A Deep Learning-Based Method and Apparatus for Mapping Microstructure of Diffusion Tensor Distribution

Abstract

This invention provides a method and apparatus for mapping tissue microstructures based on diffusion tensor distribution using deep learning. The method includes the following steps: preprocessing diffusion magnetic resonance imaging (DMRI) data; determining target voxels in the DMRI data using an FSL tool; dividing the DMRI data into target data blocks based on the target voxels; unfolding the three-dimensional target data blocks into a two-dimensional matrix; processing the two-dimensional matrix into embedded feature vectors; inputting the embedded feature vectors into a Transformer encoder to obtain encoded feature vectors; inputting the encoded feature vectors into a decoder, the decoder including a sparse dictionary mapping branch and a prediction branch, the sparse dictionary mapping branch outputting brain component distribution vectors, and the prediction branch outputting regression calculation parameters; calculating multiple brain component mapping pixels at the target voxel locations based on the brain component distribution vectors and regression calculation parameters, and constructing microstructure parameter maps corresponding to various brain components.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and apparatus for mapping tissue microstructure based on diffusion tensor distribution using deep learning. Background Technology

[0002] Diffusion magnetic resonance imaging (dMRI) is a non-invasive imaging technique that characterizes the microstructure of tissues based on the random diffusion behavior of water molecules. It has been widely used in the clinical diagnosis and mechanistic research of brain diseases. Traditional diffusion tensor imaging (DTI) assumes that the diffusion of water molecules within each voxel can be described by a single diffusion tensor. Through tensor fitting, parameters such as mean diffusivity (MD) and fractional anisotropy (FA) can be obtained, and a relatively mature clinical application procedure has been established.

[0003] However, in actual brain tissue, a single voxel often contains multiple microstructural components simultaneously, such as the intersection of white matter fiber bundles with different orientations, the mixing of gray matter and white matter or cerebrospinal fluid, and complex changes in cell density and morphology within the lesion area. In these cases, a single diffusion tensor model cannot simultaneously characterize the diffusion characteristics of multiple components; its fitting results only reflect a weighted average of multiple diffusion behaviors, making it difficult to accurately depict the microstructural heterogeneity within a voxel, leading to the loss of crucial information. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for mapping the microstructure of an organization based on the diffusion tensor distribution using deep learning, in order to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a method for mapping tissue microstructure based on diffusion tensor distribution using deep learning, the method comprising the following steps: Acquire diffusion magnetic resonance imaging data, which includes diffusion-weighted signal data under linear diffusion coding and spherical diffusion coding; The diffusion magnetic resonance imaging data is preprocessed, and the target voxels in the diffusion magnetic resonance imaging data are determined based on the FSL tool. Target data blocks are then divided in the diffusion magnetic resonance imaging data based on the target voxels. The three-dimensional target data block is unfolded into a two-dimensional matrix, and the two-dimensional matrix is ​​processed sequentially through a feature embedding layer and a position encoding layer to obtain an embedded feature vector. The embedded feature vector is input into the Transformer encoder, which uses an attention mechanism to encode the features and obtain the encoded feature vector. The encoded feature vector is input into the decoder, which includes a sparse dictionary mapping branch and a prediction branch. The sparse dictionary mapping branch outputs a brain component distribution vector, and the prediction branch outputs regression calculation parameters. Based on the brain component distribution vector and regression calculation parameters, the mapping pixels of various brain components at the target voxel locations are calculated, and microstructure parameter maps corresponding to various brain components are constructed based on the brain component mapping pixels at all target voxel locations.

[0006] Using the above scheme, this scheme directly learns the nonlinear mapping relationship between diffusion magnetic resonance signals and diffusion tensor distribution tissue microstructure parameter maps by constructing an end-to-end deep learning model, replacing the traditional random inversion process based on Monte Carlo sampling. The parameters of diffusion tensor distribution imaging are essentially obtained by statistical analysis of the multi-component diffusion behavior within voxels, and their values ​​have a stable but highly nonlinear correspondence with the diffusion magnetic resonance signals. Furthermore, this scheme employs a dual-channel decoder, outputting brain component distribution vectors through the sparse dictionary mapping branch and regression calculation parameters through the prediction branch. This allows for the simultaneous characterization of multi-component diffusion characteristics through the outputs of both channels, accurately depicting the microstructural heterogeneity within voxels, and finally combining the outputs of both channels to calculate the target image.

[0007] In some embodiments of the present invention, in the step of determining the target voxel in the diffusion magnetic resonance imaging data based on the FSL tool, the FSL tool is used to determine the voxel of the brain tissue portion in the diffusion magnetic resonance imaging data as the target voxel.

[0008] In some embodiments of the present invention, in the step of dividing the target data block in the diffusion magnetic resonance imaging data based on the target voxel, a three-dimensional data region of a preset size is divided as the target data block with the position of the target voxel in the diffusion magnetic resonance imaging data as the center.

[0009] In some embodiments of the present invention, in the step of processing the two-dimensional matrix sequentially through a feature embedding layer and a position encoding layer to obtain an embedded feature vector, the feature embedding layer is provided with at least one fully connected layer connected in sequence, and the two-dimensional matrix outputs a first feature sequence through the last fully connected layer of the feature embedding layer; the position encoding layer adds a preset classification label vector to the beginning of the first feature sequence to obtain an initial position vector, processes the initial position vector using three-dimensional sine and cosine encoding to obtain a position vector, and concatenates the position vector and the first feature sequence to obtain the embedded feature vector.

[0010] In some embodiments of the present invention, in the step of inputting the embedded feature vector into a Transformer encoder, wherein the Transformer encoder uses an attention mechanism to encode features and obtain an encoded feature vector, the Transformer encoder includes multiple Transformer blocks arranged in sequence, and the encoded feature vector is output through the last Transformer block.

[0011] In some embodiments of the present invention, in the step of inputting the embedded feature vector into a Transformer encoder, wherein the Transformer encoder uses an attention mechanism to encode features and obtain an encoded feature vector, the Transformer encoder includes multiple Transformer blocks arranged in sequence. In the processing of each Transformer block, the input data is mapped using three independent linear layers to obtain a query vector, a key vector, and a value vector. In some embodiments of the present invention, in the step of inputting the encoded feature vector into a decoder, the decoder including a sparse dictionary mapping branch and a prediction branch, the sparse dictionary mapping branch outputting a brain component distribution vector, and the prediction branch outputting regression calculation parameters, the prediction branch outputs regression calculation parameters through a fully connected layer.

[0012] In some embodiments of the present invention, in the step of inputting the encoded feature vector into a decoder, the decoder including a sparse dictionary mapping branch and a prediction branch, the sparse dictionary mapping branch outputting a brain component distribution vector, and the prediction branch outputting regression calculation parameters, the sparse dictionary mapping branch is sequentially configured with multiple sub-processing units and a layer normalization layer. The input segment of the decoder is residually connected to the output of each sub-processing unit. The encoded feature vector is concatenated with the output of each sub-processing unit and output to the next sub-processing unit. Each sub-processing unit includes a threshold layer and a convolutional layer connected in sequence. The threshold layer is used to set values ​​whose absolute value is less than a threshold to 0. The convolutional layer is used to perform convolution processing. The layer normalization layer is connected to the last sub-processing unit and outputs a brain component distribution vector.

[0013] In some embodiments of the present invention, the brain component distribution vector is provided with three feature dimensions, the values ​​of which are the white matter ratio parameter, the gray matter ratio parameter, and the cerebrospinal fluid ratio parameter, respectively; the regression calculation parameters include the isotropic diffusion expectation, the anisotropic diffusion expectation, the variance of the anisotropic diffusion expectation, and the variance of the isotropic diffusion expectation. In the step of constructing a microstructure parameter map corresponding to various brain components based on the brain component mapping pixels of all target voxel locations, the parameter of each dimension in the brain component distribution vector is multiplied by each regression calculation parameter to obtain multiple brain component mapping pixels, and a microstructure parameter map is constructed based on each type of brain component mapping pixel of all target voxels.

[0014] In some embodiments of the present invention, in the step of constructing a microstructure parameter map based on each brain component mapping pixel of all the target voxels, the position of the corresponding brain component mapping pixel is determined based on the position of the target voxel in all voxels, and the corresponding position is rendered using the brain component mapping pixel to construct the corresponding type of microstructure parameter map.

[0015] A second aspect of the present invention also provides a deep learning-based diffusion tensor distribution organization microstructure mapping device, the device including a computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the device implementing the steps of the method described above.

[0016] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned deep learning-based diffusion tensor distribution organization microstructure mapping method.

[0017] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0018] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0020] Figure 1This is a schematic diagram of one embodiment of the diffusion tensor distribution organization microstructure mapping method based on deep learning of the present invention; Figure 2 This is a comparative diagram of the proposed solution and existing technical solutions; Figure 3 This is a schematic diagram of the architecture of this solution. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0022] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0023] To overcome the limitations of single-tensor models, diffusion tensor distribution (DTD) imaging was proposed during implementation. This technique no longer assumes only one diffusion tensor within a voxel, but instead models diffusion behavior within a voxel as a probability distribution of multiple diffusion tensors. Specifically, the DTD method inverts diffusion signals with multiple b-values ​​and multiple diffusion encoding methods to estimate a set of diffusion tensors and their corresponding probability weights, thereby obtaining the diffusion tensor distribution function within the voxel. After obtaining the diffusion tensor distribution, a series of microstructural parameters can be further calculated, such as the average diffusion rate distribution, anisotropy distribution, and their statistics, to quantitatively describe the proportions and heterogeneity characteristics of different diffusion components within the voxel. Compared to traditional DTI, DTD can distinguish different diffusion components at the voxel scale, demonstrating stronger characterization capabilities for complex brain tissue structures and lesion microenvironments.

[0024] In existing technologies, the mainstream approach to solving diffusion tensor distribution (DTD) imaging parameters is the numerical inversion method based on Monte Carlo sampling. This method combines random sampling with statistical estimation to model the multi-component diffusion tensor that may exist within a voxel. Its typical implementation process is as follows: First, based on the diffuse magnetic resonance signals acquired using multiple b-values ​​and multiple diffusion encoding methods, random sampling is performed within a predefined diffusion tensor parameter space. Each sampling corresponds to a set of candidate diffusion tensor parameters, used to simulate the diffusion behavior of a potential diffusion component within a voxel.

[0025] Secondly, by substituting the aforementioned candidate diffusion tensor parameters into the diffusion signal model, the corresponding theoretical diffusion signal is calculated and compared with the actual measured signal. Based on this, an iterative optimization strategy is employed to filter and update the sampling results, gradually reducing the error between the generated signal and the measured signal, thereby achieving the numerical inversion of the diffusion tensor distribution. This process typically requires numerous repeated iterations in the parameter space to ensure the stability of the inversion results.

[0026] Subsequently, to reduce the uncertainty introduced by random sampling, multiple independent Monte Carlo inversions need to be performed on the same voxel. For the diffusion tensor distribution result obtained from each inversion, its statistic (e.g., median or mean) is calculated, and this statistic is used as the final diffusion tensor distribution estimate for that voxel.

[0027] Finally, after obtaining the diffusion tensor distribution within the voxel, a series of microstructure parameter maps are calculated based on each diffusion tensor and its corresponding weight in the distribution, which are used to quantitatively characterize the overall level and heterogeneity of diffusion properties within the voxel.

[0028] The aforementioned DTD parameter inversion method based on Monte Carlo sampling has a clear physical modeling foundation and a relatively complete theoretical framework. It can recover the diffusion information of multiple components within a voxel to a certain extent. Therefore, it is widely regarded as a representative scheme for solving DTD parameters in the prior art, and it is also the prior art that is closest to the technical solution of this invention.

[0029] Currently, diffusion tensor distribution imaging methods based on Monte Carlo sampling inversion have at least the following technical drawbacks in practical applications.

[0030] (1) High computational complexity and long processing time.

[0031] Existing methods require extensive random sampling within a high-dimensional diffusion tensor parameter space for each voxel, followed by multiple iterative optimizations to obtain stable inversion results. To reduce errors caused by randomness, it is often necessary to perform multiple independent inversions on the same voxel and then statistically summarize the results. This process significantly increases the computational load per voxel; at the whole-brain scale, data processing time for a single subject can typically reach several hours or even longer, making it difficult to meet the application requirements for rapid processing or near real-time feedback in clinical examinations.

[0032] (2) It consumes a lot of memory and storage resources, which is not conducive to large-scale application. During the inversion process, to obtain reliable statistical results, it is necessary to save intermediate results and final distribution parameters obtained from multiple Monte Carlo sampling and iterations. This not only increases runtime memory usage but also significantly increases data storage requirements. When the sample size increases or multi-center studies are conducted, the aforementioned resource consumption will be further amplified, limiting the widespread application of this method in large-sample or routine clinical procedures.

[0033] (3) Independent inversion on a voxel-by-voxel basis, ignoring spatial continuity constraints. Existing Monte Carlo inversion methods typically use voxels as the smallest processing unit, with parameter estimation performed independently for each voxel, without explicitly introducing constraints on spatial continuity or structural correlation between adjacent voxels. In regions with low signal-to-noise ratios or complex structures, this can easily lead to problems such as amplified spatial noise, discontinuous parameter maps, and unclear microstructure boundaries, thus affecting the stability and interpretability of the results.

[0034] In summary, existing diffusion tensor distribution imaging methods based on Monte Carlo inversion still have significant shortcomings in terms of computational efficiency, resource consumption, and spatial consistency, making it difficult to balance accuracy with practical application requirements.

[0035] like Figure 1 and Figure 3 As shown in (A), Figure 3 (A) is a schematic diagram of the overall processing architecture of this solution. This invention proposes a method for mapping microstructures based on diffusion tensor distribution using deep learning. The steps of the method include: Step S100: Acquire diffusion magnetic resonance imaging data, which includes diffusion-weighted signal data under linear diffusion coding and spherical diffusion coding. In some embodiments of this invention, diffusion magnetic resonance imaging (dMRI) indirectly reflects the microstructural information of tissues by measuring the Brownian motion of water molecules in biological tissues. Its core principle is that water molecules exhibit different diffusion behaviors in different tissues; by applying a diffusion-sensitive gradient field, the diffusion characteristics of water molecules can be detected.

[0036] Linear diffusion coding and spherical diffusion coding are two different ways of applying diffusion-sensitive gradient fields to detect the diffusion characteristics of water molecules in different directions. Linear diffusion encoding refers to applying a diffusion-sensitive gradient field in a single direction to detect the diffusion characteristics of water molecules in that direction. This encoding method is the basis of diffusion-weighted imaging (DWI), but it can only detect the diffusion ability of water molecules in a specific direction and cannot comprehensively describe the diffusion behavior of water molecules in three-dimensional space.

[0037] Spherical diffusion coding refers to applying diffusion-sensitive gradient fields in multiple directions in three-dimensional space to comprehensively detect the diffusion characteristics of water molecules in three-dimensional space. This coding method is the basis of diffusion tensor imaging (DTI) and diffusion kurtosis imaging (DKI) and can describe the anisotropic diffusion behavior of water molecules.

[0038] Specifically, the acquisition of diffusion magnetic resonance imaging (DMRI) data includes at least diffusion-weighted signals under linear diffusion encoding and spherical diffusion encoding, covering multiple b-value settings. The DMRI data constitutes the raw signal input for diffusion tensor distribution imaging, used to characterize the multi-component diffusion behavior within voxels.

[0039] Step S200: Preprocess the diffusion magnetic resonance imaging data, determine the target voxels in the diffusion magnetic resonance imaging data based on the FSL tool, and divide the diffusion magnetic resonance imaging data into target data blocks based on the target voxels. FSL (FMRIB Software Library) is an open-source neuroimaging analysis toolkit developed by Oxford University in the UK. Voxel processing is one of its core functions, which refers to the quantitative analysis of the smallest three-dimensional unit (voxel) in brain images to uncover abnormal features of brain structure and function. It is a key technology for neuroscience research and clinical diagnosis.

[0040] In the specific implementation process, the target voxel is a voxel of brain tissue from which non-brain tissues such as the skull and scalp have been removed.

[0041] In the specific implementation process, in the step of preprocessing the diffuse magnetic resonance imaging data, the acquired diffuse magnetic resonance data is preprocessed, including image registration, denoising, and intensity normalization.

[0042] In the step of dividing the target data block in the diffusion magnetic resonance imaging data based on the target voxel, the whole brain diffusion data is divided into multiple three-dimensional local data blocks according to a preset size. Each data block is centered on the target voxel and contains multiple neighboring voxels in its spatial neighborhood, thereby forming an input sample containing spatial context information.

[0043] Step S300: Expand the three-dimensional target data block into a two-dimensional matrix, and process the two-dimensional matrix sequentially through a feature embedding layer and a position encoding layer to obtain an embedded feature vector; In the specific implementation process, in the step of unfolding the three-dimensional target data block into a two-dimensional matrix, each three-dimensional target data block is unfolded according to the voxel dimension and signal channel dimension to form a two-dimensional matrix representation. The original two-dimensional matrix is ​​mapped to a high-dimensional feature space through at least one fully connected mapping or equivalent linear mapping structure as a feature embedding layer. Simultaneously, to avoid the loss of spatial information during the unfolding process, positional encoding information is introduced to explicitly represent the relative spatial positions of different voxels within the data block.

[0044] Step S400: The embedded feature vector is input into the Transformer encoder, which uses an attention mechanism to encode features and obtain an encoded feature vector. Using the above approach, the embedded feature vectors are input into an encoding network containing a self-attention structure. By modeling the correlations between local voxels and between different signal features, high-level feature representations incorporating spatial context information are extracted. This step enables the model to adaptively focus on voxels and signal features that contribute significantly to the discrimination of tissue microstructures.

[0045] Step S500: Input the encoded feature vector into the decoder. The decoder includes a sparse dictionary mapping branch and a prediction branch. The sparse dictionary mapping branch outputs a brain component distribution vector, and the prediction branch outputs regression calculation parameters. Specifically, such as Figure 2 As shown, this scheme constructs an end-to-end deep learning model to directly learn the nonlinear mapping relationship between the diffuse magnetic resonance signal and the diffusion tensor distribution microstructure parameter map, in order to replace the traditional random inversion process based on Monte Carlo sampling. Specifically, the parameters of diffusion tensor distribution imaging are essentially obtained statistically from the diffusion behavior of multiple components within voxels, and their values ​​have a stable but highly nonlinear correspondence with the diffusion magnetic resonance signal. Traditional methods explicitly solve this relationship through random sampling and iterative inversion, while this approach uses a deep neural network to model this mapping in a data-driven manner. During the training phase, Monte Carlo inversion results are used as a supervision signal, enabling the model to approximate this complex mapping function.

[0046] This scheme, through the aforementioned dual-branch structure, can simultaneously balance the overall parameter prediction accuracy and the ability to express the diffusion characteristics of multiple components.

[0047] Step S600: Calculate the mapping pixels of various brain components at the target voxel locations based on the brain component distribution vector and regression calculation parameters, and construct microstructure parameter maps corresponding to various brain components based on the brain component mapping pixels at all target voxel locations.

[0048] In the specific implementation process, in the step of calculating the target voxel location of multiple brain component mapping pixels based on the brain component distribution vector and regression calculation parameters, the total expected parameters obtained by the direct prediction branch are fused with the weights of each distribution component output by the sparse dictionary mapping branch, and the tissue microstructure parameters corresponding to different diffusion intervals are obtained by weighted calculation.

[0049] Using the above scheme, this scheme directly learns the nonlinear mapping relationship between diffusion magnetic resonance signals and diffusion tensor distribution tissue microstructure parameter maps by constructing an end-to-end deep learning model, replacing the traditional random inversion process based on Monte Carlo sampling. The parameters of diffusion tensor distribution imaging are essentially obtained by statistical analysis of multi-component diffusion behavior within voxels, and their values ​​have a stable but highly nonlinear correspondence with diffusion magnetic resonance signals. Furthermore, this scheme employs a dual-channel decoder, outputting brain component distribution vectors through the sparse dictionary mapping branch and regression calculation parameters through the prediction branch. This allows for the simultaneous characterization of multi-component diffusion characteristics through the outputs of both channels, and the calculation of the target image by combining the outputs of both channels.

[0050] In some embodiments of the present invention, in the step of determining the target voxel in the diffusion magnetic resonance imaging data based on the FSL tool, the FSL tool is used to determine the voxel of the brain tissue portion in the diffusion magnetic resonance imaging data as the target voxel.

[0051] In some embodiments of the present invention, in the step of dividing the target data block in the diffusion magnetic resonance imaging data based on the target voxel, a three-dimensional data region of a preset size is divided as the target data block with the position of the target voxel in the diffusion magnetic resonance imaging data as the center.

[0052] In the specific implementation process, in the step of dividing a three-dimensional data region of a preset size as the target data block with the position of the target voxel in the diffusion magnetic resonance imaging data as the center, the target data block is constructed based on voxels within a one-hop range with the position of the target voxel in the diffusion magnetic resonance imaging data as the center.

[0053] Using the above scheme, this scheme models local spatial neighborhoods: by jointly inputting the signals of the voxel spatial neighborhood, it explicitly utilizes the structural correlation between adjacent voxels.

[0054] In some embodiments of the present invention, in the step of processing the two-dimensional matrix sequentially through a feature embedding layer and a position encoding layer to obtain an embedded feature vector, the feature embedding layer is provided with at least one fully connected layer connected in sequence, and the two-dimensional matrix outputs a first feature sequence through the last fully connected layer of the feature embedding layer; the position encoding layer adds a preset classification label vector to the beginning of the first feature sequence to obtain an initial position vector, processes the initial position vector using three-dimensional sine and cosine encoding to obtain a position vector, and concatenates the position vector and the first feature sequence to obtain the embedded feature vector.

[0055] In some embodiments of the present invention, in the step of inputting the embedded feature vector into a Transformer encoder, wherein the Transformer encoder uses an attention mechanism to encode features and obtain an encoded feature vector, the Transformer encoder includes multiple Transformer blocks arranged in sequence, and the encoded feature vector is output through the last Transformer block.

[0056] like Figure 3 As shown in (B), Figure 3 (B) is a schematic diagram of the Transformer block in this scheme. In some embodiments of the present invention, in the step of inputting the embedded feature vector into the Transformer encoder, the Transformer encoder uses an attention mechanism to encode the features and obtain the encoded feature vector, the Transformer encoder includes multiple Transformer blocks arranged in sequence. In the processing of each Transformer block, the input data is mapped using three independent linear layers to obtain the query vector, key vector and value vector.

[0057] In practical implementation, this solution uses the self-attention structure of the Transformer encoder to adaptively weight the importance of different voxels and different signal channels, thereby enhancing the ability to model key information.

[0058] like Figure 3 As shown in (C), Figure 3 (C) is a schematic diagram of the decoder structure of this scheme. In some embodiments of the present invention, in the step of inputting the encoded feature vector into the decoder, the decoder includes a sparse dictionary mapping branch and a prediction branch, the sparse dictionary mapping branch outputs a brain component distribution vector, and the prediction branch outputs regression calculation parameters, the prediction branch outputs regression calculation parameters through a fully connected layer.

[0059] By adopting the above scheme, the model's ability to express the diffusion characteristics of multiple components is improved by constraining the combination relationships of different tissue components through a sparse dictionary.

[0060] In practical implementation, this solution generates the target parameter map with only one forward propagation during the model inference stage, without any random sampling, iterative optimization or statistical summarization operations, thereby significantly reducing computational complexity and improving processing efficiency.

[0061] In some embodiments of the present invention, in the step of inputting the encoded feature vector into a decoder, the decoder including a sparse dictionary mapping branch and a prediction branch, the sparse dictionary mapping branch outputting a brain component distribution vector, and the prediction branch outputting regression calculation parameters, the sparse dictionary mapping branch is sequentially configured with multiple sub-processing units and a layer normalization layer. The input segment of the decoder is residually connected to the output of each sub-processing unit. The encoded feature vector is concatenated with the output of each sub-processing unit and output to the next sub-processing unit. Each sub-processing unit includes a threshold layer and a convolutional layer connected in sequence. The threshold layer is used to set values ​​whose absolute value is less than a threshold to 0. The convolutional layer is used to perform convolution processing. The layer normalization layer is connected to the last sub-processing unit and outputs a brain component distribution vector.

[0062] In some embodiments of the present invention, the brain component distribution vector is provided with three feature dimensions, the values ​​of which are the white matter ratio parameter, the gray matter ratio parameter, and the cerebrospinal fluid ratio parameter, respectively; the regression calculation parameters include the isotropic diffusion expectation, the anisotropic diffusion expectation, the variance of the anisotropic diffusion expectation, and the variance of the isotropic diffusion expectation. In the step of constructing a microstructure parameter map corresponding to various brain components based on the brain component mapping pixels of all target voxel locations, the parameter of each dimension in the brain component distribution vector is multiplied by each regression calculation parameter to obtain multiple brain component mapping pixels, and a microstructure parameter map is constructed based on each type of brain component mapping pixel of all target voxels.

[0063] In some embodiments of the present invention, in the step of constructing a microstructure parameter map based on each brain component mapping pixel of all the target voxels, the position of the corresponding brain component mapping pixel is determined based on the position of the target voxel in all voxels, and the corresponding position is rendered using the brain component mapping pixel to construct the corresponding type of microstructure parameter map.

[0064] In the specific implementation process, in the step of multiplying the parameters of each dimension of the brain component distribution vector with each regression calculation parameter to obtain multiple brain component mapping pixels, the multiplied values ​​are normalized to between 0 and 1. Then, the normalized values ​​are mapped to pixel values ​​through value mapping. The value range corresponding to the normalized values ​​can be determined by using a preset mapping table, and the pixel values ​​corresponding to the value range are used as brain component mapping pixels.

[0065] like Figure 3 As shown in (D), Figure 3 (D) is a schematic diagram of the input and output data of this scheme. In the specific implementation process, in the step of multiplying the parameter of each dimension in the brain component distribution vector with each regression calculation parameter to obtain the mapping pixels of multiple brain components; , This represents the expected value of white matter-related isotropic diffusion rate; Specifically, due to The proportion parameters of brain white matter mainly correspond to the highly anisotropic white matter fiber structure; by The constructed microstructure parameter map reflects the average diffusion rate of white matter-related diffusion components within the voxel; it is closely related to axon density, myelin sheath integrity, and the degree of extracellular space restriction; under pathological conditions such as tumor infiltration, demyelination, or edema, E[D_iso]_bin1 often increases, indicating weakened white matter microstructure constraint.

[0066] It provides equivalent MD information at the level of white matter diffusion components, but its statistical basis comes from DTD, rather than a single tensor assumption.

[0067] , This represents the expected value of the gray matter-related isotropic diffusion rate; by The microstructure parameter map is constructed to correspond to the expected isotropic diffusion rate of gray matter. It mainly reflects the characteristics of gray matter tissue and can describe the average diffusion level of water molecules in neuronal cell bodies, dendrites and local microvascular environments. Compared with white matter, its value is usually slightly higher, but its anisotropy is significantly reduced. It is also more sensitive to changes in cortical structure, changes in cell density or local inflammation.

[0068] , This represents the expected rate of isotropic diffusion related to cerebrospinal fluid. by Constructing microstructural parameter maps corresponding to the expected isotropic diffusivity of cerebrospinal fluid (CSF) mainly corresponds to CSF ​​or free water-like components: the values ​​are close to the free water diffusivity; they can be used to quantify the proportion of CSF contamination or edema components within voxels; they are usually significantly elevated in tumor necrosis areas or periventricular regions.

[0069] , This represents the expectation of white matter-related anisotropy. by Microstructural parameter maps are constructed to characterize the overall orderliness of white matter fiber structures: higher values ​​reflect consistent axonal arrangement and obvious directional diffusion characteristics; they are closely related to fiber bundle integrity and orientation consistency; values ​​usually decrease in cases of fiber destruction, increased crossings, or tumor infiltration.

[0070] , This represents the expected value of gray matter-related anisotropy. by Construct a microstructure parameter map to correspond to the anisotropy expectation of the diffusion components in gray matter: the values ​​are usually low, reflecting the approximately isotropic diffusion environment in gray matter; they are affected by factors such as the arrangement of neuronal cell bodies and the layered structure of the cortex; slight changes may occur when gray matter is invaded or its structure is disordered.

[0071] , Indicates the expectation of cerebrospinal fluid-related anisotropy; by Constructing microstructure parameter maps corresponds to the expected anisotropy of cerebrospinal fluid: the values ​​are theoretically close to zero, which is consistent with the approximately isotropic diffusion characteristics of free water; a significant increase in values ​​usually indicates partial volume effects or interbinary aliasing. , This represents the variance of white matter-related isotropic diffusivity; by Microstructure parameter maps were constructed to characterize the dispersion of diffusion components in white matter samples in terms of diffusion rate: the variance of anisotropic diffusion expectation; reflecting the difference in diffusion constraints between different white matter components within a voxel; originating from differences in fiber diameter, crossed fibers, or local pathological infiltration, i.e., an increase in the value indicates enhanced heterogeneity of white matter microstructure.

[0072] , This represents the variance of the isotropic diffusivity related to gray matter; by Microstructure parameter maps were constructed to describe the discreteness of the internal diffusion rate distribution of diffusion components in gray matter samples, reflecting the complexity of gray matter tissue and the degree of mixing of different cellular components; it also showed a certain sensitivity to abnormal cortical structure and uneven cell density.

[0073] , This represents the variance of the isotropic diffusivity associated with cerebrospinal fluid; by Construct microstructure parameter maps to characterize the consistency of diffusion rate of CSF or free water sample diffusion components: the values ​​are usually low, which is consistent with the assumption of uniformity of free diffusion environment; if they are significantly increased, it may indicate that CSF is mixed with tissue components or that there is uneven local edema.

[0074] , This represents the variance of white matter-related anisotropy. by Microstructure parameter maps were constructed to assess the heterogeneity of white matter-like diffuse components at the anisotropic level: the values ​​were highly sensitive to fiber orientation dispersion and intra-bundle structural inconsistencies; they were typically elevated in cases of white matter damage, fiber breakage, or tumor invasion.

[0075] , This represents the variance of gray matter correlation anisotropy. by Construct microstructural parameter maps to describe the fluctuation range of anisotropy levels in gray matter: the values ​​are generally small, reflecting the relatively stable diffusion morphology of gray matter; they may increase slightly in pathological conditions due to structural disorder.

[0076] , Indicates the variance of anisotropy related to cerebrospinal fluid; by Microstructure parameter maps were constructed to characterize the anisotropic dispersion of the diffusion components in free water samples: numerically, the dispersion is close to zero; abnormally high values ​​may indicate the existence of a non-ideal free diffusion environment.

[0077] 1. This scheme provides a method for solving the diffusion tensor distribution parameters without the need for Monte Carlo random sampling inversion, and can directly and quickly obtain the diffusion tensor distribution parameter map from the diffusion magnetic resonance signal or its derived features; 2. This approach significantly reduces computation time and computational resource consumption. While ensuring the reliability of the results, it shortens the computation time of the whole-brain diffusion tensor distribution parameters, thereby improving the applicability of the method in clinical and scientific research. 3. This approach ensures the quantitative accuracy of parametric maps while introducing spatial information constraints to improve the spatial continuity of the results, reduce noise interference, and enhance the structural clarity of tissue microstructure boundaries and lesion areas; 4. This solution provides a tissue microstructure mapping technique that enables the obtained diffusion tensor distribution parameters to exhibit good stability and generalization ability in different individuals and under different disease states.

[0078] The beneficial effects of this plan include: 1. Computational efficiency is significantly improved. This approach constructs an end-to-end deep learning model to directly predict the diffusion tensor distribution parameter map from the diffusion magnetic resonance signal, completely avoiding the extensive random sampling, parameter space search, and repetitive iterative optimization processes required in traditional methods. Compared to Monte Carlo inversion-based implementations, it significantly reduces the processing time for whole-brain data under the same hardware conditions.

[0079] 2. This solution significantly reduces computational resource consumption. Since the inference phase does not require saving and statistically analyzing multiple inversion results, this solution only needs to retain the model weights and intermediate features required for one forward computation, which significantly reduces runtime memory usage and data storage requirements. It is especially suitable for application scenarios involving multiple subjects, large-scale data, or routine clinical procedures.

[0080] 3. It exhibits superior continuity and stability in the parametric graph space. This scheme introduces three-dimensional local data blocks as basic processing units and combines them with attention mechanisms to model spatial context information. It effectively utilizes the structural correlation between adjacent voxels, thereby suppressing the noise amplification problem caused by voxel-by-voxel independent inversion, making the generated tissue microstructure parameter map more continuous, smooth and with clear structural boundaries in space.

[0081] 4. Highly applicable in clinical practice and easy to integrate into systems. The reasoning process of this solution is deterministic, with stable and repeatable output. The overall processing flow is simple and time-saving, making it suitable for integration into existing clinical image processing workstations or automated analysis pipelines, providing fast and reliable parameter support for clinical auxiliary diagnosis and preoperative assessment.

[0082] 5. This solution has good generalization ability and wide applicability. Even when trained using only data from healthy subjects, the method described in this protocol can still stably generate reasonable diffusion tensor distribution parameter maps on diffusion magnetic resonance data containing lesions, indicating that the method has good generalization ability to different tissue states and is applicable to healthy people and people with various brain diseases.

[0083] 6. The usage process is highly simplified, and the deployment and application threshold is low. The diffusion tensor distribution parameter calculation method described in this solution uses a trained deep learning model as the core calculation unit. In actual use, it is only necessary to download or load the corresponding model code and model parameter file, and input the diffusion magnetic resonance imaging data to be processed into the model to automatically complete the calculation and output of the diffusion tensor distribution parameters.

[0084] Compared to existing Monte Carlo inversion-based techniques, this solution eliminates the need for complex algorithm configuration, parameter initialization, or multi-round calculations, and also eliminates the need for manual intervention in the inversion process, significantly simplifying the overall usage steps. This feature makes the invention easier to integrate into existing image processing platforms, research analysis workflows, or clinical workstations, facilitating wider application and reducing actual usage costs.

[0085] This invention also provides a deep learning-based diffusion tensor distribution organization microstructure mapping device. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0086] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned deep learning-based diffusion tensor distribution organization microstructure mapping method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0087] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0088] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0089] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for mapping microstructures of an organization based on diffusion tensor distribution using deep learning, characterized in that, The steps of the method include: Acquire diffusion magnetic resonance imaging data, which includes diffusion-weighted signal data under linear diffusion coding and spherical diffusion coding; The diffusion magnetic resonance imaging data is preprocessed, and the target voxels in the diffusion magnetic resonance imaging data are determined based on the FSL tool. Target data blocks are then divided in the diffusion magnetic resonance imaging data based on the target voxels. The three-dimensional target data block is unfolded into a two-dimensional matrix, and the two-dimensional matrix is ​​processed sequentially through a feature embedding layer and a position encoding layer to obtain an embedded feature vector. The embedded feature vector is input into the Transformer encoder, which uses an attention mechanism to encode the features and obtain the encoded feature vector. The encoded feature vector is input into the decoder, which includes a sparse dictionary mapping branch and a prediction branch. The sparse dictionary mapping branch outputs a brain component distribution vector, and the prediction branch outputs regression calculation parameters. Based on the brain component distribution vector and regression calculation parameters, the mapping pixels of various brain components at the target voxel locations are calculated, and microstructure parameter maps corresponding to various brain components are constructed based on the brain component mapping pixels at all target voxel locations.

2. The method for mapping microstructures based on diffusion tensor distribution according to claim 1, characterized in that, In the step of determining the target voxel in the diffusion magnetic resonance imaging data based on the FSL tool, the FSL tool is used to determine the voxel of the brain tissue portion in the diffusion magnetic resonance imaging data as the target voxel.

3. The method for mapping microstructures based on diffusion tensor distribution according to claim 1, characterized in that, In the step of dividing the target data block in the diffusion magnetic resonance imaging data based on the target voxel, a three-dimensional data region of a preset size is divided as the target data block with the position of the target voxel in the diffusion magnetic resonance imaging data as the center.

4. The method for mapping microstructures based on diffusion tensor distribution according to claim 1, characterized in that, In the step of processing the two-dimensional matrix sequentially through a feature embedding layer and a position encoding layer to obtain an embedded feature vector, the feature embedding layer is provided with at least one fully connected layer connected in sequence, and the two-dimensional matrix outputs a first feature sequence through the last fully connected layer of the feature embedding layer; the position encoding layer adds a preset classification label vector to the beginning of the first feature sequence to obtain an initial position vector, processes the initial position vector using three-dimensional sine and cosine encoding to obtain a position vector, and concatenates the position vector and the first feature sequence to obtain the embedded feature vector.

5. The method for mapping microstructures based on diffusion tensor distribution according to claim 1, characterized in that, In the step of inputting the embedded feature vector into the Transformer encoder, the Transformer encoder uses an attention mechanism to encode the features and obtain the encoded feature vector. The Transformer encoder includes multiple Transformer blocks arranged in sequence, and the encoded feature vector is output through the last Transformer block.

6. The method for mapping microstructures based on diffusion tensor distribution according to claim 1, characterized in that, In the step of inputting the encoded feature vector into the decoder, which includes a sparse dictionary mapping branch and a prediction branch, wherein the sparse dictionary mapping branch outputs a brain component distribution vector and the prediction branch outputs regression calculation parameters, the prediction branch outputs regression calculation parameters through a fully connected layer.

7. The method for mapping microstructures based on diffusion tensor distribution according to claim 6, characterized in that, In the step of inputting the encoded feature vector into the decoder, which includes a sparse dictionary mapping branch and a prediction branch, wherein the sparse dictionary mapping branch outputs a brain component distribution vector and the prediction branch outputs regression calculation parameters, the sparse dictionary mapping branch is sequentially configured with multiple sub-processing units and a layer normalization layer. The input segment of the decoder is residually connected to the output of each sub-processing unit. The encoded feature vector is concatenated with the output of each sub-processing unit and output to the next sub-processing unit. Each sub-processing unit includes a threshold layer and a convolutional layer connected in sequence. The threshold layer is used to set values ​​whose absolute value is less than a threshold to 0. The convolutional layer is used to perform convolution processing. The layer normalization layer is connected to the last sub-processing unit and outputs a brain component distribution vector.

8. The method for mapping microstructures of diffusion tensor distribution based on deep learning according to any one of claims 1 to 7, characterized in that, The brain component distribution vector has three feature dimensions, with values ​​representing the proportions of white matter, gray matter, and cerebrospinal fluid, respectively. The regression calculation parameters include the expected value of isotropic diffusion, the expected value of anisotropic diffusion, the variance of the expected value of anisotropic diffusion, and the variance of the expected value of isotropic diffusion. In the step of constructing microstructure parameter maps corresponding to various brain components based on the brain component mapping pixels of all target voxel locations, the parameter of each dimension in the brain component distribution vector is multiplied by each regression calculation parameter to obtain multiple brain component mapping pixels. A microstructure parameter map is constructed based on the brain component mapping pixels of all target voxels for each type of brain component.

9. The method for mapping microstructures based on diffusion tensor distribution according to claim 8, characterized in that, In the step of constructing a microstructure parameter map based on the brain component mapping pixels of all the target voxels, the position of the corresponding brain component mapping pixel is determined based on the position of the target voxel in all voxels, and the corresponding position is rendered using the brain component mapping pixels to construct the microstructure parameter map of the corresponding type.

10. A deep learning-based diffusion tensor distribution organization microstructure mapping device, characterized in that, The device includes a computer device, which includes a processor and a memory, wherein computer instructions are stored in the memory, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method as described in any one of claims 1 to 9.

Images