Thalamic nucleus segmentation method and device based on structure connection and evidence accumulation clustering
Through the multi-process flow fusion and multi-view clustering methods, the problems of poor contrast and single process limitations in thalamic nuclear segmentation are solved, and more accurate and stable thalamic nuclear segmentation is achieved.
Patent Information
- Application Number
- CN202510253383.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-05
- Publication Date
- 2025-07-29
AI Technical Summary
The prior art is difficult to accurately divide the thalamic nuclei in standard MRI images, especially due to poor contrast between the thalamic nuclei and the limitations of a single processing flow, resulting in unstable and inaccurate segmentation results.
By fusing multiple processing processes, including T1-weighted image segmentation, fiber tracking, fiber bundle segmentation and multi-view clustering, a structural connection feature matrix is established, and spectral clustering is combined with the co-correlation matrix to generate high-quality thalamic nuclear segmentation results.
It significantly improves the stability and accuracy of thalamic nuclear segmentation, reduces the impact of systematic errors in fiber tracking and fiber bundle segmentation, and enhances the understanding and recognition of thalamic nuclear structure.
Smart Images
Figure CN120388028A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of medical image processing, and particularly to a thalamic nucleus segmentation method and device based on structural connection and evidence accumulation clustering. Background Art
[0002] The thalamus is a core hub in the brain neural network, integrating and transmitting sensory, motor, and cognitive information through white matter pathways. Understanding the complex structure of the thalamus and accurately dividing thalamic nuclei is crucial because these thalamic nuclei have different functions and pathological significance, directly affecting clinical treatment strategies and neuroscience research. However, in vivo standard structural magnetic resonance imaging (sMRI) data (such as T1-weighted or T2-weighted images), the contrast between thalamic nuclei is poor, making it difficult for traditional atlas-based methods to accurately divide these thalamic nuclei.
[0003] To accurately locate thalamic nuclei, researchers have proposed various valuable methods, which can be classified according to different magnetic resonance imaging (MRI) modalities. sMRI-based methods mainly utilize the high-contrast features provided by new imaging techniques to develop thalamic nucleus segmentation methods. These techniques include susceptibility-weighted imaging (SWI), quantitative susceptibility mapping (QSM), and white matter-suppressed magnetization-prepared rapid acquisition gradient echo imaging (WMn-MPRAGE). Although these methods have achieved remarkable results, their complex imaging protocols, reconstruction algorithms, long acquisition times, and sensitivity to motion limit their application in routine clinical practice. To solve this problem, researchers have proposed synthetic methods to enhance the contrast by synthesizing special modality images from standard T1-weighted images, thereby achieving thalamic nucleus segmentation. However, these methods still rely on the initial contrast of T1w (i.e., T1-weighted) images, making high-quality synthesis challenging. Diffusion magnetic resonance imaging (dMRI) is the most commonly used modality in thalamic nucleus segmentation methods and is widely used in various thalamic nucleus segmentation methods because it is easy to acquire and the local fiber direction features derived from it can effectively distinguish sub-region structures. However, these methods only focus on local information and thus are difficult to accurately divide thalamic nuclei with similar fiber direction distributions but projecting to different brain regions.
[0004] Thalamic nucleus segmentation based on structural connection information driven by dMRI may be a more general and effective strategy, as it not only utilizes the high-contrast features driven by standard imaging modalities but also benefits from the inherent different connection patterns between thalamic nuclei, making it naturally suitable for this task. These methods are usually based on unsupervised learning and rely on a single processing flow to obtain the structural connection feature matrix and generate thalamic nucleus segmentation results. However, due to the inherent assumptions and limitations of current technologies, a single processing flow cannot be universally applicable to all subjects and scenarios, resulting in unreliable structural connection information and limited ability to divide complex thalamic nucleus structures with different connection patterns. This further leads to poor stability of the method, difficulty in capturing consistent thalamic nuclei between different subjects, and inaccurate thalamic nucleus localization. Summary of the Invention
[0005] The embodiments of the present application provide a method and device for thalamic nucleus segmentation based on structural connection and evidence accumulation clustering. By fusing the results of multiple processing flows, a single, robust, and accurate thalamic nucleus segmentation result is obtained. This design can not only effectively reduce the direct impact of systematic errors in intermediate steps such as fiber tracking or fiber bundle segmentation on the final segmentation result but also integrate richer structural connection information, thus better coping with the complexity of thalamic nucleus structures and significantly improving the segmentation quality.
[0006] On the one hand, the embodiments of the present application provide a method for thalamic nucleus segmentation based on structural connection and evidence accumulation clustering, including:
[0007] Dividing the T1-weighted nuclear magnetic resonance image into multiple regions of interest, and combining the regions of interest to obtain a tissue map and a brain region file;
[0008] Performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a probabilistic tracking method to obtain a first fiber tracking result; performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a deterministic fiber tracking method to obtain a second fiber tracking result;
[0009] Using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the first fiber tracking result respectively to obtain four first fiber bundle segmentation results; using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the second fiber tracking result respectively to obtain four second fiber bundle segmentation results;
[0010] Calculating the structural connection features of the first fiber bundle segmentation results and the second fiber bundle segmentation results respectively, and establishing eight structural connection feature matrices;
[0011] Combine eight structural connection feature matrices to obtain multiple view groups;
[0012] Perform multi-view clustering on each view group to obtain corresponding basic clusters;
[0013] Evaluate the number of times a voxel belongs to the same cluster in the basic clusters and establish a co-association matrix;
[0014] Perform spectral clustering on the co-association matrix to obtain the thalamic nucleus segmentation result.
[0015] On the other hand, an embodiment of the present application also provides a thalamic nucleus segmentation device based on structural connection and evidence accumulation clustering, including:
[0016] A preprocessing module for dividing the T1-weighted nuclear magnetic resonance image into multiple regions of interest, and combining the regions of interest to obtain a tissue map and a brain region file;
[0017] A fiber tracking module for performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a probabilistic tracking method to obtain a first fiber tracking result; and performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a deterministic fiber tracking method to obtain a second fiber tracking result;
[0018] A fiber bundle segmentation module for respectively using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the first fiber tracking result to obtain four first fiber bundle segmentation results; and respectively using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the second fiber tracking result to obtain four second fiber bundle segmentation results;
[0019] A feature matrix establishment module for respectively calculating the structural connection features of the first fiber bundle segmentation result and the second fiber bundle segmentation result and establishing eight structural connection feature matrices;
[0020] A view group formation module for combining the eight structural connection feature matrices to obtain multiple view groups;
[0021] A first clustering module for performing multi-view clustering on each view group to obtain corresponding basic clusters; there are two methods for multi-view clustering: anchor-based multi-view K-means clustering and anchor-based multi-view spectral clustering, and these two clustering methods will be randomly selected for the generation of basic clusters;
[0022] A co-association matrix establishment module for evaluating the number of times a voxel belongs to the same cluster in the basic clusters and establishing a co-association matrix;
[0023] The second clustering module is used to perform spectral clustering on the co - association matrix to obtain the thalamic nucleus segmentation result.
[0024] The thalamic nucleus segmentation method and device based on structural connection and evidence accumulation clustering in this application have the following advantages:
[0025] 1. By introducing multiple processing flows, this application not only reduces the impact of systematic errors caused by fiber tracking or fiber bundle segmentation on the final thalamic nucleus segmentation result, but also combines more structural connection information to cope with the complexity of the thalamic nucleus structure.
[0026] 2. This application introduces a view group strategy into the framework to generate a basic clustering by combining multiple structural connection feature matrices. This strategy not only enhances the diversity of the basic clustering, but also can capture the commonalities and complementarities between the structural connection feature matrices, thereby reducing the impact of noise in a single matrix while generating high - quality evidence.
[0027] 3. Experiments on the Human Connectome Project (HCP) Test - retest dataset show that the method of this application is superior to the single - flow - based method in terms of the stability and accuracy of thalamic nucleus segmentation. Brief Description of the Drawings
[0028] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0029] Figure 1 It is the architecture diagram of the thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering provided by the embodiments of this application. Detailed Embodiments
[0030] The following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the drawings in the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, rather than all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.
[0031] Figure 1An architectural diagram of a thalamic nucleus segmentation method based on structural connectivity and evidence accumulation clustering provided in an embodiment of the present application. An embodiment of the present application provides a thalamic nucleus segmentation method based on structural connectivity and evidence accumulation clustering, the overall process of which is: first, a cluster set is generated, and a variety of technical variants are explored in each step of the thalamic nucleus segmentation processing flow based on structural connectivity, and they are combined into multiple processing flows to generate diverse basic clusters. Secondly, evidence is collected, and these basic clusters are integrated into a co-correlation matrix according to the evidence accumulation clustering (EAC) method. Finally, the natural cluster structure is restored, and spectral clustering is applied to the co-correlation matrix to obtain the final thalamic nucleus segmentation result.
[0032] The method of the present application specifically comprises the following steps:
[0033] S100, dividing the T1-weighted MRI image into multiple regions of interest, and combining the regions of interest to obtain a tissue map and a brain region file.
[0034] Illustratively, before dividing the T1-weighted MRI image into regions of interest, the T1-weighted MRI image is preprocessed.
[0035] The present embodiment minimally preprocessed the sMRI data from each subject's two scans using the standard HCP preprocessing process. Specifically, the sMRI preprocessing process included: 1) distortion correction, 2) bias field correction, 3) brain extraction, and 4) registration between the individual T1w image and the MNI152 T1w template image.
[0036] After completing the preprocessing of the T1w images, CINet was used to segment 112 regions of interest (ROIs) for each subject using the T1w images as input. By combining them, the individual's tissue map (including white matter, gray matter, and cerebrospinal fluid), Desikan-Killiany brain area, and subcortical gray matter brain area were obtained. The Desikan-Killiany brain area and subcortical gray matter brain area will constitute the brain area file.
[0037] S200, performing fiber tracking on the high-angle-resolution diffusion magnetic resonance image and the tissue map using a probabilistic tracking method to obtain a first fiber tracking result; performing fiber tracking on the high-angle-resolution diffusion magnetic resonance image and the tissue map using a deterministic fiber tracking method to obtain a second fiber tracking result.
[0038] For example, before using the high-angle-resolution diffusion-weighted MRI image for fiber tracking, the high-angle-resolution diffusion-weighted MRI image is preprocessed.
[0039] In the preprocessing process of this application embodiment, the minimum preprocessing was performed on the dMRI data of each subject's two scans using the standard HCP preprocessing pipeline. Specifically, the preprocessing process for dMRI included: 1) denoising, 2) removing Gibbs artifacts, 3) performing distortion correction with Topup, 4) motion and eddy current correction, 5) bias field correction, and 6) registration between the individual T1w image and the individual high angular resolution diffusion magnetic resonance imaging (HARDI).
[0040] After completing the preprocessing of the high angular resolution diffusion magnetic resonance imaging, the method of performing fiber tracking on the high angular resolution diffusion magnetic resonance imaging and the tissue map using the probabilistic tracking method was mainly executed using the MRtrix3 toolbox. The processing process included:
[0041] Estimating the fiber orientation distribution (FODs) of each voxel using the multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD) diffusion model under the default parameter state;
[0042] Inputting the fiber orientation distribution and the tissue map into the probabilistic tracking algorithm to calculate the second-order integral (iFOD2) on the fiber orientation distribution, and combining the anatomical constrained tractography (ACT) framework to generate multiple whole-brain nerve fibers for each subject. In this application embodiment, the number was 8 million;
[0043] Filtering the whole-brain nerve fibers using the spherical deconvolution-based fiber tractogram filtering (SIFT) algorithm to retain a part of the nerve fibers. In this application embodiment, the number was 2 million.
[0044] Furthermore, the method of performing fiber tracking on the high angular resolution diffusion magnetic resonance imaging and the tissue map using the deterministic fiber tracking method was executed using the white matter analysis (WMA) library. The execution process included:
[0045] Combining the white matter mask in the tissue map with the thalamus mask in the brain region file to construct a unified tracking region mask;
[0046] Performing filtered estimation (UKF) deterministic fiber tracking on all voxels with an anisotropy fraction (FA) greater than the threshold within the tracking region mask. In this application embodiment, the threshold was 0.1, and the Two-tensor version of UKF deterministic fiber tracking was performed on voxels with an FA value greater than 0.1.
[0047] The tracking termination condition was set as: stop tracking when the FA value was lower than 0.08 or the normalized average signal (i.e., the sum of the normalized signals in all gradient directions) was lower than 0.06. To ensure data consistency, this application set 20 seeds within each voxel, so that the number of nerve fibers finally obtained for all subjects was also close to 2 million.
[0048] Further, after obtaining the first fiber tracking result and the second fiber tracking result, a thalamus mask is used to extract the nerve fibers passing through the thalamus from the first fiber tracking result and the second fiber tracking result, the nerve fibers that are flipped within a set range are removed, and the QuickBundles clustering algorithm is used in combination with a set threshold to eliminate some clusters of nerve fibers. In the embodiment of the present application, the nerve fibers that are flipped by 180 degrees within a range of 30 millimeters are removed, and the threshold here is 5.
[0049] S300, for the first fiber tracking result, respectively use the parcellation-based fiber bundle segmentation method, the white matter analysis fiber bundle segmentation method, the deep fiber clustering (DFC) fiber bundle segmentation method, and the hierarchical fiber clustering (FAST) fiber bundle segmentation method to obtain four first fiber bundle segmentation results; for the second fiber tracking result, respectively use the parcellation-based fiber bundle segmentation method, the white matter analysis fiber bundle segmentation method, the deep fiber clustering fiber bundle segmentation method, and the hierarchical fiber clustering fiber bundle segmentation method to obtain four second fiber bundle segmentation results.
[0050] Exemplarily, the white matter analysis method, the deep fiber clustering method, and the hierarchical fiber clustering method all belong to the fiber bundle segmentation methods based on fiber clustering.
[0051] For the parcellation-based fiber bundle segmentation method, the present application operates on the two fiber tracking results respectively: using the brain region file obtained by combining the Desikan-Killiany brain region and the subcortical gray matter brain region to extract the nerve fibers of the two fiber tracking results, and obtaining the nerve fiber bundles connecting each brain region to the thalamus.
[0052] For the fiber bundle segmentation method based on fiber bundles, this application operates on the fiber tracking results of two types respectively. First, the fiber tracking results of each subject are deformed into the unified MNI152 template space by using the deformation file obtained when registering the individual space to the MNI152 template space in the preprocessing. Second, 30,000 nerve fibers passing through the thalamus are randomly selected from each subject and combined into a group of fiber tracking results of nerve fibers passing through the thalamus. Then, the fiber bundle segmentation is completed for the group of fiber tracking results of nerve fibers passing through the thalamus by using three methods, namely WMA, DFC, and FAST, respectively, to obtain their respective group fiber bundle atlases. Finally, the individual fiber bundle identification of each subject is completed according to the group fiber bundle atlases obtained by each method, and three different nerve fiber bundles passing through the thalamus are obtained for each subject. It should be noted that the number of clusters for both the WMA and DFC methods is set to 150, and the maximum distance threshold for FAST is set to 40 mm. These parameters are optimized through range search to ensure good generalization ability among different subjects.
[0053] S400, calculate the structural connection features of the first fiber bundle segmentation result and the second fiber bundle segmentation result respectively, and establish eight structural connection feature matrices.
[0054] Exemplarily, this application calculates the structural connection features of each voxel within the thalamus mask. When constructing the feature vector of each voxel, this application simultaneously considers the existence of crossing fiber bundles and the number of crossing fibers. The relationship between the fiber bundle and the thalamic voxel is described by the structural connection feature matrix M, and its definition is as follows:
[0055]
[0056] Among them, q~p means that if any nerve fiber in fiber bundle p passes through voxel q, then this relationship holds. After the above steps, eight independent structural connection feature matrices can be constructed for eight different fiber bundle results.
[0057] S500, combine the eight structural connection feature matrices to obtain multiple view groups.
[0058] Exemplarily, this application introduces a view group strategy in the framework. By combining multiple structural connection feature matrices, basic clusters are generated. This strategy not only enhances the diversity of the basic clusters but also can capture the commonalities and complementarities between the structural connection feature matrices, thus generating high-quality evidence while reducing the influence of noise in a single matrix. In the framework of this application, multiple structural connection feature matrices are randomly combined into view groups, and each view group contains 1 to 8 matrices. The total number of view groups is flexible. In this application, it is set to 2000.
[0059] S600. Perform multi-view clustering on each view group to obtain corresponding basic clusters.
[0060] Exemplarily, use either the anchor-based multi-view K-means clustering algorithm or the anchor-based multi-view spectral clustering algorithm to cluster each view group.
[0061] Furthermore, to enhance the diversity of the basic clusters and capture multi-scale information, the embodiments of this application set the number of clusters in the range of 4 to 20. For each view group, randomly select a value within this range to generate the basic clusters.
[0062] S700. Evaluate the number of times a voxel belongs to the same cluster in the basic clusters and establish a co-association matrix.
[0063] Exemplarily, the co-association matrix is expressed as:
[0064]
[0065] where C(i,j) represents the value of the voxel (i,j) in the co-association matrix, n ij represents the number of times the voxel (i,j) belongs to the same cluster in the basic clusters, and N represents the total number of basic clusters.
[0066] Through the above evidence accumulation mechanism, this application integrates the results of multiple processing flows, thereby obtaining a more robust and reliable similarity measure of the structural connection patterns between thalamic voxels. This enables better handling of the complexity of thalamic sub-regions and reduces the significant impact of systematic errors introduced by a single processing flow.
[0067] S800. Perform spectral clustering on the co-association matrix to obtain the thalamic nucleus segmentation result.
[0068] The embodiments of this application also include a thalamic nucleus segmentation device based on structural connection and evidence accumulation clustering. The device includes the following modules:
[0069] A preprocessing module for dividing the T1-weighted nuclear magnetic resonance image into multiple regions of interest, combining the regions of interest to obtain a tissue map and a brain region file;
[0070] A fiber tracking module for performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a probabilistic tracking method to obtain a first fiber tracking result; performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a deterministic fiber tracking method to obtain a second fiber tracking result;
[0071] The fiber bundle segmentation module is used to obtain four first fiber bundle segmentation results for the first fiber tracking result respectively by using the partition-based fiber bundle segmentation method, the white matter analysis fiber bundle segmentation method, the deep fiber clustering fiber bundle segmentation method, and the hierarchical fiber clustering fiber bundle segmentation method; and obtain four second fiber bundle segmentation results for the second fiber tracking result respectively by using the partition-based fiber bundle segmentation method, the white matter analysis fiber bundle segmentation method, the deep fiber clustering fiber bundle segmentation method, and the hierarchical fiber clustering fiber bundle segmentation method.
[0072] The feature matrix establishment module is used to calculate the structural connection features of the first fiber bundle segmentation result and the second fiber bundle segmentation result respectively, and establish eight structural connection feature matrices.
[0073] The view group formation module is used to combine the eight structural connection feature matrices to obtain multiple view groups.
[0074] The first clustering module is used to perform multi-view clustering on each view group to obtain the corresponding basic clustering; there are two methods for multi-view clustering: anchor-based multi-view K-means clustering and anchor-based multi-view spectral clustering, and these two clustering methods will be randomly selected for the generation of the basic clustering.
[0075] The co-association matrix establishment module is used to evaluate the number of times a voxel belongs to the same cluster in the basic clustering and establish a co-association matrix.
[0076] The second clustering module is used to perform spectral clustering on the co-association matrix to obtain the thalamic nucleus segmentation result.
[0077] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present application.
[0078] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these changes and modifications.
Claims
1. A thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering, characterized in that Including: Dividing the T1-weighted nuclear magnetic resonance (NMR) image into multiple regions of interest (ROIs), and combining the ROIs to obtain a tissue map and a brain region file; Performing fiber tracking on the high angular resolution diffusion NMR image and the tissue map using a probabilistic tracking method to obtain a first fiber tracking result; Performing fiber tracking on the high angular resolution diffusion NMR image and the tissue map using a deterministic fiber tracking method to obtain a second fiber tracking result; Using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the first fiber tracking result respectively to obtain four first fiber bundle segmentation results; using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the second fiber tracking result respectively to obtain four second fiber bundle segmentation results; Calculating the structural connection features of the first fiber bundle segmentation result and the second fiber bundle segmentation result respectively, and establishing eight structural connection feature matrices; Combining the eight structural connection feature matrices to obtain multiple view groups; Performing multi-view clustering on each of the view groups to obtain corresponding basic clusters; Evaluating the number of times a voxel belongs to the same cluster in the basic clusters to establish a co-association matrix; Performing spectral clustering on the co-association matrix to obtain a thalamic nucleus segmentation result.
2. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 1, characterized in that Before dividing the T1-weighted NMR image into the ROIs, preprocessing the T1-weighted NMR image first; before using the high angular resolution diffusion NMR image for fiber tracking, preprocessing the high angular resolution diffusion NMR image first.
3. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 2, characterized in that The preprocessing performed on the T1-weighted NMR image includes: Successively performing distortion correction, bias field correction, brain extraction, and image registration on the T1-weighted NMR image in the structural NMR image; The preprocessing performed on the high angular resolution diffusion NMR image includes: Successively performing denoising, Gibbs artifact removal, distortion correction, motion and eddy current correction, bias field correction, and image registration on the high angular resolution diffusion NMR image.
4. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 1, characterized in that The method for performing fiber tracking on the high angular resolution diffusion NMR image and the tissue map using a probabilistic tracking method includes: Estimating the fiber orientation distribution of each voxel using a multi-shell multi-tissue constrained spherical deconvolution diffusion model; Inputting the fiber orientation distribution and the tissue map into the probabilistic tracking algorithm iFOD2, and combining an anatomical constraint fiber bundle tracking framework to generate whole-brain nerve fibers for each subject; Filtering the whole-brain nerve fibers using a fiber bundle map filtering algorithm based on spherical deconvolution to retain a part of the nerve fibers.
5. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 1, characterized in that The method for performing fiber tracking on the high angular resolution diffusion NMR image and the tissue map using a deterministic fiber tracking method includes: Combining the white matter mask in the tissue map with the thalamus mask in the brain region file to construct a unified tracking region mask; Perform UKF deterministic fiber tracking on all voxels with anisotropic fractions greater than the threshold within the range of the tracking region mask to generate whole-brain nerve fibers for each subject.
6. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 1, characterized in that After obtaining the first fiber tracking result and the second fiber tracking result, use the thalamus mask to extract the nerve fibers passing through the thalamus from the first fiber tracking result and the second fiber tracking result, remove the nerve fibers that are flipped within the set range, and use the QuickBundles fiber clustering algorithm combined with the set threshold to eliminate some clusters of nerve fibers.
7. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 1, wherein Perform multi-view clustering on each of the view groups using either the anchor-based multi-view K-means clustering algorithm or the anchor-based multi-view spectral clustering algorithm.
8. The thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to claim 1, characterized in that The co-association matrix is expressed as: Among them, C(i,j) represents the value of the voxel (i,j) in the co-correlation matrix, and n ij represents the number of times the voxel (i,j) belongs to the same cluster in the basic clustering, and N represents the total number of the basic clustering.
9. An apparatus for applying the thalamic nucleus segmentation method based on structural connection and evidence accumulation clustering according to any one of claims 1-8, characterized in that, Including: A preprocessing module for dividing the T1-weighted nuclear magnetic resonance image into multiple regions of interest, and combining the regions of interest to obtain a tissue map and a brain region file; A fiber tracking module for performing fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a probabilistic tracking method to obtain a first fiber tracking result; Perform fiber tracking on the high angular resolution diffusion nuclear magnetic resonance image and the tissue map using a deterministic fiber tracking method to obtain a second fiber tracking result; A fiber bundle segmentation module for respectively using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the first fiber tracking result to obtain four first fiber bundle segmentation results; and respectively using a partition-based fiber bundle segmentation method, a white matter analysis fiber bundle segmentation method, a deep fiber clustering fiber bundle segmentation method, and a hierarchical fiber clustering fiber bundle segmentation method on the second fiber tracking result to obtain four second fiber bundle segmentation results; A feature matrix establishment module for respectively calculating the structural connection features of the first fiber bundle segmentation result and the second fiber bundle segmentation result, and establishing eight structural connection feature matrices; A view group formation module for combining the eight structural connection feature matrices to obtain multiple view groups; A first clustering module for performing multi-view clustering on each of the view groups to obtain corresponding basic clusters; the methods of multi-view clustering include two types: anchor-based multi-view K-means clustering and anchor-based multi-view spectral clustering, and one of these two clustering methods will be randomly selected for the generation of the basic clusters; A co-association matrix establishment module for evaluating the number of times a voxel belongs to the same cluster in the basic clusters and establishing a co-association matrix; A second clustering module for performing spectral clustering on the co-association matrix to obtain a thalamus nucleus segmentation result.