A method for reconstructing the directional distribution of white matter nerve fibers in the brain and related apparatus
Patent Information
- Application Number
- CN202311697227.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-11
- Publication Date
- 2026-10-09
- Estimated Expiration
- 2043-12-11
AI Technical Summary
[0003]本发明的目的在于提供一种重建脑白质神经纤维方向分布的方法及相关装置，以解决临床环境中获取到的往往是单壳低角分辨率数据，MSMT-CSD并不适用于此类数据的问题
[0029]This invention utilizes deep learning technology to reconstruct fiber bundle orientation distribution, achieving higher accuracy compared to traditional methods. It fully considers the spatial relationships of voxels, enabling high-quality reconstruction of fiber bundle orientation distribution even using low-angular-resolution diffusion imaging data. Furthermore, the fiber bundle orientation distribution reconstruction network is entirely based on multilayer perceptrons, achieving excellent performance with a relatively small number of parameters. Additionally, the invention employs 3D patches as input to the fiber bundle reconstruction network, allowing for flexible handling of data acquired using different protocols.
Smart Images
Figure CN117671153B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of magnetic resonance medical image processing, and specifically relates to a method and related apparatus for reconstructing the directional distribution of nerve fibers in the white matter of the brain. Background Technology
[0002] Diffusion magnetic resonance imaging (dMRI) is a method that uses magnetic resonance imaging to measure the movement and diffusion of water molecules in biological tissues, providing a non-invasive way to study white matter fiber bundle structure. The fiber orientation distribution function (fODF / FOD) estimated from dMRI data can be used to interpret the structure of nerve fibers in the brain. Diffusion tensor imaging (DTI) is often used to estimate local fODF, but this method only considers the isotropic diffusion of water molecules and cannot explain complex fiber structures (such as multiple fiber bundle crossings, bends, and kissing). To overcome the limitations of DTI, high angular resolution diffusion imaging (HARDI) is used to estimate fODF. HARDI uses multi-b-value, multi-gradient directional pulse sequences to acquire diffuse signals, providing detailed fiber bundle orientation information. Based on HARDI data, the Multi-Shell Multi-Tissue Constrained Spherical Deconvolution (MSMT-CSD) method can accurately calculate fODF, thus resolving complex fibrous structures. However, due to limitations in scanner acquisition protocols and acquisition time, clinical settings often yield single-shell, low-angular-resolution data, making MSMT-CSD unsuitable for such data. Therefore, an effective tool for accurately estimating fODF from LARDI data remains lacking in clinical practice, and reconstructing high-quality FOD from clinically relevant LARDI data is a crucial area for further research. Summary of the Invention
[0003] The purpose of this invention is to provide a method and related apparatus for reconstructing the directional distribution of nerve fibers in brain white matter, in order to solve the problem that clinical data often consists of single-shell low-angular-resolution data, which MSMT-CSD is not suitable for.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] In a first aspect, the present invention provides a method for reconstructing the directional distribution of nerve fibers in the white matter of the brain, comprising:
[0006] Acquired and preprocessed diffusion magnetic resonance imaging (dMRI) data, calculated the orientation distribution function (FOD) image of the single-shell fiber bundle, and divided it into 3D patches.
[0007] A FOD image feature extraction model is constructed, and FOD patch data is used as input to the model for preliminary feature extraction.
[0008] Construct a 3D coordinate attention model, using the initially extracted features as input, to extract the spatial positional relationships of the input features;
[0009] A spherical harmonic coefficient estimation model is constructed, which takes features containing spatial location relationships as input and outputs the FOD after 3DPatch center voxel reconstruction.
[0010] The FOD image feature extraction model, the 3D coordinate attention model, and the spherical harmonic coefficient estimation model are cascaded together to form a complete FOD reconstruction network. The complete model is trained, and the trained model is used to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data.
[0011] Optionally, acquire diffusion magnetic resonance imaging (dMRI) data and perform preprocessing:
[0012] The required training and testing data were obtained from the HCP dataset published by the Human Connectome Project (HCP). Multi-shell FOD was calculated using a multi-shell, multi-organism constrained spherical deconvolution algorithm; this data served as the "ground truth" for supervised model training. The original data was downsampled, starting from b = 0 s / mm. 2 The shell and b = 1000s / mm 2 The shell extracts the first 1 direction and the first 32 directions, and then merges them into a LARDI dataset. Based on the downsampled data, the single-shell FOD is calculated using the single-shell three-organization constrained spherical deconvolution algorithm. This data is used as the training set for the model. When training the model, a 9×9×9 3D patch needs to be divided for each voxel in the training set as the input of the FOD reconstruction network.
[0013] Optionally, for clinical quality dMRI data, the dMRI data first needs to be denoised, artifact removed, motion corrected, distortion corrected, and eddy current corrected. Then, the single-shell three-tissue constrained spherical deconvolution algorithm is used to calculate the single-shell FOD image from the processed data. When reconstructing the single-shell FOD into a high-quality FOD, a 9×9×9 3DPatch needs to be partitioned for each voxel of the single-shell FOD as the input of the FOD reconstruction network.
[0014] Optionally, a FOD image feature extraction model can be constructed, using FOD patch data as input to perform preliminary feature extraction:
[0015] The FOD image feature extraction module consists of five multilayer perceptron (MLP)-mixer modules connected in series, with a skip connection between the modules. Each MLP-mixer module consists of two components connected in series: a token-mixing component and a channel-mixing component. Each component is preceded by a LayerNorm normalization layer and a skip connection. The token-mixing component is used to extract features between different voxels, and the channel-mixing component is used to extract features between different spherical harmonic coefficients (SHCs). This module is used to initially extract features from the input data, and the output preliminary features are used as input to the next module.
[0016] Optionally, a 3D coordinate attention model can be constructed, using the initially extracted features as input, to extract the spatial relationships of the input features:
[0017] The 3D coordinate attention module consists of a pooling layer, an MLP layer, an activation function layer, and a fusion layer. The pooling layer compresses global information into different dimensional directions. The MLP encodes the local spatial position information and channel information of voxels. The activation function layer calculates the weight coefficients for different dimensional directions. The fusion layer performs global information interaction and is used to further extract the spatial positional relationships of voxels. The new features containing voxel spatial position information are used as the input of the next module.
[0018] Optionally, a spherical harmonic coefficient estimation model is constructed, taking features containing spatial location relationships as input, and outputting the FOD after 3DPatch center voxel reconstruction:
[0019] The spherical harmonic coefficient estimation model is implemented by adaptive pooling operations. It uses attention-based pooling operations instead of global average pooling operations. By calculating the similarity between the features of the central voxel and all features, the softmax function is used to assign different weight coefficients to different voxels. During the pooling operation, the reconstructed FOD is obtained by weighting and summing according to the corresponding weights.
[0020] Optionally, a complete FOD reconstruction network can be constructed by cascading the data, training a complete model, and using the trained model to estimate a high-quality fiber bundle orientation distribution function (FOD) from LADRI data.
[0021] Single-shell FOD is used as the training set for the model, while multi-shell FOD is used as the "ground truth" for supervised model training. The model only inputs a 3D patch from a single-shell FOD at a time and outputs the version after central voxel reconstruction. The loss function is calculated between the reconstructed voxel and the corresponding "ground truth". By minimizing the loss function, the parameter weights of the FOD reconstruction network are adjusted to improve the model performance.
[0022] In a second aspect, the present invention provides a system for reconstructing the directional distribution of white matter fibers in the brain, comprising:
[0023] The data receiving module is used to acquire diffusion magnetic resonance imaging (dMRI) data and perform preprocessing, calculate the orientation distribution function (FOD) image of the single-shell fiber bundle and divide it into 3D patches;
[0024] The model building module is used to build a FOD image feature extraction model, which takes FOD patch data as input to the model for preliminary feature extraction; build a 3D coordinate attention model, which takes the preliminarily extracted features as input to extract the spatial positional relationship of the input features; and build a spherical harmonic coefficient estimation model, which takes features containing spatial positional relationships as input and outputs the FOD after reconstruction of the 3D patch center voxel.
[0025] The FOD reconstruction module is used to connect the FOD image feature extraction model, the 3D coordinate attention model, and the spherical harmonic coefficient estimation model to form a complete FOD reconstruction network, train the complete model, and use the trained model to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data.
[0026] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a method for reconstructing the directional distribution of white matter fibers in the brain.
[0027] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a method for reconstructing the directional distribution of white matter fibers in the brain.
[0028] Compared with the prior art, the present invention has the following technical effects:
[0029] This invention utilizes deep learning technology to reconstruct fiber bundle orientation distribution, achieving higher accuracy compared to traditional methods. It fully considers the spatial relationships of voxels, enabling high-quality reconstruction of fiber bundle orientation distribution even using low-angular-resolution diffusion imaging data. Furthermore, the fiber bundle orientation distribution reconstruction network is entirely based on multilayer perceptrons, achieving excellent performance with a relatively small number of parameters. Additionally, the invention employs 3D patches as input to the fiber bundle reconstruction network, allowing for flexible handling of data acquired using different protocols. Attached Figure Description
[0030] Figure 1 The overall logic block diagram of this invention.
[0031] Figure 2 A schematic diagram of the overall network structure of this invention.
[0032] Figure 3 A schematic diagram of the FOD feature extraction module of this invention.
[0033] Figure 4 A schematic diagram of the MLP structure used in this invention.
[0034] Figure 5 A schematic diagram of the 3D coordinate attention module of this invention.
[0035] Figure 6 A schematic diagram of the attention-based pooling module of this invention.
[0036] Figure 7 A schematic diagram comparing the fiber bundle orientation distribution reconstructed by this invention with that of other methods.
[0037] Figure 8 Flowchart of this invention. Detailed Implementation
[0038] This section will describe exemplary embodiments in detail, examples of which are illustrated in the accompanying drawings. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.
[0039] Please see Figures 1 to 8 A method for reconstructing the directional distribution of nerve fibers in brain white matter, comprising:
[0040] Acquired and preprocessed diffusion magnetic resonance imaging (dMRI) data, calculated the orientation distribution function (FOD) image of the single-shell fiber bundle, and divided it into 3D patches.
[0041] A FOD image feature extraction model is constructed, and FOD patch data is used as input to the model for preliminary feature extraction.
[0042] Construct a 3D coordinate attention model, using the initially extracted features as input, to extract the spatial positional relationships of the input features;
[0043] A spherical harmonic coefficient estimation model is constructed, which takes features containing spatial location relationships as input and outputs the FOD after 3DPatch center voxel reconstruction.
[0044] The FOD image feature extraction model, the 3D coordinate attention model, and the spherical harmonic coefficient estimation model are cascaded together to form a complete FOD reconstruction network. The complete model is trained, and the trained model is used to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data.
[0045] Specifically:
[0046] An example of reconstructing the fiber bundle orientation distribution according to the present invention, the overall process is as follows: Figure 1 As shown, the overall architecture of the model is as follows: Figure 2 As shown, the steps are as follows:
[0047] Data preprocessing:
[0048] The dataset used for training the model was downloaded from the HCP dataset, publicly available from the Human Connectome Project (HCP). The HCP dataset contains data acquired using a 1.25mm × 1.25mm × 1.25mm 3T scanner. The acquired data underwent a rigorous data preprocessing process to provide high-quality, highly consistent, and highly comparable brain imaging data, making it ideal as both a training and testing set for the model. Each subject's data consists of three shells, with b = 1000 s / mm. 2 b = 2000s / mm 2 b = 3000s / mm 2 Each shell sampled signals in 90 gradient directions. Additionally, at b = 0 s / mm... 2 The shell also collected signals in 18 gradient directions. Therefore, the HARDI data size for each subject was 145×174×145×288. Based on this data, the multi-shell FOD calculated using MSMT-CSD can be used as the "ground truth" for supervised model training. Furthermore, to train the model, a clinical-quality LARDI dataset needs to be simulated. The LARDI data is downsampled from the HCP's HARDI data. The specific downsampling process is as follows: first, from b = 0 s / mm... 2The signal in the first gradient direction is extracted from the shell, and then from b = 1000 s / mm 2 The signals from the first 32 gradient directions of the shell are extracted, and the two sets of signals are merged to form a new LARDI dataset. Based on this data, the single-shell FOD is calculated using SS3T-CSD and can be used as the training set for the model. When training the model, the single-shell FOD needs to be divided into 9×9×9 3D patches as input to the FOD reconstruction network.
[0049] For clinical-quality dMRI data, the first step is to perform denoising, artifact removal, motion correction, distortion correction, and eddy current correction. Then, a single-shell three-tissue constrained spherical deconvolution algorithm is used to calculate the single-shell FOD image from the processed data. When reconstructing the single-shell FOD into a high-quality FOD, a 9×9×9 3D patch needs to be defined for each voxel of the single-shell FOD as the input to the FOD reconstruction network.
[0050] Constructing a FOD feature extraction module:
[0051] FOD feature extraction module, such as Figure 3 As shown in the diagram, this module consists of five MLP-Mixer modules chained together, with a jump connection between them. Each MLP-Mixer module comprises two components chained together: a token-mixing component and a channel-mixing component. Each component is preceded by a LayerNorm layer and a jump connection. The token-mixing component extracts features between different voxels, the channel-mixing component extracts features between different SHCs, and the jump connection is used to address the gradient vanishing problem. The structure of the MLP is as follows: Figure 4 As shown, it consists of two fully connected (FC) layers, with a GELU activation function between them. The FOD feature extraction model receives sequential data; the 3D patch data needs to be re-stretched into sequential data before being input into the model. Each voxel corresponds to a token, and the SHC of the voxel represents the original encoded features of the token. The input and output sizes of each MLP-Mixer module remain unchanged; therefore, the input and output sizes of the entire FOD feature extraction model also remain unchanged. Specifically, the token-mixing component uses an MLP with a 729 input layer, 1024 hidden layers, and a 729 output layer; the channel-mixing component uses an MLP with a 45 input layer, 1024 hidden layers, and a 45 output layer. The FOD feature extraction module is used to initially extract features from the input data, and the output preliminary features serve as the input for the next module.
[0052] Constructing a 3D coordinate attention module:
[0053] 3D coordinate attention module such as Figure 5 As shown, this module consists of a pooling layer, an MLP layer, an activation function layer, and a fusion layer. The pooling layer performs average pooling operations in different dimensions, compressing cross-channel information and spatial location into the x, y, and z directions. Each direction yields new features that contain local channel and spatial information. The MLP layer performs channel encoding and position encoding on the three sets of features respectively. The activation function layer calculates weight coefficients for different dimensions and channel directions based on the encoded features. The fusion layer multiplies the three sets of weight coefficients with the input to obtain new features with global position and channel information. This module is also an MLP-based model, and the MLP structure is consistent with the MLP structure used in the feature extraction model. Since the 3D coordinate attention model accepts 3D data, the feature sequence output by the feature extraction model needs to be reconstructed into a 3D patch, maintaining consistent input and output sizes. The three MLP structures used for channel encoding and position encoding are identical, with an input layer size of 729, a hidden layer size of 1024, and an output layer size of 1024. The 3D coordinate attention module is used by this module to further extract the spatial positional relationships of voxels, and the resulting new features containing voxel spatial positional information serve as the input to the next module.
[0054] Constructing the SHC estimation module:
[0055] This module is implemented using adaptive pooling. This operation replaces global average pooling with attention-based pooling, fully considering the importance of different voxels. Different weights are assigned to the pooling operation by calculating the feature similarity between the central voxel and all voxels. The similarity between voxels is measured by calculating the inner product of the feature vectors of the central voxel and the feature vectors of all voxels. To prevent excessively large inner products, all inner products are divided by a scaling factor before being applied using softmax to distribute the pooling weights. The weighted sum of all features using these weights yields the reconstructed central voxel FOD. The calculation formula for attention-based pooling is as follows:
[0056]
[0057] Where, q center Yes, it is the central voxel feature vector, and K is the feature vector of all voxels. T It is the transpose of K. It is the scaling factor, and softmax is the normalization function.
[0058] Model training:
[0059] The FOD reconstruction network consists of a FOD feature extraction module, a 3D coordinate attention module, and an attention-based pooling module connected in series. A single-shell FOD is used as the training set, while a multi-shell FOD serves as the "ground truth" for supervised model training. The model takes only one 3D patch from a single-shell FOD as input each time and outputs the reconstructed version of the central voxel. The reconstructed voxel and its corresponding "ground truth" are compared to calculate the loss function. By minimizing the loss function, the parameter weights of the FOD reconstruction network are adjusted to improve model performance.
[0060] The model settings in this embodiment of the invention are as follows:
[0061] 1. Model initialization: Xavier is used as a method for initializing model weights during training. Initializing weight coefficients can help to properly initialize neural network parameters, thereby promoting network training and performance improvement.
[0062] 2. Loss Function: The model is trained using a mean squared error loss based on spherical harmonics. The formula for calculating this loss function is as follows:
[0063]
[0064] Among them, c i and These are the "ground truth" and the estimated spherical harmonic coefficients, with a total of 45 spherical harmonic coefficients.
[0065] 3. Training details: Adam is used as the optimizer for model training, and a cosine annealing learning rate adjustment strategy is used.
[0066] In another embodiment of the present invention, a system for reconstructing fiber bundle orientation distribution is provided, which can be used to implement the above-described method for reconstructing fiber bundle orientation distribution. Specifically, the system includes:
[0067] The data receiving module is used to acquire diffusion magnetic resonance imaging (dMRI) data and perform preprocessing, calculate the orientation distribution function (FOD) image of the single-shell fiber bundle and divide it into 3D patches;
[0068] The model building module is used to build a FOD image feature extraction model, which takes FOD patch data as input to the model for preliminary feature extraction; build a 3D coordinate attention model, which takes the preliminarily extracted features as input to extract the spatial positional relationship of the input features; and build a spherical harmonic coefficient estimation model, which takes features containing spatial positional relationships as input and outputs the FOD after reconstruction of the 3D patch center voxel.
[0069] The FOD reconstruction module is used to connect the FOD image feature extraction model, the 3D coordinate attention model, and the spherical harmonic coefficient estimation model to form a complete FOD reconstruction network, train the complete model, and use the trained model to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data.
[0070] To verify the effectiveness of this invention, the inventors conducted a qualitative comparative experiment on the HCP public dataset using the method of the embodiments of this invention, along with SS3T-CSD and MSMT-CSD. Figure 6 As shown, the proposed model has demonstrated superior performance compared to SS3T-CSD. In quantitative experiments, the model's performance was evaluated in the white matter region using the angular correlation coefficient (ACC) and mean angle error (MAE) as evaluation metrics. ACC and MAE measure the closeness of the reconstructed FOD to the "ground truth" from the perspectives of spherical harmonic coefficient and fiber bundle, respectively. The quantitative results are shown in Table 1, demonstrating excellent performance on both quantitative metrics in the white matter region.
[0071] Table 1 shows the evaluation results using the angular correlation coefficient (ACC) and mean angular error (MAE).
[0072]
[0073] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a method for reconstructing the directional distribution of white matter fiber tracts in the brain.
[0074] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for reconstructing the directional distribution of nerve fibers in brain white matter in the above embodiments.
[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.
Claims
1. A method for reconstructing the directional distribution of nerve fibers in brain white matter, characterized in that, include: Acquired and preprocessed diffusion magnetic resonance imaging (dMRI) data, calculated the orientation distribution function (FOD) image of the single-shell fiber bundle, and divided it into 3D patches. A FOD image feature extraction model is constructed, and FOD patch data is used as input to the model for preliminary feature extraction. Construct a 3D coordinate attention model, using the initially extracted features as input, to extract the spatial positional relationships of the input features; A spherical harmonic coefficient estimation model is constructed, which takes features containing spatial positional relationships as input and outputs the FOD after the 3D Patch center voxel reconstruction. The FOD image feature extraction model, the 3D coordinate attention model, and the spherical harmonic coefficient estimation model are connected in series to form a complete FOD reconstruction network. The complete model is trained, and the trained model is used to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data. A FOD image feature extraction model is constructed, using FOD patch data as input for preliminary feature extraction, specifically including: The FOD image feature extraction model consists of five multilayer perceptron (MLP)-mixer modules connected in series, with a skip connection between the modules. Each MLP-mixer module consists of two components connected in series: a token-mixing component and a channel-mixing component. Each component is preceded by a LayerNorm normalization layer and a skip connection. The token-mixing component is used to extract features between different voxels, and the channel-mixing component is used to extract features between different spherical harmonic coefficients (SHCs). This is used to initially extract features from the input data, and the output preliminary features are used as input for the next model. Construct a 3D coordinate attention model, taking the initially extracted features as input, and extract the spatial positional relationships of the input features, specifically including: The 3D coordinate attention model consists of a pooling layer, an MLP layer, an activation function layer, and a fusion layer. The pooling layer compresses global information into different dimensions. The MLP encodes the local spatial position information and channel information of voxels. The activation function layer calculates the weight coefficients for different dimensions. The fusion layer performs global information interaction and is used to further extract the spatial positional relationships of voxels. The new features containing voxel spatial position information are used as the input of the next model. A spherical harmonic coefficient estimation model is constructed, taking features containing spatial location relationships as input, and outputting the field of view (FOD) after 3D Patch center voxel reconstruction, specifically including: The spherical harmonic coefficient estimation model is implemented by adaptive pooling operations. It uses attention-based pooling operations instead of global average pooling operations. By calculating the similarity between the features of the central voxel and all features, the softmax function is used to assign different weight coefficients to different voxels. During the pooling operation, the reconstructed FOD is obtained by weighting and summing according to the corresponding weights.
2. The method for reconstructing the directional distribution of nerve fibers in brain white matter according to claim 1, characterized in that, Acquiring and preprocessing diffusion-weighted magnetic resonance imaging (dMRI) data specifically includes: For the required training and testing data, the HCP dataset, publicly available from the Human Connectome Project (HCP), was obtained. A multi-shell FOD was calculated using a multi-shell, multi-organism constrained spherical deconvolution algorithm; this data served as the "ground truth" for supervised model training. The original data was downsampled, and samples were taken from... shell and The shell extracts the first 1 direction and the first 32 directions, and then merges them into a LARDI dataset. Based on the downsampled data, the single-shell FOD is calculated using the single-shell three-organization constrained spherical deconvolution algorithm. This data is used as the training set for the model. When training the model, a 9×9×9 3D patch needs to be divided for each voxel in the training set as the input of the FOD reconstruction network.
3. The method for reconstructing the directional distribution of nerve fibers in brain white matter according to claim 2, characterized in that, For clinical-quality dMRI data, the first step is to perform denoising, artifact removal, motion correction, distortion correction, and eddy current correction on the dMRI data. Then, the single-shell three-tissue constrained spherical deconvolution algorithm is used to calculate the single-shell FOD image from the processed data. When reconstructing the single-shell FOD into a high-quality FOD, a 9×9×9 3D patch needs to be divided for each voxel of the single-shell FOD as the input of the FOD reconstruction network.
4. The method for reconstructing the directional distribution of nerve fibers in brain white matter according to claim 1, characterized in that, A complete FOD reconstruction network is constructed by cascading the data, a complete model is trained, and the trained model is used to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data. Specifically, this includes: Single-shell FOD is used as the training set for the model, while multi-shell FOD is used as the "ground truth" for supervised model training. The model only inputs a 3D patch from a single-shell FOD at a time and outputs the version after central voxel reconstruction. The loss function is calculated between the reconstructed voxel and the corresponding "ground truth". By minimizing the loss function, the parameter weights of the FOD reconstruction network are adjusted to improve model performance.
5. A system for reconstructing the directional distribution of nerve fibers in brain white matter, characterized in that, include: The data receiving module is used to acquire diffusion magnetic resonance imaging (dMRI) data and perform preprocessing, calculate the orientation distribution function (FOD) image of the single-shell fiber bundle and divide it into 3D patches; The model building module is used to build a FOD image feature extraction module, which takes FOD patch data as input to perform preliminary feature extraction; a 3D coordinate attention module is built, which takes the preliminarily extracted features as input to extract the spatial position relationship of the input features; and a spherical harmonic coefficient estimation module is built, which takes features containing spatial position relationships as input and outputs the FOD after reconstruction of the 3DPatch center voxel. The FOD reconstruction module is used to connect the FOD image feature extraction module, the 3D coordinate attention module, and the spherical harmonic coefficient estimation module to form a complete FOD reconstruction network, train a complete model, and use the trained model to estimate the high-quality fiber bundle orientation distribution function (FOD) from LADRI data. A FOD image feature extraction module is constructed, which uses FOD patch data as input for preliminary feature extraction, specifically including: The FOD image feature extraction module consists of five multilayer perceptron (MLP)-mixer modules connected in series, with a skip connection between the modules. Each MLP-mixer module consists of two components connected in series: a token-mixing component and a channel-mixing component. Each component is preceded by a LayerNorm normalization layer and a skip connection. The token-mixing component is used to extract features between different voxels, and the channel-mixing component is used to extract features between different spherical harmonic coefficients (SHCs). This module is used to initially extract features from the input data, and the output preliminary features are used as input to the next module. Construct a 3D coordinate attention module, taking the initially extracted features as input, to extract the spatial relationship of the input features, specifically including: The 3D coordinate attention module consists of a pooling layer, an MLP layer, an activation function layer, and a fusion layer. The pooling layer compresses global information into different dimensional directions. The MLP encodes the local spatial position information and channel information of voxels. The activation function layer calculates the weight coefficients for different dimensional directions. The fusion layer performs global information interaction and is used to further extract the spatial positional relationships of voxels. The new features containing voxel spatial position information are used as the input of the next module. A spherical harmonic coefficient estimation module is constructed, which takes features containing spatial positional relationships as input and outputs the Foreground Object Dimension (FOD) after 3D Patch center voxel reconstruction, specifically including: The spherical harmonic coefficient estimation module is implemented by adaptive pooling operations. It uses attention-based pooling operations instead of global average pooling operations. By calculating the similarity between the features of the central voxel and all features, the softmax function is used to assign different weight coefficients to different voxels. During the pooling operation, the reconstructed FOD is obtained by weighting and summing according to the corresponding weights.
6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for reconstructing the directional distribution of nerve fibers in the white matter of the brain as described in any one of claims 1 to 4.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for reconstructing the directional distribution of nerve fibers in the white matter of the brain as described in any one of claims 1 to 4.