An individualized brain atlas partitioning system based on a multi-dimensional morphological lateralization inverse divergence network

Abstract

The application relates to the technical field of neural image processing, and particularly discloses a brain atlas division system based on a multi-modal multi-dimensional lateralization index similarity network. The method first performs spatial uniform random sampling on the left hemisphere cortical surface of an individual, and extracts the 5-layer neighborhood of the sampling points by using the grid topological connection relationship; then, according to the cross-hemisphere vertex correspondence, the symmetric neighborhood is positioned in the right hemisphere, and the lateralization index (LI) distribution of the cortical features is calculated; by kernel density estimation modeling and morphological counter divergence algorithm, the LI-MIND correlation matrix representing the whole brain symmetry is constructed; finally, the spectral clustering algorithm is used for feature decomposition and dimension reduction of the matrix, the optimal clustering number is determined according to the contour coefficient, and the smooth individualized brain region division atlas is generated. By introducing the topological neighborhood and the lateralization distribution characteristics, the problem that the traditional brain atlas cannot effectively capture the individual organization left-right hemisphere difference is solved, and the brain region division scheme depending on the lateralization information is provided.

Description

Technical Field

[0001] This invention relates to the fields of neuroimaging processing, biomedical engineering, and computer vision, specifically to an individualized brain atlas segmentation system based on a multidimensional lateralization index similarity network. Background Technology

[0002] Brain atlases (brain maps) are a fundamental framework for understanding brain functional organization patterns, exploring neural circuit mechanisms, and conducting clinical neuropsychiatric research. Traditional brain atlas divisions typically rely on population-average templates (such as Desikan-Killiany or Glasser maps), which assume a high degree of consistency in brain anatomy and functional organization across all individuals. However, recent studies have revealed significant individual heterogeneity at the apex level within the cerebral cortex, manifested in functional connectivity patterns, cortical morphology, and cognitive performance. Therefore, constructing precise, individual-specific brain atlases is crucial for precision medicine and personalized brain network analysis.

[0003] The left and right hemispheres of the brain exhibit significant lateralization in their anatomical structure and functional division. This lateralization is not only an important marker of human brain evolution but also closely related to higher functions such as language, logic, and spatial cognition. However, existing individualized brain partitioning methods often focus on temporal correlations or morphological similarities within a single hemisphere, neglecting the synergistic interactions between symmetrical regions of the left and right hemispheres and the unique value of the lateralization index (LI) in characterizing individual tissue heterogeneity.

[0004] Furthermore, current brain mapping methods often employ simple Euclidean distance neighborhoods when extracting local features, which are ill-suited to the complex manifold structure of the cerebral cortex and prone to feature contamination across sulci and gyri. Simultaneously, single-modality data cannot comprehensively characterize the multidimensional biological properties of the cortex (such as thickness, myelination, and connectivity), limiting the robustness and accuracy of brain atlases in a multimodal fusion context. Therefore, how to integrate multidimensional lateralized information and combine it with the cortical topological manifold structure to achieve high-precision individualized brain mapping is a pressing technical problem in the field of neuroimaging. Summary of the Invention

[0005] In order to generate individualized brain region maps that incorporate cortical feature deviation information, this invention makes the following two innovations:

[0006] Innovation Point 1: A multidimensional lateralization feature anti-divergence association network based on region-independent fusion was constructed. This invention breaks away from the dependence of traditional brain network analysis on predefined anatomical partition templates (Atlas). By performing spatially uniform random sampling across the entire brain cortex and combining it with 5-layer topological neighborhood extraction to obtain local multidimensional lateralization features, data-driven sampling independent of any prior atlas is achieved. This method utilizes kernel density estimation and morphological anti-divergence algorithms to model the local lateralization information of each sampling point as a probability distribution and calculate the statistical similarity between distributions, thereby constructing a whole-brain LI-MIND association matrix. This region-independent modeling approach can accurately capture subtle, irregularly distributed lateralization patterns on the cortex, effectively avoiding the spatial bias caused by traditional partition templates and significantly improving the sensitivity of brain feature description.

[0007] Innovation Point Two: This invention achieves personalized brain atlas segmentation driven by the similarity of a multidimensional lateralized feature inverse divergence correlation network. It proposes a brain region segmentation framework centered on an individual lateralized similarity network. Unlike traditional deformable registration methods based on group templates, this scheme uses the generated multidimensional lateralized feature inverse divergence correlation matrix as an individual-specific graph structure, employing spectral clustering algorithms and adaptive parameter optimization mechanisms for feature decomposition and dimensionality reduction. By calculating the silhouette coefficient, the optimal number of partitions best matching the individual's brain tissue characteristics is automatically determined, and the clustering labels are mapped back to the anatomical cortical surface in real time. The brain atlas generated in this process highly respects the individual brain's anatomical manifold and left-right hemisphere heterogeneity, producing a brain partitioning scheme with clear boundaries, high anatomical coherence, and individual specificity, providing a reliable anatomical reference framework for subsequent precision medicine and personalized brain network research.

[0008] The technical solution adopted in this invention is mainly achieved through the following processing flow:

[0009] Step 1: Resting-state functional magnetic resonance imaging (rs-fMRI) preprocessing: Obtain the resting-state functional magnetic resonance imaging (rs-fMRI) data of the subjects and perform standard preprocessing procedures on the image data, including head motion correction, temporal tomography correction, spatial normalization to standard template space, bandpass filtering to remove physiological noise and low-frequency drift, and regression to remove obfuscating signals such as white matter and cerebrospinal fluid. Then, the preprocessed volume data is projected onto the cortical surface grid to obtain the blood oxygen level dependent (BOLD) signal matrix of vertex number × time point number.

[0010] Step 2: Hierarchical segmentation based on vertex spatial coordinates: Utilizing the coordinate positions of each vertex on the cortical surface in standard 3D space, hierarchical segmentation is performed according to the coordinate axes. Specifically, firstly, the vertices are divided into M1 groups based on the value of the first coordinate axis (e.g., the x-axis). Then, within each group, it is further divided into M2 subgroups based on the quantile of the second coordinate axis (e.g., the y-axis). Finally, each subgroup is divided into M3 blocks based on the third coordinate axis (e.g., the z-axis), ultimately dividing the cortical surface into... Each space block records the block-to-block index of each vertex and the centroid coordinates of each block, which are used for spatial distance offset calculation in subsequent block-level global interaction modeling.

[0011] Step 3: Self-supervised encoder pre-training: Using a mask reconstruction self-supervised learning strategy, some time points of the input time series are randomly masked, and a hierarchical encoder is trained to reconstruct the masked signal, so that the encoder learns robust spatiotemporal representation capabilities.

[0012] Step 4: Intra-block independent temporal feature extraction: Input the vertex time series in each block into a one-dimensional convolutional network to extract the temporal dynamic features of each vertex independently, and obtain the local feature matrix of all vertices in the whole brain.

[0013] Step 5: Block-level global interaction modeling: The vertex features in each block are aggregated into a block-level representation vector through attention pooling. The spatial distance bias matrix is ​​calculated using the block centroid coordinates. The block-level representation is input into a Transformer encoder with spatial bias to establish long-distance functional connections between blocks, resulting in a globally enhanced block-level representation.

[0014] Step 6, Vertex-level Differentiated Context Integration: Each vertex generates a query vector based on its local features. The attention weights for all blocks are calculated through a cross-block attention mechanism, and the weighted aggregation of the globally enhanced block-level representation breaks the limitation of uniform features within blocks in traditional partitioning methods. This allows different vertices within the same anatomical region to aggregate global information differentially based on their own functional characteristics, generating highly individualized fused feature representations. Then, through a gating modulation mechanism, the scaling factor, offset, and gating strength are learned to achieve adaptive fusion of local features and global context, resulting in the final fused feature representation.

[0015] Step 7: Construction of a multi-relationship graph: Based on the fusion features obtained in Step 6, calculate the functional similarity between vertices, select several vertices with the most relevant functions to construct the connection relationship of the graph, and establish the functional topological skeleton; then calculate the geodesic distance of connected vertices on the cortical surface, and use the Gaussian kernel function to convert the distance into edge weights, thereby constructing a weighted cortical graph that integrates long-distance functional associations and local spatial constraints.

[0016] Step 8: Two-stage cluster center initialization: The improved K-means++ algorithm is used to initialize the cluster centers. When iteratively selecting each new center, a medium-sized candidate set with the greatest functional differences is first selected based on feature space distance (Top-M selection). Then, weighted probability sampling is performed within the candidate set based on geodesic distance. This two-stage strategy ensures that the initial centers are functionally representative while maximizing their physical spatial dispersion.

[0017] Step 9: Graph Attention Feature Enhancement: Input the fused features into the Graph Attention Network (GAT), use edge weights to achieve neighborhood-aware feature aggregation and enhancement, and at the same time protect the original feature information through residual connections.

[0018] Step 10, Adaptive Soft Assignment Calculation: Using the temperature prediction network, dynamically predict the temperature parameters based on the current iteration round and the changing state of the multi-objective loss value; calculate the feature distance from each vertex to all cluster centers, and perform soft assignment calculation by combining the temperature parameters and the Softmax activation function to obtain the membership probability of each vertex to each cluster.

[0019] Step 11: Construction and optimization of multi-objective loss function: Construct a multi-objective loss function that includes cluster compactness loss, contrastive learning loss, distribution entropy loss and center separation loss. Jointly optimize the parameters of graph attention network, temperature prediction network and learnable cluster centers. Through iterative optimization, achieve a balance between functional homogeneity, spatial continuity and inter-class discriminability in the partitioning results.

[0020] Step 12: Generation of individualized brain region map: Based on the final soft assignment matrix, hard labels are assigned to each vertex by taking the maximum membership probability to obtain an individualized brain cortex region map. This map divides the cortical surface into a preset number of functionally consistent and spatially continuous brain regions.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] This invention proposes an individualized brain atlas partitioning system based on a multi-dimensional morphological lateralization inverse divergence network. Compared with existing partitioning strategies, in terms of lateralization feature representation, this invention introduces multimodal morphological indicators to construct a lateralization index (LI) distribution, filling the technical gap in previous partitioning schemes that ignored interhemispheric asymmetry, and significantly improving the biological validity of brain atlases in capturing brain functional specialization and anatomical heterogeneity. In terms of region-independent modeling, this invention uses spatially uniform sampling combined with region-independent LI-MIND correlation matrix construction, realizing data-driven analysis that does not rely on any predefined prior templates. This effectively overcomes the spatial bias problem of traditional methods being limited by fixed anatomical atlas boundaries, and can more accurately capture individual-specific, irregularly distributed cortical tissue units. In terms of anatomical manifold constraints, this invention adopts a 5-layer neighborhood extraction mechanism based on grid topological connectivity, ensuring that feature sampling strictly follows the geometric manifold of the cerebral cortex, effectively avoiding signal contamination caused by traditional Euclidean distance neighborhoods at the sulci and gyri, and ensuring the anatomical accuracy of partition boundaries. Meanwhile, this invention utilizes kernel density estimation modeling and morphological inverse divergence algorithm to elevate point-to-point numerical comparison to a statistical similarity measure between neighborhood distributions, which greatly improves the system's robustness to image noise and registration errors.

[0023] The accompanying drawings form part of this invention. The drawings are provided to offer a further understanding of the invention, and the illustrative embodiments thereof, together with their descriptions, are used to explain this disclosure and do not constitute an undue limitation of the invention.

[0024] Figure 1 This is a schematic diagram of the overall process of a brain atlas classification system based on a multi-dimensional morphological lateralization inverse divergence network provided by the present invention.

[0025] Figure 2 This is a schematic diagram of the process for generating cortical surface and morphological indicators from raw T1-weighted structured magnetic resonance data in this invention.

[0026] Figure 3 This is a schematic diagram illustrating the process of lateralization index extraction, morphological lateralization anti-divergence similarity network construction, and clustering-based cortical partitioning in this invention. Detailed Implementation

[0029] The following detailed description, in conjunction with the accompanying drawings, provides a specific embodiment of the brain mapping system based on a multi-dimensional morphological lateralization inverse divergence network proposed in this invention. Figure 1This is a flowchart illustrating the LI-MIND matrix construction method according to an embodiment of this application. The technical solution of this invention mainly includes two stages: the first stage is morphological feature spatial mapping and lateralization feature extraction, which aims to generate accurate anatomical structures and morphological lateralization feature representations for the vertices of the cerebral cortex; the second stage is the construction of a lateralization morphological similarity network, which uses the extracted lateralization features to complete the correlation matrix calculation between the functional / structural regions of the whole brain; the third stage is the implementation of an individualized brain atlas division system, which performs individualized brain region division based on the constructed lateralization morphological similarity network.

[0030] Phase 1: Spatial Mapping of Hemisphere Morphological Features

[0031] Reference Figure 2 Steps S1 to S3 are performed to achieve mapping of anatomical space and precise pairing between hemispheres.

[0032] S1: Cortical Reconstruction Preprocessing

[0033] The data used in this embodiment comes from the publicly available dataset HCP. The data includes high-resolution T1w structural magnetic resonance images. A fully automated cortical reconstruction workflow was run using the recon-all command in FreeSurfer software. First, the raw T1w data was preprocessed, including motion correction and bias field correction; skull stripping was performed using a hybrid watershed algorithm. After separating the white matter tissue, the white matter surface and pial surface were reconstructed based on voxel intensity gradients; after topological defect correction, a high-precision individual three-dimensional cortical surface mesh was output. Finally, based on the spatial distance between the white matter surface and the pial surface, five basic morphological features of each vertex were extracted: cortical thickness, surface area, cortical volume, mean curvature, and sulcal depth, serving as input for subsequent analysis.

[0034] S2: Homologous vertex registration between hemispheres

[0035] To enable subsequent lateralization feature calculations, a precise one-to-one mapping relationship between the left and right hemisphere vertices needs to be established. This step utilizes FreeSurfer's mris_left_right_register bilateral registration algorithm to register the individual's left and right hemispheres to a perfectly symmetrical standard template space (fsaverage_sym). This algorithm performs non-linear geometric alignment by aligning the folding patterns of cortical sulci and gyri, thereby determining the vertices of a given left hemisphere. The corresponding anatomical location on the right hemisphere And record the whole brain paired vertex index list.

[0036] S3: Morphological index estimation of homogeneous vertices

[0037] Considering registration errors and the slight natural asymmetry in interhemispheric anatomy, directly comparing indices of corresponding vertices may introduce noise. Therefore, this invention employs a K-nearest neighbor (KNN) spatial smoothing strategy to enhance the robustness of index calculation. For a given cortical vertex... Find its homologous pairing point in the contralateral hemisphere. Using the geodesic distance algorithm, the distance is searched on the opposite grid surface. The 7 nearest neighbors (i.e.) ), marked as , and The geodetic distances are respectively Morphological features of these 7 neighboring vertices (based on cortical thickness) (For example) Perform a weighted summation to obtain the original paired vertices. Smoothed baseline morphological indicators :

[0038]

[0039] Among them, weight We employ inverse spatial geodesic distance weighting to assign higher confidence scores to spatially neighboring nodes:

[0040]

[0041] Phase Two: Construction of Lateralized Morphological Similarity Networks and Individualized Atlases

[0042] Reference Figure 3 Then, proceed with steps S4 to S7 to construct the LI-MIND association matrix and build individualized maps.

[0043] S4: Vertex Lateralization Feature Calculation

[0044] First, for indices with negative representations (mean curvature and sulcus depth), sign correction is performed by adding a positive value to the original indices to ensure that all vertex indices are positive. Then, for each given cortical vertex, its original morphological indices (denoted as...) are combined... The weighted smoothing index of the contralateral hemisphere homology matching vertices calculated in S3 (denoted as ) and the index of the contralateral hemisphere homology matching vertices calculated in S3. ), calculate the LI of the given cortical vertex. The calculation formula is as follows:

[0045]

[0046] when When the lateralization lateralization feature is lateralized, it indicates that the morphological feature is biased towards the hemisphere containing the given cortical vertex; conversely, it is biased towards the contralateral hemisphere. The final output is a high-resolution, multidimensional morphological lateralization index brain map at the whole-brain vertex level, providing individualized feature representations for subsequent matrix construction.

[0047] S5: Random vertex sampling and neighborhood extraction based on topological neighborhood

[0048] This step aims to construct a LI-MIND matrix based on random sampling.

[0049] First, obtain the individual left hemisphere cortical mesh model M after surface reconstruction processing. lh = (V lh , F lh ), where V lh Let F represent the set of vertices. lh This represents the set of topological relationships of triangular facets.

[0050] Secondly, the neighborhood of random sampling in one hemisphere is determined. The first step is to preset the number of sampling vertices to N=1000 on the surface of the left hemisphere cortex. The second step is uniform cortical sampling. To ensure the spatial representativeness of the sampling points in the complex cortical folds, a random sampling algorithm based on Euclidean distance constraints is adopted. From V lh The first vertex p1 is randomly selected from the remaining vertices, and then each new sampling point p1 is selected from the remaining vertices. i All must satisfy the condition that the selected sampling point set {p1, ..., p} is consistent with the selected sampling point set {p1, ..., p}. i-1 The shortest surface distance between the points is greater than a preset threshold sigma (set to 10 mm in this embodiment) until N sampling points are selected, thus constructing the left hemisphere sampling point set P. lh = {p i | i=1, 2, ..., N}. The third step is a recursive search of the topological neighborhood (5 layers). For P... lh Each sampling point p in i By traversing the triangular facet index table F lhExtract its topological neighborhood N(p) i The specific steps are as follows:

[0051] ① Define adjacency relations: Define the set of vertices directly connected to vertex v as Adj(v).

[0052] ② Recursive extraction:

[0053] Let R0 = {p i};

[0054] The first-level neighborhood R1 = {v | v∈Adj(R0)} (excluding R0);

[0055] The k-th layer neighborhood R k = {v | v∈Adj(R k-1 (excluding R0, R1, ... R) k-1 (the union of the sets);

[0056] ③ Construct a 5-layer topology set: Iterate to k=5, then the sampling point p i The 5-ring neighborhood vertex set is represented as N(p i ) = .

[0057] Next, symmetric hemispherical mapping and neighborhood localization. Using the cross-hemispherical vertex mapping function f obtained in the previous step... map :V lh→ V rh To achieve mirror positioning of the sampling center and its neighborhood:

[0058] ① Right hemisphere sampling point location: Calculate the sampling point set P corresponding to the right hemisphere. rh ={f map (p i ) | p i ∈P lh}

[0059] ② Right hemisphere neighborhood construction: In the right hemisphere cortical grid M rh Above, with f map (p i Centered on , and following the topological recursive search algorithm described in the neighborhood determination step of random sampling in one hemisphere, the 5-ring neighborhood vertex set N corresponding to the right hemisphere is independently extracted. ),in ∈ P rh .

[0060] S6: LI-MIND Matrix Construction. First, based on the neighborhood obtained in S5 and the lateralization features obtained in S4, extract the multidimensional LI indices of all vertices within any two random sampling points (A, B), forming two multidimensional distributions. and Secondly, quantization is performed using the morphological inverse divergence MIND algorithm. and The similarity reflects the strength of lateralized connectivity between brain regions. In this implementation, divergence is calculated based on Kullback-Leibler divergence (KL Divergence), and the MIND value is defined as the inverse measure of KL divergence:

[0061]

[0062] A larger MIND value indicates a more similar lateralized morphological distribution between the two brain regions, and a stronger morphological covariance connection. To eliminate directional bias, the algorithm performs whole-brain calculations based on both left and right hemisphere partitions, following the above process, and averages the results of the two sets of symmetric matrices to obtain the final dimension. The LI MIND symmetric incidence matrix (in this implementation case) ).

[0063] S7: Spectral Clustering of Brain Regions Based on LI-MIND Matrix

[0064] Using the NxN dimensional whole-brain symmetric LI-MIND association matrix calculated in step 6 as the similarity matrix W, an individual brain network graph model is constructed. First, the degree matrix D is calculated, with its diagonal elements as follows: Based on this, a symmetric normalized Laplace matrix is ​​constructed:

[0065]

[0066] Where I is the identity matrix. For Perform eigenvalue decomposition, extract the eigenvectors corresponding to the k smallest eigenvalues, and construct the feature space matrix. To eliminate the influence of vector magnitude, L2 normalization is performed on each row of U, calculated using the following formula:

[0067]

[0068] Within the normalized feature space, the K-means clustering algorithm is used for brain region segmentation. The number of clusters is iterated within a preset range, and the average silhouette coefficient is calculated. The average silhouette coefficient is calculated as follows:

[0069]

[0070] Furthermore, the value corresponding to the maximum mean of s(i) is selected as the optimal number of clusters. , where a(i) is the average distance within a cluster and b(i) is the average distance between the nearest dissimilar clusters. Labels are assigned to each sampling point based on the clustering results, and the nearest neighbor interpolation algorithm is used to map the labels back to the vertices of the entire cortex. Finally, a majority voting method is used to morphologically smooth the label distribution on the cortical surface, correcting label bias for isolated vertices, ultimately generating a well-defined and anatomically coherent individualized brain region atlas.

[0071] Furthermore, this invention also provides a system (device) for individualized partitioning based on the LI-MIND matrix. Although the structural block diagram of this system is not shown separately in the accompanying drawings, its functional modules correspond one-to-one with the steps described above. The system specifically includes:

[0072] The structural image preprocessing and cortical reconstruction module is used to perform the above step S1, process and obtain the three-dimensional cortical surface mesh generated based on T1w data;

[0073] The interhemispheric nonlinear registration module is used to perform the above step S2 and establish the spatial mapping relationship between the common vertices of the left and right hemispheres.

[0074] The spatial weighted feature estimation module is used to perform the above step S3, and calculate the morphological benchmark index of the opposite homologous vertex using K-nearest neighbor and geodesic distance;

[0075] The lateralization index calculation module is used to perform the above step S4 to generate whole-brain vertex-level lateralization features.

[0076] The random vertex sampling and neighborhood extraction module based on topological neighborhood is used to perform the above step S5 to obtain neighborhood partitioning based on random vertices.

[0077] The inverse divergence matrix construction module is used to perform the above step S6, calculate the similarity of brain region feature distributions through KL divergence and generate the final LI-MIND association matrix;

[0078] The LI-MIND matrix-based spectral clustering brain region segmentation module is used to perform step S7 above to obtain an anatomically coherent individualized brain region segmentation atlas.

[0079] The specific implementation methods of each module can be found in the detailed description of the above steps and embodiments, and will not be repeated here.

Claims

1. A brain parcellation system based on multi-dimensional morphological lateralization inverse divergence similarity network, characterized in that, The method includes: The data acquisition and reconstruction module is used to acquire multimodal medical imaging data of the subjects and reconstruct individual left and right hemisphere cortical surface mesh models; The spatial uniform sampling module is used to perform spatial uniform random sampling on the surface of the left hemisphere cortex to determine a set of sampling points containing $N$ vertices; The topological neighborhood extraction module is used to extract the $k$-ring neighborhood vertex set of each sampling point based on the grid topological connection relationship, and locate the symmetric neighborhood vertex set in the right hemisphere according to the cross-hemispherical vertex correspondence. The lateralization feature calculation module is used to extract multi-dimensional morphological features of vertices within the symmetric neighborhood and calculate the multi-dimensional lateralization index (LI) distribution corresponding to each sampling point. The similarity network construction module is used to calculate the similarity between the neighborhood distributions of any two sampling points using the morphological inverse divergence algorithm, and to construct a multimodal, multidimensional lateralized exponential similarity network (LI-MIND). The partition map generation module is used to perform spectral clustering analysis based on the LI-MIND network, determine the optimal number of clusters, and map the results back to the cortical surface to generate individualized brain maps.

2. The method of claim 1, wherein, When extracting the k-th layer neighborhood, the topological neighborhood extraction module performs the following operations: defines the 0th layer neighborhood R0 centered on the sampling point; and recursively searches the k-th layer neighborhood using the following calculation formula: Where Adj(v) is the set of vertices directly connected to vertex v, and "\" represents the set difference operation; in this embodiment, k is 5.

3. The method of claim 1, wherein, The multi-dimensional morphological features extracted by the lateralization feature calculation module include at least two features from cortical thickness, surface area, mean curvature, curvature, and degree of myelination. The formula for calculating the lateralization index LI is as follows: wherein, and respectively represent morphological characteristic values of the corresponding vertices of the left and right hemispheres.

4. The system of claim 1, wherein, The similarity network (LI-MIND) construction module performs the following operations when calculating similarity: The kernel density estimation method is used to transform the LI value in the neighborhood of each seed point into a probability density distribution function P; The similarity between seed point i and seed point j is calculated using a morphological anti-divergence formula : wherein, is with the Jensen-Shannon divergence or the KL divergence between 5. The system of claim 1, wherein, The spectral clustering partitioning and optimization module includes the following sub-steps when performing clustering analysis: Computing a symmetric normalized laplacian matrix where W is the LI-MIND matrix and D is the degree matrix. right Perform eigenvalue decomposition, extract the eigenvectors corresponding to the k smallest non-zero eigenvalues, and then... Normalization; K-means clustering with silhouette coefficient determining the optimal number of clusters.

6. The system of claim 1, wherein, The system also includes a smoothing post-processing submodule, used for: The clustering labels of seed points are diffused to all vertices of the cortex using the nearest neighbor interpolation algorithm; The majority voting method was used to spatially smooth the label distribution on the cortical surface. By statistically analyzing the frequency of labels appearing in vertices and their topological neighborhoods, isolated vertex labels were corrected to optimize brain region boundaries.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the functions of the system as described in any one of claims 1 to 6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it performs the functions of the system as described in any one of claims 1 to 6.

Images