Deep learning-based collaborative dual-path microstructure estimation method and system

Abstract

The invention discloses a deep learning-based collaborative dual-path microstructure estimation method and system, and relates to the technical field of medical image processing. The objective of the invention is to solve the problems that an existing microstructure estimation method based on deep learning has defects in a modeling mechanism, especially lacks an effective information transmission mechanism between an x space and a q space, resulting in insufficient structure expression ability of a model, low estimation precision and the like. According to the technical key points, a collaborative double-path hybrid neural network is designed to enhance feature expression in a heterogeneous domain. By embedding a three-dimensional x space into q space representation, context information of voxels in the x space is considered, so that compatible learning among heterogeneous domains is realized; meanwhile, the qx learning branch transmits angle information of the q space back to the x space, and the feature expression ability of the x space is further improved. And finally, the fusion module based on the attention mechanism further enhances the robustness of the model by adaptively weighting the output of the double paths. A large number of experiments based on HCP data show that the quality of the microstructure index diagram is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, to a method and system for estimating microscopic tissue structure based on a deep learning-based collaborative dual-path approach. Background Technology

[0002] Diffusion magnetic resonance imaging (dMRI) is a non-invasive imaging technique that provides crucial information about the microstructure of brain tissue by measuring signal attenuation caused by the anisotropic diffusion of water molecules within the tissue. dMRI is widely used in tracking white matter fibers, reconstructing brain networks, and diagnosing and researching neurodegenerative diseases.

[0003] To gain a deeper understanding of brain tissue microstructure, several advanced diffusion models have been developed in recent years, such as diffusion kurtosis imaging (DKI), neural directed diffusion and density imaging (NODDI), and spherical mean technique (SMT). These models typically rely on densely sampled multidirectional diffusion data in q-space and employ complex nonlinear processes to fit microscopic parameters. However, these high-density sampling strategies are difficult to apply in clinical practice due to time constraints, patient compliance, and equipment limitations, resulting in a trade-off between accuracy and usability.

[0004] To address these issues, researchers have gradually introduced deep learning methods to attempt to quickly and efficiently estimate microstructural parameters from dMRI data under q-space downsampling conditions (i.e., fewer gradient directions). Although existing methods have made some progress in reconstruction accuracy, most methods still only model in a single spatial dimension, such as capturing diffusion angle features only from q-space, ignoring the spatial dependencies between voxels in x-space. Even the hybrid modeling frameworks proposed in recent years—HGT (Hybrid Graph Transformer) and 3D-HGT—only perform unidirectional feature transfer from q-space to x-space, failing to establish a bidirectional collaborative mechanism between space and diffusion domain, thus limiting the model's ability to comprehensively express structural and diffusion information in the heterogeneous joint space (xq-space).

[0005] For example, document CN115775626A discloses a multimodal brain network computing method, device, equipment, and storage medium. This method is used to train a brain disease prediction model. It extracts brain region structural features and brain region functional features from magnetic resonance diffusion tensor imaging data and functional magnetic resonance imaging data, respectively. Then, it uses a graph representation diffusion learning network to separate common and unique features from these features. Finally, based on alignment algorithms and adaptive weighting techniques, it effectively fuses the common and unique features of the multimodal data, fully exploring the complementary information between the multimodal data. This allows the model to learn effective disease-related features during training, resulting in a more accurate and better predictive model for brain diseases. However, this existing technology does not mention the modeling mechanism, information transmission mechanism, or model structure representation.

[0006] In summary, although existing artificial intelligence technologies have made some progress in estimating microscopic tissue structures, their shortcomings in accuracy and generalization ability prevent them from fully meeting actual clinical needs. Therefore, there is an urgent need for a new estimation model that can fully utilize information from both the x-space and q-space under fewer gradient conditions to achieve high-quality NODDI estimation. Summary of the Invention

[0007] The technical problem to be solved by this invention is:

[0008] Existing deep learning-based methods for estimating microstructure still have shortcomings in their modeling mechanisms, especially the lack of an effective information transfer mechanism between the x-space and q-space. This results in insufficient ability of the model to express the structure, low estimation accuracy, and limited robustness, making it difficult to meet the clinical application needs in low-sampling and highly complex structural scenarios.

[0009] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0010] This invention provides a deep learning-based collaborative dual-path microstructure estimation method, comprising the following steps:

[0011] Step 1: Data Acquisition and Gold Standard Generation. Download the publicly available brain dMRI dataset from the Human Connectome Project (HCP) and use the AMICO tool to generate ICVF (Neuron Density Index), ISOVF (Isotropic Volume Fraction), and ODI (Neuron Orientation Distribution Index) to construct the NODDI gold standard dataset;

[0012] Step 2: Construct a brain white matter microstructure estimation model based on dual-path heterogeneous spatial modeling, including an xq space learning module, a qx space learning module, and a feature fusion module. The xq space learning module uses a 3D convolutional neural network to extract spatial context features in x space and constructs a q space graph structure based on angle correlation. Angle dependency modeling is then performed in q space using a graph convolutional neural network (ChebNet), thereby achieving feature embedding from spatial structure to diffuse structure. The qx space learning module constructs a graph neural network using q space direction signals. After extracting angle features through graph convolution, it uses the Swin Transformer V2 module to achieve long-range dependency modeling in 3D space, and reversely realizes semantic transfer and feature enhancement from q space to x space. The fusion module integrates the heterogeneous information extracted from the xq and qx paths and generates the final microstructure parameters through convolution operations and linear mapping.

[0013] Step 3: The normalized dMRI data and the corresponding NODDI gold standard index are used to form training pairs. A 16×16×16 patch is extracted using a sliding window method to train the model. Mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for parameter updates. During training, the model's performance on the independent validation set is simultaneously verified to ensure its fitting and generalization abilities.

[0014] Step 4: Input the dMRI image data to be estimated into the trained microstructure estimation model, extract heterogeneous features via the xq and qx paths respectively, and output a microstructure parameter map after fusion. The output includes key indicators such as ICVF (neuron density index), ISOVF (isotropic volume fraction), and ODI (neuronal orientation distribution index), forming a complete microstructure estimation map.

[0015] Step 5: Quantitatively compare the predicted microstructure maps with the NODDI gold standard. Use Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) to evaluate model performance. Simultaneously, combine brain region masking for quantitative analysis of local structures to verify the model's ability to represent complex brain structures.

[0016] If the evaluation results show that the model is inadequate in specific indicators or local brain regions, the network structure or feature fusion strategy is adjusted accordingly, and the model is retrained and validated after improvement to continuously improve the ability of the microstructure estimation model to represent complex brain tissue microstructures until it meets expectations.

[0017] During the retraining process, the performance of the microstructure estimation model on the validation set is continuously evaluated through multiple rounds of iteration, and the parameters of the best-performing model are saved. Finally, the result of the best microstructure estimation model on the test set is used as the final output.

[0018] Furthermore, step one includes descrambling and normalizing the acquired diffusion magnetic resonance (dMRI) data, and using a sliding window method to divide the data into three-dimensional patches of fixed size 16×16×16.

[0019] Furthermore, the mean squared error is used as the loss function, and gradient descent is combined to train the constructed collaborative dual-path model to achieve efficient estimation of the microstructure of brain white matter. The training loss function... Includes branch loss term With fusion loss term The specific expression is:

[0020] Among them, the fusion loss term is , representing the mean squared error loss between the final fused output Y of the model and the true value GT of the microstructure; the branch loss term is:

[0021]

[0022] Where L(.) represents the mean squared error loss function, These represent the prediction results output by the xq branch and the qx branch, respectively. The weighting factor 'a' is used to adjust the weight of the branch loss in the total loss and is manually set according to experimental requirements; in this study, it is set to 0.5.

[0023] Furthermore, regarding the construction of the adjacency matrix for graph convolution in step two, specifically, this invention uses a graph structure G to represent the geometric structure of the q-space in dMRI data. This graph is composed of a binary affinity matrix A = { } defines each element This represents the affinity relationship between two q-space data points with indices i and j. The affinity matrix B is calculated as follows:

[0024] Furthermore, the functional implementation process of the fusion enhancement module is as follows:

[0025] First, regarding the two features mentioned above... Perform an averaging operation to obtain its average characteristics. Given that different features contribute differently to the final result, this paper introduces an attention mechanism to measure the contribution weight of different features at each voxel level. Specifically, to generate an attention map for a specific feature, the attention map is... and and The concatenated features are then fed into a convolutional layer to generate the initial attention weights.

[0026] in, This indicates a convolutional layer targeting this feature. The parameters of this convolutional layer are represented by the Sigmoid activation function. In other words, for each feature, this invention designs a separate convolutional layer. Different attention weights are generated. To ensure that the fused features have a consistent numerical range, the generated attention weights are normalized using the Softmax function.

[0027] The attention weights are multiplied element-wise with their corresponding feature maps, and finally, the adaptively fused feature maps are obtained by weighted summation of the individual feature maps. :

[0028]

[0029] Furthermore, in step two, the feature extraction in the x-space from the qx module employs a densely connected SwinTransformer V2 (referred to as the Swin V2 Block) as the backbone network. This Swin V2 Block consists of three densely connected Swin Transformer V2 layers. The embedding layer of this invention comprises two 3D convolutional neural networks (3D CNNs), followed by a GELU activation function layer and a normalization layer, where the kernel size and stride are both set to 1. The embedding dimension is set to 96, and the window size is set to 4.

[0030] Furthermore, in step two, the graph convolutional neural network is constructed using three-dimensional patches as nodes. These nodes represent patches with fixed spatial dimensions in the input image, rather than single pixels, thereby encoding contextual information from the x space in the q space.

[0031] Furthermore, the graph convolutional network employs the Chebyshev spectral graph convolution method, using a k-order Chebyshev polynomial to approximate the normalized Laplacian matrix, the expression of which is:

[0032]

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This invention designs a collaborative dual-path hybrid neural network to enhance feature representation in heterogeneous domains. By embedding the 3D x-space into the q-space representation and considering the contextual information of voxels in the x-space, the collaborative dual-path hybrid neural network achieves compatible learning between heterogeneous domains. Simultaneously, the qx learning branch transmits the angular information from the q-space back to the x-space, further enhancing the feature representation capability of the x-space. Finally, an attention-based fusion module further enhances the robustness of the model by adaptively weighting the outputs of the dual paths. The method described in this invention does not rely on high-density sampling conditions and can achieve high-quality reconstruction of the NODDI index with fewer directions. Extensive experiments based on HCP data demonstrate that this invention significantly improves the quality of microstructure index maps.

[0035] This invention effectively solves the defects in the modeling mechanism of existing deep learning-based microstructure estimation methods. It can effectively transfer information between x-space and q-space, greatly improve the model's ability to express the structure, improve estimation accuracy and robustness, and fully meet the clinical application needs in low-sampling and highly complex structural scenarios. Attached Figure Description

[0037] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0038] Figure 1 This is a diagram of the overall structure of the cooperative dual-path microstructure estimation network constructed in this invention.

[0039] Figure 2 This refers to the dense Swing V2 Block in the qx module of this invention;

[0040] Figure 3 This is a schematic diagram of the attention weight generation and weighting process in the feature fusion module of this invention;

[0041] Figure 4 This is a comparison chart of the model estimation output of this invention (under 30 gradients) and the NODDI gold standard;

[0042] Figure 5 This is a comparison chart of the model estimation output of this invention (under 60 gradients) and the NODDI gold standard. Detailed Implementation

[0044] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with the accompanying drawings. Figure 1-5 Specific embodiments of the present invention will be described in detail below.

[0046] Specific Implementation Plan 1: (e.g.) Figures 1 to 3 As shown, the present invention provides a method for estimating the microstructure of a collaborative dual-pathway microstructure, comprising the following steps:

[0047] Step 1: Data Acquisition and Gold Standard Generation. Download the publicly available brain dMRI dataset from the Human Connectome Project (HCP) and use the AMICO tool to generate ICVF (Neuron Density Index), ISOVF (Isotropic Volume Fraction), and ODI (Neuron Orientation Distribution Index) to construct the NODDI gold standard dataset;

[0048] Step 2: Construct a brain white matter microstructure estimation model based on dual-path heterogeneous spatial modeling, including an xq space learning module, a qx space learning module, and a feature fusion module. The xq space learning module uses a 3D convolutional neural network to extract spatial context features in x space and constructs a q space graph structure based on angle correlation. Angle dependency modeling is then performed in q space using a graph convolutional neural network (ChebNet), thereby achieving feature embedding from spatial structure to diffuse structure. The qx space learning module constructs a graph neural network using q space direction signals. After extracting angle features through graph convolution, it uses the Swin Transformer V2 module to achieve long-range dependency modeling in 3D space, and reversely realizes semantic transfer and feature enhancement from q space to x space. The fusion module integrates the heterogeneous information extracted from the xq and qx paths and generates the final microstructure parameters through convolution operations and linear mapping.

[0049] Step 3: The normalized dMRI data and the corresponding NODDI gold standard index are used to form training pairs. A 16×16×16 patch is extracted using a sliding window method to train the model. Mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for parameter updates. During training, the model's performance on the independent validation set is simultaneously verified to ensure its fitting and generalization abilities.

[0050] Step 4: Input the dMRI image data to be estimated into the trained microstructure estimation model, extract heterogeneous features via the xq and qx paths respectively, and output a microstructure parameter map after fusion. The output includes key indicators such as ICVF (neuron density index), ISOVF (isotropic volume fraction), and ODI (neuronal orientation distribution index), forming a complete microstructure estimation map. Step 5: Quantitatively compare the predicted microstructure maps with the NODDI gold standard, and evaluate model performance using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Simultaneously, perform quantitative analysis of local structures using brain region masks to verify the model's ability to represent complex brain structures.

[0051] If the evaluation results show that the model is inadequate in specific indicators or local brain regions, the network structure or feature fusion strategy is adjusted accordingly, and the model is retrained and validated after improvement to continuously improve the ability of the microstructure estimation model to represent complex brain tissue microstructures until it meets expectations.

[0052] During the retraining process, the performance of the microstructure estimation model on the validation set is continuously evaluated through multiple rounds of iteration, and the parameters of the best-performing model are saved. Finally, the result of the best microstructure estimation model on the test set is used as the final output.

[0053] Step one includes descrambling and normalizing the acquired diffusion magnetic resonance (dMRI) data, and then using a sliding window method to divide the data into three-dimensional patches of fixed size 16×16×16.

[0054] Step three uses mean squared error as the loss function and combines it with gradient descent to train the constructed collaborative dual-path model, thereby achieving efficient estimation of the microstructure of brain white matter. The training loss function... Includes branch loss term With fusion loss term The specific expression is:

[0055]

[0056] Among them, the fusion loss term is , representing the mean squared error loss between the final fused output Y of the model and the true value GT of the microstructure; the branch loss term is:

[0057]

[0058] Where L(.) represents the mean squared error loss function, These represent the prediction results output by the xq branch and the qx branch, respectively. The weighting factor 'a' is used to adjust the weight of the branch loss in the total loss and is manually set according to experimental requirements; in this study, it is set to 0.5.

[0059] The construction of the adjacency matrix for graph convolution in step two is defined by the included angle. Specifically, this invention uses a graph structure G to represent the geometric structure of the q-space in dMRI data. This graph consists of a binary affinity matrix A = { } defines each element This represents the affinity relationship between two q-space data points with indices i and j. The affinity matrix A is calculated as follows:

[0060] The functional implementation process of the fusion module is as follows:

[0061] First, regarding the two features mentioned above... Perform an averaging operation to obtain its average characteristics. Given that different features contribute differently to the final result, this paper introduces an attention mechanism to measure the contribution weight of different features at each voxel level. Specifically, to generate an attention map for a specific feature, the attention map is... and and The concatenated features are then fed into a convolutional layer to generate the initial attention weights.

[0062]

[0063] in, This indicates a convolutional layer targeting this feature. The parameters of this convolutional layer are represented by the Sigmoid activation function. In other words, for each feature, this invention designs a separate convolutional layer. Different attention weights are generated. To ensure that the fused features have a consistent numerical range, the generated attention weights are normalized using the Softmax function.

[0064] The attention weights are multiplied element-wise with their corresponding feature maps, and finally, the adaptively fused feature maps are obtained by weighted summation of the individual feature maps. :

[0065]

[0066] The method for estimating microstructure using a collaborative dual-path approach according to claim 1 is characterized in that, in step 2, the feature extraction of the x-space from the qx module employs a densely connected Swin Transformer V2 (referred to as the Swin V2 Block) as the backbone network. This Swin V2 Block consists of three densely connected Swin Transformer V2s. The embedding layer of this invention consists of two 3D convolutional neural networks (3D CNNs), followed by a GELU activation function layer and a normalization layer, wherein the kernel size and stride are both set to 1. The embedding dimension is set to 96, and the window size is set to 4.

[0067] In step two, the graph convolutional neural network is constructed using three-dimensional patches as nodes. These nodes represent patches with fixed spatial dimensions in the input image, rather than individual pixels, thereby encoding contextual information from the x space in the q space.

[0068] The graph convolutional network employs the Chebyshev spectral graph convolution method, using a K-order Chebyshev polynomial to approximate the normalized Laplacian matrix. Its expression is:

[0069]

[0070] Example:

[0071] The visualization results of this invention are based on the HCP dataset, which includes 28 subjects, of whom 18 were used for training, 2 for validation, and 8 for testing. Figure 4 and Figure 5 The results shown are all from the test set. For ease of comparison, we selected brain regions from the test set where the predicted values ​​differed significantly from the actual values ​​to highlight the performance differences between different methods in microstructure estimation. We compared our method with current state-of-the-art models, including MLP, GCNN, MEDN, MEDN++, UNet, UNet++, CNN, HGT, and 3D-HGT. All experiments were conducted under identical conditions. All images were presented under the same slice location, intensity range, and color scale conditions to ensure comparability between methods. Visualization was performed using Mango to visually demonstrate the differences between the models in microstructure estimation. Figure 4 The first line gives the prediction results for each indicator, and the second line shows the difference between the predicted value and the actual value; Figure 5The focus is then on presenting the error plot. The error plot shows that the prediction error of this invention is significantly lower than that of other methods, indicating that its prediction results are more accurate. Combining quantitative and qualitative analysis, this invention outperforms the comparative model in overall performance. Based on the method proposed in this invention, a collaborative dual-path microstructure estimation system is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the collaborative dual-path microstructure estimation method during runtime.

[0072] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the above-described steps of cooperative dual-path microstructure estimation when invoked by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0073] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0074] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0075] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for estimating microorganisms based on deep learning through collaborative dual-path approaches, characterized in that, Includes the following steps: Step 1: Data Acquisition and Gold Standard Generation: Download the publicly available brain dMRI dataset from the Human Connectome Project (HCP), and use the AMICO tool to generate three microstructure indicators: neuronal density index ICVF, isotropic volume fraction ISOVF, and neuronal orientation distribution index ODI, to construct the NODDI gold standard dataset. Step 2: Construct a model based on collaborative dual-path microstructure estimation, including an xq-space learning module, a qx-space learning module, and a feature fusion module. The xq-space learning module uses a 3D convolutional neural network to extract spatial context features in x-space and constructs a q-space graph structure based on angle correlation. Angle dependency modeling is then performed in q-space using a graph convolutional neural network (ChebNet), thereby achieving feature embedding from spatial structure information to diffuse angle structure. The qx-space learning module constructs a graph neural network using q-space direction signals. After extracting angle features through graph convolution, it uses the Swin Transformer V2 module to achieve long-range dependency modeling in 3D space, and reversely realizes semantic transfer and feature enhancement from q-space angle information to x-space. The fusion module integrates the heterogeneous information extracted from the xq and qx paths, introduces an attention mechanism, performs weighted integration of dual-path information, and generates the final microstructure parameters through convolution operations and linear mapping. Step 3: The normalized dMRI data and the corresponding NODDI gold standard index are used to form training pairs to train the model based on collaborative dual-path microstructure estimation constructed in Step 2. A 16×16×16 patch is extracted using a sliding window method to train the model. Mean squared error (MSE) is used as the loss function, and the Adam optimizer is used to update the parameters. During the training process, the performance of the model on the independent validation set is verified simultaneously to ensure the model's fitting ability and generalization ability. Step 4: Input the dMRI image data to be estimated into the trained microstructure estimation model. Extract heterogeneous features via the xq and qx paths respectively, fuse them, and output a microstructure parameter map. The output includes key indicators such as neuron density index ICVF, isotropic volume fraction ISOVF, and neuron orientation distribution index ODI, forming a complete microstructure estimation map. Step 5: Quantitatively compare the predicted microstructure map with the NODDI gold standard. Use peak signal-to-noise ratio PSNR and structural similarity index SSIM to evaluate model performance. Simultaneously, combine brain region masking for quantitative analysis of local structures to verify the model's ability to express complex brain region structures. If the evaluation results show that the model is inadequate in specific indicators or local brain regions, the network structure or feature fusion strategy is adjusted accordingly, and the model is retrained and validated after improvement to continuously improve the ability of the microstructure estimation model to represent complex brain tissue microstructures until it meets expectations. During the retraining process, the performance of the microstructure estimation model on the validation set is continuously evaluated through multiple rounds of iteration, and the parameters of the best-performing model are saved. Finally, the result of the best microstructure estimation model on the test set is used as the final output.

2. The method for estimating microstructure using a synergistic dual-path approach according to claim 1, characterized in that, Step one includes descrambling and normalizing the acquired diffusion magnetic resonance (dMRI) data, and then using a sliding window method to divide the data into three-dimensional patches of fixed size 16×16×16.

3. The method for estimating microstructure using a cooperative dual-path approach according to claim 1 or 2, characterized in that, Step three uses mean squared error as the loss function and combines it with gradient descent to train the constructed collaborative dual-path model, thereby achieving efficient estimation of the microstructure of brain white matter. The training loss function... Includes branch loss term With fusion loss term The specific expression is: ； Among them, the fusion loss term is , representing the mean squared error loss between the final fused output Y of the model and the true value GT of the microstructure; the branch loss term is: ， Where L(.) represents the mean squared error loss function, These represent the prediction results output by the xq branch and the qx branch, respectively. The weighting factor 'a' is used to adjust the weight of the branch loss in the total loss and is manually set according to experimental requirements; in this study, it is set to 0.

5.

4. The method for estimating microstructure using a synergistic dual-path approach according to claim 3, characterized in that, The adjacency matrix of the graph convolution described in step two is defined by the included angle. A graph structure G is used to represent the geometric structure of the q-space in the dMRI data. This graph structure G consists of a binary affinity matrix A = { } defines each element The affinity matrix A represents the affinity relationship between two q-space data points with indices i and j. The affinity matrix A is calculated as follows: .

5. The method for estimating microstructure using a synergistic dual-path approach according to claim 1 or 4, characterized in that, The functional implementation process of the fusion module is as follows: First, regarding the two features mentioned above... Perform an averaging operation to obtain its average characteristics. Given that different features contribute differently to the final result, an attention mechanism is introduced to measure the contribution weight of different features at each voxel level. To generate an attention map for a specific feature, the attention map is... and and The concatenated features are then fed into a convolutional layer to generate the initial attention weights. ； in, This indicates a convolutional layer targeting this feature. This represents the parameters of the convolutional layer. The activation function is Sigmoid; separate convolutional layers are designed for each feature. Different attention weights are generated; to ensure that the fused features have a consistent numerical range, the generated attention weights are normalized using the Softmax function. ; The attention weights are multiplied element-wise with their corresponding feature maps, and finally, the adaptively fused feature maps are obtained by weighted summation of the individual feature maps. : 。 6. The method for estimating microstructure using a synergistic dual-path approach according to claim 5, characterized in that, In step 2, the features in the x space extracted by the qx module are used as the backbone network of a densely connected Swin Transformer V2, which is called the Swin V2 Block. The Swin V2 Block consists of three densely connected Swin Transformer V2s. The embedding layer consists of two 3D convolutional neural networks (3D CNNs), followed by a GELU activation function layer and a normalization layer. The kernel size and stride are both set to 1, the embedding dimension is set to 96, and the window size is set to 4.

7. The method for estimating microstructure using a synergistic dual-path approach according to claim 6, characterized in that, In the graph convolutional network, graph convolution is constructed using three-dimensional patches as nodes. These nodes represent patches with fixed spatial dimensions in the input image, rather than individual pixels, thereby encoding contextual information from the x space in the q space.

8. The method for estimating microstructure using a synergistic dual-path approach according to claim 1, characterized in that, The graph convolutional network employs the Chebyshev spectral graph convolution method, using a k-order Chebyshev polynomial to approximate the normalized Laplacian matrix. Its expression is: ； Where L represents the normalized Laplacian matrix, For the k-th order Chebyshev polynomial, Here, x is the learning parameter, and x is the input signal.

9. A system for collaborative dual-path microstructure estimation based on deep learning, characterized in that: The system has a program module corresponding to the steps of any one of claims 1-8, and executes the steps in the method for estimating a collaborative dual-path microstructure based on deep learning when it is run.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the method for deep learning-based collaborative dual-path microstructure estimation as described in any one of claims 1-8.

Images