to re-weight the complementary interaction components. The expression is: Y = Conv 1×1 (X), Z = GELU(Y + γ s ⊙(Y - GAP(Y))) where γ s is a learnable scaling factor, GAP(·) represents the global average pooling operation, and GELU(·) represents the activation function; The multi-stage gated aggregation module employs three parallel deep convolutional layers to encode low-order, middle-order, and high-order features at different dilation rates, and then concatenates the outputs of these layers into a unified multi-stage context Y C = ConcatY l , Y m , Y h ), and fuses the context and input features using the gated mechanism of SiLU: Z = SiLU(Conv 1×1 (X)) ⊙ SiLU(Conv 1×1 (Y C )) Among them, SiLU is the Swish activation function, Concat is the channel merging operation, Y l , Y m , Y h are the low-order, middle-order, and high-order features respectively, and Conv 1×1 is a convolution operation with a 1×1 convolution kernel.

7. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment according to claim 6, characterized in that, The dilation rates of low-order, middle-order, and high-order feature encodings are 1, 2, and 3.

8. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment according to claim 1, characterized in that, The microenvironment measurement metrics include: isotropic volume fraction, neurite density, neurite direction dispersion, free water content, axonal volume fraction, intrinsic diffusion coefficient, mean signal kurtosis, mean signal diffusion, fractional anisotropy, perfusion fraction, axonal water content, hindered axial diffusion coefficient, hindered radial diffusion coefficient, axonal diffusion coefficient, axial kurtosis, mean kurtosis, radial kurtosis, axial diffusion coefficient, mean diffusion coefficient, radial diffusion coefficient, intra-neurite volume fraction, intrinsic diffusion coefficient, mean squared displacement, and return-to-origin probability.

9. A diffusion magnetic resonance imaging method for multi-scale measurement of brain tissue microenvironment according to claim 8, characterized in that, It also includes: When other parameters need to be estimated, use the training data to retrain the microstructure dictionary, and after the training adjustment, transfer the representation to estimate other microstructure parameters.

10. A diffusion magnetic resonance imaging system for implementing the method according to any one of claims 1-9, characterized in that, It includes: A data processing module that obtains the spherical mean signal of multi-b value diffusion magnetic resonance data, performs parameterization and normalization, and constructs a data training set; The micro-structure dictionary training module constructs a micro-structure dictionary including a feature extraction module and a multi-scale aggregation module. The feature extraction module is based on the Vision Mamba-convolution hybrid architecture and is used to extract multi-scale structure information. The multi-scale aggregation module aggregates the extracted multi-scale structure information to obtain a compact feature representation; The joint representation module trains the micro-structure dictionary through a data training set, and uses the trained micro-structure dictionary to jointly represent multiple micro-structure indicators to obtain a micro-environment measurement indicator.

Images