SIND network construction method and system based on KD-Tree multi-source feature fusion

Abstract

The invention discloses an SIND network construction method and system based on KD-Tree multi-source feature fusion, and the method comprises the steps: obtaining individual QSM data, mapping the individual QSM data to a preset standardized template space to obtain a standardized brain image, and segmenting the brain image into different brain regions; encoding the QSM data of each brain region into multi-dimensional vector data, and constructing a KD tree of each brain region based on a variance-sensitive dynamic division strategy and median segmentation; calculating KL divergence of any two brain regions based on a KD tree, and determining an SIND value between the two brain regions according to the KL divergence; and carrying out dimension reduction analysis on the SIND value by adopting diffusion mapping, and determining the magnetic sensitive gradient characteristics of the brain region. According to the method, efficient data processing, multi-source feature fusion, KL divergence measurement and diffusion mapping analysis of the KD Tree are combined, the magnetic sensitive gradient features of the brain regions can be accurately analyzed, the structure and function relation between the brain regions is disclosed, and a powerful tool is provided for brain science research.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic resonance imaging, and particularly to a method and system for constructing a SIND network based on multi-source feature fusion of KD-Tree. Background Art

[0002] Quantitative susceptibility mapping (QSM) is a magnetic resonance imaging (MRI) technique that can accurately and quantitatively measure the spatial distribution of magnetic susceptibility. QSM is mainly used to analyze the distribution of iron elements in tissues or structures. Especially in the field of brain science, it can provide detailed information about substances such as hemoglobin, iron deposition, and calcification. By characterizing tissue magnetic susceptibility differences, it provides new ideas for biomarker discovery of neurodegenerative diseases such as Parkinson's disease and multiple system atrophy.

[0003] Currently, the QSM network construction method has the following technical bottlenecks: Firstly, the feature dimension is single. Existing technologies (such as the Gaussian-KL method) only use the total QSM imaging data and only construct a single positive / negative magnetic susceptibility network, which cannot effectively fuse the complementary information of the two magnetic susceptibility characteristics and ignores important information such as myelin metabolism carried by negative magnetic susceptibility. Secondly, the operation efficiency is low. The traditional KL (Kullback-Leibler) divergence calculation method relies on Gaussian kernel function smoothing, and the time complexity reaches O(n2). For high-precision analysis tasks such as whole-brain partitioning, the amount of calculation is huge. Thirdly, the calculation accuracy is insufficient. The traditional KL divergence calculation method uses a Gaussian kernel with a fixed bandwidth for probability density estimation, resulting in the loss of small-scale features and being unable to adapt to the heterogeneous feature distributions of different brain regions.

[0004] Therefore, there is an urgent need to propose a method for constructing a SIND (Susceptibility Inverse Divergence) network based on multi-source feature fusion of KD-Tree, which can accurately analyze the magnetic susceptibility gradient features of brain regions, reveal the structural and functional relationships between brain regions, improve the calculation efficiency and data analysis function, and provide a powerful tool for brain science research and clinical applications. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for constructing a SIND network based on multi-source feature fusion of KD-Tree to solve the technical problems that the existing QSM network method only constructs a single positive / negative magnetic susceptibility network, ignores the magnetic susceptibility characteristics, and has high complexity in brain region features, large calculation amount when processing large-scale data, and insufficient accuracy.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for constructing a SIND network based on KD-Tree multi-source feature fusion, including:

[0008] Obtain individual QSM data, map the individual QSM data into a preset standardized template space to obtain a standardized brain image, and segment the brain image into different brain regions;

[0009] Encode the QSM data of each brain region into multi-dimensional vector data, and construct a KD-tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median splitting;

[0010] Calculate the KL divergence between any two brain regions based on the KD-tree, and determine the SIND value between the two brain regions according to the KL divergence;

[0011] Perform dimensionality reduction analysis on the SIND value by using diffusion mapping to determine the susceptibility gradient feature of the brain region.

[0012] Further, the encoding of the QSM data of each brain region into multi-dimensional vector data includes:

[0013] Use the voxel set V = {v1,..v i ,v N} to represent each brain region;

[0014] Among them, each voxel is composed of a susceptibility vector vi = (QSM + ,QSM - ,QSM) in three dimensions, where QSM + represents the paramagnetic feature, QSM - represents the diamagnetic feature, and QSM represents the comprehensive magnetic susceptibility signal.

[0015] Further, the construction of the KD-tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median splitting includes:

[0016] Calculate the susceptibility variances in three dimensions, determine the dimension with the largest variance, and select the dimension with the largest variance in the current subspace for partitioning at each level;

[0017] After sorting the data in the current dimension, select the median as the splitting point.

[0018] Further, when constructing the KD-tree for each brain region, the Best-Bin-First (BBF) algorithm is used to accelerate the nearest neighbor search to reduce the query complexity.

[0019] Further, the calculation of the KL divergence between any two brain regions based on the KD-tree includes:

[0020] The calculation formula for the KL divergence between any two brain regions i and j is:

[0021]

[0022] Among them, Ω i represents the set of all voxels in the i-th brain region, and N i =|Ω i | represents the total number of voxels in the i-th brain region, and P i (v k ) represents the local density estimation probability centered on voxel v k . represents the probability of the nearest neighbor density estimation of v k in the j-th brain region.

[0023] Furthermore, the determination of the SIND value between the two brain regions according to the KL divergence includes:[[]]

[0024] The calculation formula for the similarity value between any two brain regions i and j is:[[]]

[0025]

[0026] where KL(i,j) represents the KL divergence between brain regions i and j.[[]]

[0027] Furthermore, the use of diffusion mapping to perform dimensionality reduction analysis on the SIND value to determine the susceptibility gradient characteristics of the brain region includes:[[]]

[0028] Converting the SIND value into angular similarity to determine the affinity matrix between brain regions;[[]]

[0029] Taking the eigenvectors corresponding to the first k largest non-zero eigenvalues of the affinity matrix as the principal component susceptibility gradients of the SIND network.[[]]

[0030] Furthermore, the conversion of the SIND value into angular similarity to determine the affinity matrix between brain regions is expressed by the formula:[[]]

[0031]

[0032] where A ij represents the affinity matrix of brain regions i and j, SIND i,j represents similarity, and σ represents the standard deviation of SIND i,j .

[0033] Furthermore, the mapping of the individual's QSM data into a preset standard template space includes:[[]]

[0034] Aligning the individual QSM data with the MNI152 template and using third-order B-spline interpolation to ensure the alignment accuracy of the anatomical boundary to eliminate the spatial deviation caused by individual differences.[[]]

[0035] Further, the segmentation of the standardized brain image into different brain regions includes:

[0036] Performing cortical segmentation and subcortical basal ganglia region segmentation on the brain image using the Brainnetome 246 template to form multiple different brain regions.

[0037] The present invention also provides a SIND network construction system based on KD-Tree multi-source feature fusion, including an acquisition and segmentation module, an encoding and construction module, a calculation and determination module, and an analysis module:

[0038] The acquisition and segmentation module is used to acquire individual QSM data, map the individual QSM data into a preset standardized template space to obtain a standardized brain image, and segment the brain image into different brain regions;

[0039] The encoding and construction module is used to encode the QSM data of each brain region into multi-dimensional vector data, and construct a KD tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median segmentation;

[0040] The calculation and determination module is used to calculate the KL divergence between any two brain regions based on the KD tree, and determine the SIND value between the two brain regions according to the KL divergence;

[0041] The analysis module is used to perform dimensionality reduction analysis on the SIND value using diffusion mapping to determine the magnetic susceptibility gradient characteristics of the brain region.

[0042] Compared with the prior art, the main advantages of the SIND network construction method and system based on KD-Tree multi-source feature fusion proposed by the present invention are:

[0043] (1) Multi-source feature fusion paradigm: It realizes the spatial joint similarity analysis of positive / negative QSM features for the first time; since positive / negative QSM features reflect the differences in regions with different magnetic susceptibility properties, through the spatial joint analysis of these features, the mutual relationship between lesions and brain regions can be understood more comprehensively, breaking the limitation of traditional single-feature analysis and enhancing the recognition ability of complex pathological changes.

[0044] (2) Dynamic topology analysis framework: This framework is not only applicable to single-case analysis, but also supports scalable network modeling across diseases (such as Alzheimer's disease, stroke, etc.), expanding the applicable range of network analysis methods and improving their application value.

[0045] (3) Precise biomarker: Revealing the changes in the brain's micro-structure through magnetic susceptibility gradient characteristics can reflect the structural changes in the brain more accurately compared with traditional anatomy-based biomarkers.

[0046] In summary, the present method combines the efficient data processing of the KD Tree, multi-source feature fusion, KL divergence measurement, and diffusion mapping analysis, and can accurately analyze the susceptibility gradient features of brain regions, revealing the structural and functional relationships between brain regions. The method of the present invention has high computational efficiency, accurate partitioning ability, and powerful data analysis function, providing a powerful tool for brain science research and clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic flowchart of the SIND network construction method based on KD-Tree multi-source feature fusion provided by the present invention;

[0048] Figure 2 It is a schematic flowchart of the data processing provided by the present invention;

[0049] Figure 3 It is a schematic structural diagram of the SIND network construction system based on KD-Tree multi-source feature fusion provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The following will specifically describe the preferred embodiments of the present invention with reference to the accompanying drawings, in which the accompanying drawings form a part of the present application and are used together with the embodiments of the present invention to explain the principle of the present invention, rather than to limit the scope of the present invention.

[0051] Before introducing the embodiments of the present invention, the inventive concept of the present invention will be described first.

[0052] The existing medical QSM network construction technology has the following problems: First, the traditional QSM network construction method (such as the Gaussian-KL method) can only construct a single positive / negative susceptibility network, and cannot effectively fuse the complementary information of the two magnetization characteristics, and the feature dimension is relatively single.

[0053] Secondly, the traditional KL divergence calculation method depends on Gaussian kernel function smoothing, as shown in formula (1):

[0054]

[0055] Where P_i is the distribution kernel function after Gaussian smoothing.

[0056] As the input scale increases, the running time of the algorithm will increase at a speed of n2, and the time complexity reaches O(n2), which is inefficient in processing large-scale high-dimensional data and difficult to adapt to refined analysis tasks such as whole-brain partitioning.

[0057] Finally, the traditional KL divergence calculation method uses a Gaussian kernel with a fixed bandwidth for probability density estimation, as shown in formula (2):

[0058]

[0059] Among them, h is the fixed bandwidth, x i is a sample point, and n and m are the number of samples.

[0060] The fixed bandwidth h cannot adapt to the heterogeneous feature distributions of different brain regions, which will lead to the loss of small-scale features.

[0061] To solve the above existing problems, this solution realizes the spatial joint similarity analysis of positive / negative QSM features through the collaborative calculation of dual-source features based on the KD Tree, and encodes the positive and negative magnetic susceptibilities into three-dimensional feature vectors (QSM+, QSM-, QSM); constructs a three-level progressive network system in the SIND network framework, providing a powerful tool for brain science research and clinical applications.

[0062] Please refer to Figure 1 , this embodiment provides a method for constructing a SIND network based on KD-Tree multi-source feature fusion, including:

[0063] Step S101: Obtain individual QSM data, map the individual QSM data to a preset standardized template space to obtain a standardized brain image, and segment the brain image into different brain regions;

[0064] Step S102: Encode the QSM data of each brain region into multi-dimensional vector data, and construct a KD tree for each brain region through a variance-sensitive dynamic partitioning strategy and median splitting;

[0065] Step S103: Calculate the KL divergence between any two brain regions based on the KD tree, and determine the SIND value between the two brain regions according to the KL divergence;

[0066] Step S104: Analyze the SIND value using the diffusion map to determine the magnetic susceptibility gradient features of the brain regions.

[0067] The method for constructing a SIND network based on KD-Tree multi-source feature fusion provided in this embodiment, by encoding the QSM data of brain regions into multi-dimensional vectors, constructing a KD tree, and combining KL divergence and diffusion map analysis, can not only accurately measure the magnetic susceptibility gradient features between brain regions, but also efficiently process and analyze large-scale high-dimensional data, providing an efficient, accurate and flexible method for in-depth research in the fields of brain science and medical imaging, etc., to reveal the relationships between brain regions and their changes under different states, and has high practical value.

[0068] As a preferred embodiment, in step S101, mapping the individual QSM data to a preset standardized template space to obtain a standardized brain image includes:

[0069] Align the individual QSM data with the MNI152 template, and use third-order B-spline interpolation to ensure the alignment accuracy of anatomical boundaries, so as to eliminate the spatial deviation caused by individual differences.

[0070] As a specific embodiment, in order to register the brain region data to the standard deformation field and realize the spatial standardization of the data, in the actual processing process, the following two steps are first performed on the obtained individual QSM data:

[0071] The first step: Rigid registration: Use the SyN algorithm of ANTs to map the individual QSM to the Age-specific MNI152 template (parameter settings: step size = 0.25, iterations = 100×100×50);

[0072] The second step: Sub-millimeter interpolation: Apply third-order B-spline interpolation to ensure the alignment accuracy of anatomical boundaries.

[0073] Through the above two-step processing, the individual differences between the subjects' cranial brains are eliminated, and a spatial reference is established for subsequent partition calculations.

[0074] As a preferred embodiment, in step S101, the brain image is segmented into different brain regions, including:

[0075] Use the Brainnetome 246 template to perform cortical segmentation and subcortical basal ganglia region segmentation on the brain image to form multiple different brain regions.

[0076] As a specific embodiment, cortical segmentation is to divide the brain image into 210 functional regions based on the cytoarchitecture probability map, and then refine the subcortex, and use fuzzy clustering to segment the basal ganglia into 36 subregions.

[0077] The specific Brainnetome 246 template used can refer to the literature: Fan L, Li H, Zhuo J, Zhang Y, Wang J, Chen L, Yang Z, Chu C, Xie S, Laird AR, Fox PT, Eickhoff SB, Yu C, Jiang T. The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional Architecture. Cereb Cortex. 2016 Aug; 26(8): 3508-26. doi: 10.1093 / cercor / bhw157..

[0078] Using the Brainnetome 246 template as a reference template can provide a consistent spatial benchmark and ensure the consistency of segmentation results among different individuals. The standardized template provides a unified comparison benchmark for different individuals and different studies, enabling effective comparison of results between different studies, facilitating cross-individual and cross-study comparisons, and improving the reproducibility and comparability of research.

[0079] It provides a solid foundation for the standardization, precise segmentation, and subsequent analysis of brain image data. Through high-precision spatial registration, detailed brain region division, and the use of a standardized template, the reliability of research results can be ensured, and the comparability and consistency across individuals and studies can be improved. These operations provide strong support for large-scale data processing, functional research, and clinical applications in neuroimaging research.

[0080] As a preferred embodiment, in step S102, encoding the QSM data of each brain region into multi-dimensional vector data includes:

[0081] Using the voxel set V = {v1,..v i .,v N} to represent each brain region;

[0082] Among them, each voxel is composed of a susceptibility vector vi = (QSM + ,QSM - ,QSM) in three dimensions, where QSM + represents the paramagnetic feature, QSM - represents the diamagnetic feature, and QSM represents the comprehensive susceptibility signal.

[0083] As a specific embodiment, by encoding the QSM data of each brain region into multi-dimensional vector data and representing it using the voxel set, the magnetic characteristics of the brain region can be carefully captured, enhancing the combination of spatial and functional information, and providing an optimized data structure for subsequent deep learning model processing.

[0084] As a preferred embodiment, in step S103, constructing the KD tree of each brain region based on the variance-sensitive dynamic partitioning strategy and median splitting includes:

[0085] Calculating the susceptibility variances in three dimensions, determining the dimension with the largest variance, and selecting the dimension with the largest variance in the current subspace for partitioning at each level;

[0086] After sorting the data in the current dimension, selecting the median as the splitting point.

[0087] As a specific embodiment, when constructing a three-dimensional feature tree for each brain region, first calculate the variances σ k ={σQsM+ , σQSM - , σQSM}, select the dimension with the largest variance Among them, σ represents the variance of the k-th dimensional data, realizing the segmentation of the maximum discrete direction; the dataset is segmented by the d-dimensional median, and the left and right subtrees are recursively constructed. By selecting the dimension with the largest variance for dataset segmentation, it can ensure that each segmentation can distribute the data into the left and right subtrees as evenly as possible, rather than simply by coordinate or dimension segmentation. Through recursive construction and the selection of the dimension with the largest variance, each segmentation can take into account the characteristics of the data distribution, which helps to maintain the balance of the tree.

[0088] It should be noted that in order to balance query efficiency and accuracy, the leaf nodes are set to contain at most 30 voxels. Setting the limit that each leaf node contains at most 30 voxels can ensure that the depth of the tree will not be too deep, thus avoiding the low query efficiency caused by too many tree levels. By restricting the leaf nodes, the tree structure can be prevented from becoming too complex, ensuring that the target area can be quickly located during query.

[0089] In some embodiments, when dividing the dimension selection strategy, cKDTree of Scipy adopts a variance-sensitive dynamic division strategy during construction: at each level, the dimension with the largest variance in the current subspace is selected for division, and this selection makes the data distribution in the divided subspace more balanced. For example, if QSM+ (assuming in dimension 1) has a larger variance than QSM- (dimension 2), dimension 1 will be preferentially divided.

[0090] In addition, in the division value determination rule, the median cut strategy is used to select the division value, that is: after sorting the data in the current dimension, the median is selected as the cut point.

[0091] Regarding the balance characteristic of the tree, cKD Tree defaults to constructing a completely balanced tree, and the two child nodes of each node contain approximately equal numbers of samples.

[0092] By selecting the dimension with the largest variance for segmentation and using the median cut strategy, the balance of the feature tree can be ensured, the query efficiency can be optimized, and at the same time, the uniformity of data distribution can be improved. The balance characteristic of the tree not only helps to improve the efficiency of query and positioning, but also enhances the stability of data processing, especially in the scenario of large-scale datasets. In addition, setting the size limit of the leaf nodes can prevent the tree from becoming overly complex, and the two child nodes of each node contain approximately equal numbers of samples, ensuring the accuracy and efficiency of the model and achieving the best balance in practical applications.

[0093] Furthermore, when constructing the KD tree for each brain region, the BBF algorithm is used to accelerate the nearest neighbor search to reduce the complexity of the query.

[0094] Specifically, the Best-Bin-First (BBF) algorithm is an optimization strategy for KD-tree search. By selecting the tree nodes most likely to contain the nearest neighbors for search, it avoids traversing the entire tree. Specifically, the BBF algorithm accelerates the search through the following steps:

[0095] First, start the search from the root node and adopt the method based on "optimal bin selection" to start from the "bin" or node closest to the query point.

[0096] Secondly, the BBF algorithm determines whether to continue searching other regions by calculating the distance between the current node and the query point and the bounding box of the current node. If the nearest neighbors of the current node are already close to the distance of the query point and the bounding boxes of other nodes cannot provide better results, the search will be pruned to avoid wasting computing resources. Combining the BBF (Best Bin First) algorithm to accelerate the nearest neighbor search, the query complexity is reduced to O(n log n).

[0097] By selecting the regions most likely to contain the nearest neighbors for search, the BBF algorithm can significantly reduce the inspection of irrelevant or distant nodes, thus accelerating the search process. Compared with traditional search methods (such as brute force or conventional KD-tree depth-first search), BBF can determine the optimal nearest neighbor points faster.

[0098] As a preferred embodiment, in step S103, calculating the KL divergence between any two brain regions based on the KD tree includes:

[0099] The calculation formula for the KL divergence between any two brain regions i and j is:

[0100]

[0101] where, Ω i represents the set of all voxels in the i-th brain region, N i =|Ω i | represents the total number of voxels in the i-th brain region, P i (v k ) represents the probability of local density estimation centered on the voxel v k , B(v k , r) represents the spherical domain centered on the voxel v k with radius r, represents the probability of nearest neighbor density estimation in the j-th brain region corresponding to v k , represents finding v in the KD Tree of the j-th brain region kThe m-th nearest neighbor distance, k represents the adaptive number of neighbors (the recommended value range is 15 ≤ k ≤ 30), and vol() represents the function for calculating the volume of the spherical domain (in three-dimensional space is ).

[0102] As a preferred embodiment, in step S103, the determining the SIND value between the two brain regions according to the KL divergence includes:

[0103] The calculation formula for the similarity value between any two brain regions i and j is:

[0104]

[0105] where KL(i, j) represents the KL divergence between brain regions i and j.

[0106] As a preferred embodiment, in step S104, the performing dimensionality reduction analysis on the SIND value by using diffusion mapping to determine the susceptibility gradient feature of the brain region includes:

[0107] Converting the SIND value into angular similarity to determine the affinity matrix between brain regions;

[0108] Taking the eigenvectors corresponding to the first k largest non-zero eigenvalues of the affinity matrix as the principal component susceptibility gradients of the SIND network.

[0109] As a preferred embodiment, the converting the SIND value into angular similarity to determine the affinity matrix between brain regions is represented by the formula:

[0110]

[0111] where A ij represents the affinity matrix of brain regions i and j, SIND i,j represents the similarity, and σ represents the standard deviation of SIND i,j .

[0112] Furthermore, performing principal component extraction on the affinity matrix, and taking the first 3 non-trivial eigenvectors (i.e., non-zero eigenvectors) of the affinity matrix as the principal components of the SIND network as the susceptibility gradients.

[0113] As Figure 2 shown, Figure 2 shows the data flow diagram during actual processing.

[0114] As Figure 3 shown, the present invention also provides a system for constructing a SIND network based on KD-Tree multi-source feature fusion, which includes an acquisition and segmentation module, an encoding and construction module, a calculation and determination module, and an analysis module:

[0115] The acquisition and segmentation module is used to acquire individual QSM data, map the individual QSM data into a preset standardized template space to obtain a standardized brain image, and segment the brain image into different brain regions;

[0116] The encoding and construction module is used to encode the QSM data of each brain region into multi-dimensional vector data, and construct a KD tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median segmentation;

[0117] The calculation and determination module is used to calculate the KL divergence between any two brain regions based on the KD tree, and determine the SIND value between the two brain regions according to the KL divergence;

[0118] The analysis module is used to perform dimensionality reduction analysis on the SIND value by using diffusion mapping to determine the susceptibility gradient characteristics of the brain region.

[0119] The method and system for constructing a SIND network based on KD-Tree multi-source feature fusion proposed by the present invention solve the problem of insufficient single susceptibility feature through the fusion of multi-source features; organize data efficiently through the KD tree structure and perform nearest neighbor search quickly, significantly improving the calculation efficiency; overcome the accuracy deficiency of traditional methods in dealing with heterogeneous features through dynamic partitioning strategies and flexible bandwidth selection. Through the method of the present invention, when processing QSM data, it can effectively improve the calculation efficiency, solve the accuracy problem, and can more comprehensively reveal the structural and functional relationships between brain regions, providing a powerful tool for brain science research.

[0120] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a SIND network based on KD-Tree multi-source feature fusion, characterized in that Including: Obtain individual QSM data, map the individual QSM data into a preset standardized template space to obtain a standardized brain image, and segment the brain image into different brain regions; Encode the QSM data of each brain region into multi-dimensional vector data, and construct a KD tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median splitting; Calculate the KL divergence between any two brain regions based on the KD tree, and determine the SIND value between the two brain regions according to the KL divergence; Perform dimensionality reduction analysis on the SIND value using the diffusion mapping algorithm to determine the susceptibility gradient characteristics of the brain regions.

2. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 1, wherein The encoding of the QSM data of each brain region into multi-dimensional vector data includes: Use the voxel set V = {v1,..v i .,v N} to represent each brain region; Among them, each voxel is composed of three-dimensional susceptibility vectors vi = (QSM + , QSM - , QSM), where QSM + represents the paramagnetic feature, QSM - represents the diamagnetic feature, and QSM represents the comprehensive susceptibility signal.

3. The method for constructing the SIND network based on KD-Tree multi-source feature fusion according to claim 2, wherein The construction of the KD tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median splitting includes: Calculate the susceptibility variances in three dimensions, determine the dimension with the largest variance, and select the dimension with the largest variance in the current subspace for partitioning at each level; After sorting the data in the current dimension, select the median as the splitting point.

4. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 3, wherein When constructing the KD tree for each brain region, the BBF algorithm is used to accelerate the nearest neighbor search to reduce the query complexity.

5. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 1, characterized in that, The calculation of the KL divergence between any two brain regions based on the KD tree includes: The calculation formula for the KL divergence between any two brain regions i and j is: where, Ω i represents the set of all voxels in the i-th brain region, and N i = |Ω i | represents the total number of voxels in the i-th brain region, and P i (v k ) represents the local density estimation probability centered on voxel v k , and represents the nearest neighbor density estimation probability of v k in the j-th brain region.

6. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 1, wherein The determination of the SIND value between the two brain regions according to the KL divergence includes: The calculation formula for the similarity value between any two brain regions i and j is: Where KL(i,j) represents the KL divergence between brain regions i and j.

7. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 1, wherein The use of diffusion mapping to perform dimensionality reduction analysis on the SIND value to determine the susceptibility gradient characteristics of the brain regions includes: Convert the SIND value to angular similarity to determine the affinity matrix between brain regions; Take the eigenvectors corresponding to the first k largest non-zero eigenvalues of the affinity matrix as the principal component susceptibility gradients of the SIND network.

8. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 7, wherein The conversion of the SIND value to angular similarity to determine the affinity matrix between brain regions is expressed by the formula: Among them, A ij represents the affinity matrix of brain regions i and j, and SIND i,j represents the similarity, and σ represents the standard deviation of SIND i,j .

9. The method for constructing a SIND network based on KD-Tree multi-source feature fusion according to claim 1, wherein The mapping of the individual's QSM data into a preset standard template space to obtain a standardized brain image and the segmentation of the brain image into different brain regions include: Align the individual QSM data with the MNI152 template, and use third-order B-spline interpolation to ensure the alignment accuracy of the anatomical boundaries to eliminate the spatial deviation caused by individual differences, thereby obtaining a standardized brain image; Use the Brainnetome 246 template to perform cortical segmentation and subcortical basal ganglia region segmentation on the standardized brain image to form multiple different brain regions.

10. A SIND network construction system based on KD-Tree multi-source feature fusion, characterized in that, Including an acquisition and segmentation module, an encoding and construction module, a calculation and determination module, and an analysis module: The acquisition and segmentation module is used to obtain individual QSM data, map the individual QSM data into a preset standardized template space to obtain a standardized brain image, and segment the brain image into different brain regions; The encoding and construction module is used to encode the QSM data of each brain region into multi-dimensional vector data, and construct a KD tree for each brain region based on a variance-sensitive dynamic partitioning strategy and median splitting; The calculation and determination module is used to calculate the KL divergence between any two brain regions based on the KD tree, and determine the SIND value between the two brain regions according to the KL divergence; The analysis module is used to perform dimensionality reduction analysis on the SIND value by using diffusion mapping to determine the susceptibility gradient characteristics of the brain region.

Images