Human brain superficial layer white matter fiber structure map construction method and system
The method addresses the challenge of accurately identifying and analyzing SWM fibers by using unsupervised clustering and deep learning to construct a detailed SWM graph, enhancing early neurodegenerative disease diagnosis.
Patent Information
- Application Number
- CN202510394490.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-31
- Publication Date
- 2025-07-15
AI Technical Summary
The prior art is difficult to extract and analyze white matter fibers in the superficial surface of the human brain with high precision, and lacks a standardized map construction method, which limits its application value in the early diagnosis of neurodegenerative diseases.
By combining geometric feature thresholds and deep learning classifiers based on whole-brain diffusion magnetic resonance imaging data, superficial white matter fibers are identified from whole-brain functional fiber clusters, and layered clustering is carried out to construct a structural map of superficial white matter fibers in the whole-brain superficial white matter fibers and analyze its functional properties.
High-precision extraction and analysis of superficial white matter fibers was achieved, and its structural and functional correlation model was constructed, providing a reliable data basis for the early diagnosis of neurodegenerative diseases.
Smart Images
Figure CN120318541A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of whole-brain image processing, and particularly relates to a method and system for constructing a human brain superficial white matter fiber structure atlas. Background Art
[0002] Superficial White Matter (SWM), as a key connection structure between the cerebral cortex and deep white matter, in which U-shaped fibers play important roles in cognitive functions, neural plasticity, and disease processes. In recent years, the progress of diffusion magnetic resonance imaging (dMRI) technology has provided a basis for white matter fiber tracking. However, existing methods mainly focus on the analysis of long-distance deep white matter bundles, and there are still gaps in the technologies for identifying, extracting, and analyzing superficial white matter. In addition, due to the characteristics of superficial white matter fibers, such as short length, high curvature (showing U-shaped bending), spatial distribution close to the cortex, and connection of adjacent brain regions, traditional fiber tracking algorithms (such as deterministic tracking or screening based on fractional anisotropy FA) are prone to missed detection or misjudgment of superficial white matter fibers due to partial volume effects, noise interference, and insufficient parameter adaptability. There are still significant technical bottlenecks in the accurate extraction of superficial white matter fibers. In addition, there is no widely recognized method for constructing a standardized superficial white matter atlas in the existing technology, it is difficult to quantify its spatial distribution law, and there are few studies systematically associating the structural characteristics and functional attributes of superficial white matter, which limits its application value in the early diagnosis of neurodegenerative diseases. Therefore, it is urgent to develop a high-precision extraction technology for superficial white matter characteristics and integrate multi-modal data to construct its structural and functional association model to fill the technical gaps in this field. Summary of the Invention
[0003] To overcome the challenges of the existing technology, the present invention provides a method and system for constructing a human brain superficial white matter fiber structure atlas, which extracts superficial white matter fibers from the whole brain, constructs a whole-brain superficial white matter fiber atlas, and analyzes the functional attributes of superficial white matter fibers, providing a high-quality data acquisition channel for research based on superficial white matter.
[0004] The technical solution to achieve the object of the present invention is as follows:
[0005] A method for constructing a human brain superficial white matter fiber structure atlas, comprising:
[0006] Step 1, based on whole-brain diffusion magnetic resonance imaging data, unsupervised generation of a finely divided whole-brain functional fiber cluster atlas;
[0007] Step 2, combining geometric feature thresholds and a deep learning classifier to identify superficial white matter fibers from the whole-brain functional fiber clusters and perform hierarchical clustering to generate a superficial white matter fiber sub-atlas with the optimal number of clusters;
[0008] Step 3: Analyze the functional properties of the superficial white matter fibers by the trajectory-weighted dynamic functional connectivity analysis method to obtain the structural diagram of the superficial white matter fibers of the human brain.
[0009] Further, step 1 specifically includes:
[0010] Step 1.1: Obtain the whole-brain diffusion magnetic resonance imaging data and perform preprocessing.
[0011] Step 1.2: Based on the preprocessed data, reconstruct the three-dimensional spatial distribution of the whole-brain white matter fibers by the fiber tracking method.
[0012] Step 1.3: Based on step 1.2, merge the whole-brain white matter fibers of multiple samples into a population whole-brain fiber set, and filter out the unreasonable fibers.
[0013] Step 1.4: Perform two consecutive clusterings on the population fibers to divide the whole-brain fiber atlas into 33,256 fiber clusters to obtain the population whole-brain fiber atlas.
[0014] Step 1.5: Map the fiber clusters obtained in step 1.4 to individuals, find the corresponding fiber clusters in each individual, and filter out the unreasonable fibers to obtain the whole-brain fiber atlas of the individual.
[0015] Further, the reconstruction of the three-dimensional spatial distribution of the whole-brain white matter fibers by the fiber tracking method in step 1.2 specifically includes:
[0016] Step 1.2.1: Use the probabilistic fiber tracking method to perform whole-brain fiber tracking on all individuals based on the first-order integral method of the fiber orientation distribution, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm.
[0017] Step 1.2.2: Use the probabilistic fiber tracking method to perform whole-brain fiber tracking on all individuals based on the second-order integral method of the fiber orientation distribution, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm.
[0018] Step 1.2.3: Use the deterministic fiber tracking method to perform whole-brain fiber tracking on all individuals based on the streamline tractography of spherical deconvolution, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm.
[0019] Step 1.2.4: Merge the fibers obtained in steps 1.2.1 - 1.2.3 as the whole-brain fibers of the individual.
[0020] Further, the two consecutive clusterings of the population fibers in step 1.4 specifically include:
[0021] First, the whole-brain fibers are cortically divided according to the Yeo 2011 functional brain network parcellation atlas, and the fibers are classified into different categories according to the different cortical functional networks they connect;
[0022] The first clustering is performed on different fiber categories, and the number of categories is where n is the total number of fibers in this fiber classification;
[0023] The second clustering is performed on each fiber subclass obtained after the first clustering, and the number of categories is where m is the total number of fibers in this fiber subclass.
[0024] Furthermore, step 2 specifically includes:
[0025] Step 2.1, according to a preset geometric feature threshold, candidate fiber clusters in the whole-brain fiber atlas are screened, and the probability that this fiber cluster belongs to the superficial white matter is calculated based on geometric features;
[0026] Step 2.2, a deep learning model classifier is established to determine whether a fiber cluster belongs to the superficial white matter, and the probability that this fiber cluster belongs to the superficial white matter is calculated based on the deep learning method;
[0027] Step 2.3, a joint analysis is performed on the two probabilities of belonging to the superficial white matter obtained in step 2.1 and step 2.2, and the fiber clusters with probabilities of belonging to the superficial white matter greater than the probability threshold under both calculation methods are retained;
[0028] Step 2.4, based on the fiber clusters retained in step 2.3, combined with the mask of the functional brain network cortical segmentation atlas, the fibers that only connect adjacent brain regions are retained;
[0029] Step 2.5, a secondary hierarchical clustering is performed on the superficial white matter fibers obtained in step 2.4 to adaptively generate a sub-atlas of superficial white matter fibers with the optimal number of clusters.
[0030] Furthermore, the geometric features in step 2.1 specifically include:
[0031] Step 2.1.1, the bending degree threshold: the ratio of the middle segment distance of the fiber to the total length is less than 2;
[0032] Step 2.1.2, the fiber length range: the ratio of the middle segment distance of the fiber to the end point distance is greater than 1.1;
[0033] Step 2.1.3, the end point spatial hemisphere position constraint: the ratio of the total length of the fiber to the end point distance is greater than 1.1, and the total length range of the fiber is between 15 mm and 85 mm.
[0034] Further, the specific steps for calculating the probability that the fiber clustering belongs to the superficial white matter based on the deep learning method in step 2.2 are as follows: Using the fibers in the ORG whole-brain fiber atlas as the training data input and the deep / superficial white matter labels in the ORG atlas as the ground truth, a fully connected network is established to predict the category of the fiber clustering. The fully connected network is a three-layer fully connected neural network. The input is a 4×3-dimensional fiber feature vector matrix. After passing through two hidden layers activated by ReLU, the final output is a 2-dimensional classification result.
[0035] Further, step 3 specifically includes:
[0036] Step 3.1, using the trajectory-weighted dynamic functional connectivity calculation method, based on the individual's superficial white matter fiber image and functional magnetic resonance image, generate the whole-brain TW-dFC image of the sample;
[0037] Step 3.2, perform independent component analysis on the whole-brain TW-dFC images of all samples to obtain the whole-brain voxel-level segmentation of the superficial white matter at the group level;
[0038] Step 3.3, calculate the number of different independent components passed through by each group's superficial white matter clustering, and calculate the uncertainty of the superficial white matter clustering according to the degree of consistency of the independent components passed through by the fibers in the clustering;
[0039] Step 3.4, calculate the degree of overlap between different independent components and the functional brain networks connected by white matter fibers, and reveal the correlation between different superficial white matter components and brain functions.
[0040] Further, the calculation formula for the uncertainty is:
[0041]
[0042] Where, is the uncertainty measure of the superficial white matter clustering kk, represents the probability that the points within the superficial white matter clustering k belong to the same M independent components, N points represents the total number of points in the superficial white matter clustering k, N(M) same IC is the number of fiber points belonging to the same M independent components.
[0043] A system for constructing a human brain superficial white matter fiber structure atlas, including:
[0044] An unsupervised partitioning unit, which unsupervised generates a fine-partitioned whole-brain functional fiber cluster atlas based on the whole-brain diffusion magnetic resonance imaging data;
[0045] A hierarchical clustering unit combines a geometric feature threshold with a deep learning classifier to identify superficial white matter fibers from the whole-brain functional fiber clusters and perform hierarchical clustering, generating a sub-atlas of superficial white matter fibers with the optimal number of clusters.
[0046] An analysis unit analyzes the functional properties of the superficial white matter fibers through a trajectory-weighted dynamic functional connectivity analysis method to obtain a structural diagram of the superficial white matter fibers in the human brain.
[0047] Compared with the prior art, the beneficial effects of the present invention are as follows: Based on the fiber tracking algorithm, the three-dimensional spatial distribution of the whole-brain white matter fibers is reconstructed to ensure that all potentially real white matter fibers are traced; by calculating the number of different independent components passed through by each population of superficial white matter clusters and calculating the uncertainty measure of the superficial white matter cluster according to the degree of consistency of the independent components passed through by the fibers in the cluster; calculating the overlap degree between different independent components and the functional brain networks connected by the white matter fibers to reveal the correlation between different superficial white matter components and brain functions, and obtaining the functional properties of different superficial white matter fibers; the present invention fills the gap in high-precision data resources in the field by constructing a multi-level superficial white matter atlas, providing a reliable research basis for deeply analyzing the correlation mechanism between the structure and function of the superficial white matter. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a schematic diagram of a method for generating a finely divided whole-brain fiber atlas proposed by the present invention.
[0049] Figure 2 It is a diagram of the whole-brain superficial white matter fiber atlas obtained by the present invention.
[0050] Figure 3 It is a diagram of the hierarchical clustering result of the whole-brain superficial white matter fibers of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0051] To enable those skilled in the art to better understand the technical solutions of the present application, the following further describes in detail the construction process of the overall superficial white matter fiber atlas provided by the present application in conjunction with the drawings and specific embodiments.
[0052] Combined with Figure 1 , a method for constructing a structural atlas of superficial white matter fibers in the human brain includes the following steps:
[0053] Step 1: Unsupervised generation of a finely divided whole-brain functional fiber cluster atlas, including merging population fibers based on a multi-modal fiber tracking algorithm and dividing them into 33,256 fiber subclasses through two rounds of clustering.
[0054] Step 2: Combine a geometric feature threshold with a deep learning classifier to identify superficial white matter fibers from the whole-brain functional fiber clusters and perform hierarchical clustering.
[0055] Step 3, analyze the functional properties of superficial white matter fibers through the trajectory-weighted dynamic functional connectivity (TW-dFC) analysis method.
[0056] The process of Step 1 is as follows:
[0057] Step 1.1, obtain whole-brain diffusion magnetic resonance imaging (dMRI) data and perform preprocessing to eliminate noise and motion artifacts;
[0058] Step 1.2, based on the fiber tracking algorithm, reconstruct the three-dimensional spatial distribution of whole-brain white matter fibers;
[0059] Use the probabilistic fiber tracking method based on the first-order integration method of fiber orientation distribution (iFOD1, First-order Integration over Fiber Orientation Distributions) to perform whole-brain fiber tracking on all individuals, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm;
[0060] Use the probabilistic fiber tracking method based on the second-order integration method of fiber orientation distribution (iFOD2, Second-order Integration over Fiber Orientation Distributions) to perform whole-brain fiber tracking on all individuals, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm;
[0061] The deterministic fiber tracking method based on spherical deconvolution streamline tractography (DET) performs whole-brain fiber tracking on all individuals, generating 1,000,000 fibers respectively, with a minimum length of 20 mm and a maximum length of 250 mm;
[0062] Merge the fibers obtained by the above three methods as the whole-brain fibers of the individual to ensure that all white matter fibers that may actually exist are traced. When using the above three fiber tracking methods, the Anatomically-Constrained Tractography (ACT) algorithm is used to ensure that the tracking conforms to anatomical constraints.
[0063] Step 1.3, merge the whole-brain white matter fibers of multiple samples to form a collective whole-brain fiber set of the population, and screen out the unreasonable fibers in the classification atlas. The calculation method for screening out the unreasonable fibers in the classification atlas is as follows:
[0064] First, encode all single fibers, and the calculation method is as follows:
[0065] For a single fiber, the number of its points is n, which can be represented by an ordered sequence of n points from P1 to P n where P i =(x i , y i , z i ). Let d i+1 represent the distance between point P i and point P i+1 . The proportion of a point in this fiber can be expressed as t i .
[0066]
[0067] By flipping and splicing the point sequence representing the fiber and the sequence of point proportions, a matrix of size (2n - 1)×3 and an array of length (2n - 1) can be obtained. From this, a matrix Y of size (2n - 1)×4 can be calculated, where Thus, the final coefficient matrix β, with a shape of (4×3), can be calculated as follows.
[0068]
[0069] The first row of β can represent the position information of the fiber, and directly use the first row as the first three dimensions among the nine dimensions of the coding result;
[0070] For the coding result of the fiber, calculate the mean and standard deviation of the data:
[0071]
[0072] where μ represents the mean value and v represents the variance. Determine the selected values according to the threshold. For example, if threshold = 1.5, the retention interval is [μ - 1.5σ, μ + 1.5σ].
[0073] Step 1.4, perform two clustering operations on the population fibers in sequence, divide the whole-brain fiber atlas into 33,256 smaller fiber clusters, and obtain the whole-brain white matter fiber atlas of the population. The clustering process is as follows:
[0074] First, perform cortical division on the whole-brain fibers according to the Yeo2011 functional brain network partition atlas, and divide the fibers into different categories according to the different cortical functional networks they connect;
[0075] The first clustering is to cluster different fiber categories, and the number of categories is where n is the total number of fibers in this fiber classification;
[0076] The second clustering performs clustering on each fiber sub-category obtained after the first clustering, and the number of categories is where m is the total number of fibers in the fiber sub-category;
[0077] Step 1.5, map the fiber clustering obtained in Step 1.4 to individuals, find the corresponding fiber clustering in each individual, and screen out the unreasonable fibers in the classification atlas to obtain the whole-brain white matter fiber atlas of the individual.
[0078] The process of identifying and hierarchically clustering the superficial white matter fibers from the whole-brain fiber atlas in Step 2 is as follows:
[0079] Step 2.1, according to the preset geometric feature thresholds, screen the candidate fiber clusters in the whole-brain fiber atlas, and calculate the probability that the fiber cluster belongs to the superficial white matter based on geometric features, where the geometric features include the bending degree threshold range, fiber length range, and endpoint spatial hemisphere position constraint; the geometric feature calculation method is as follows:
[0080] a. Set the bending degree threshold: the ratio of the middle segment distance of the fiber (the distance from the fiber starting point to the midpoint plus the distance from the midpoint to the end point) to the total length (the actual total length of the fiber) is less than 2;
[0081] b. The ratio of the middle segment distance of the fiber to the endpoint distance is greater than 1.1: This ratio indicates that the fiber is not a strictly straight line but has a certain degree of bending, which conforms to the morphological characteristics of the superficial white matter fibers;
[0082] c. The ratio of the total length of the fiber to the endpoint distance is greater than 1.1: This condition further ensures that the fiber is not a completely straight line but has a certain degree of bending, excluding overly regular long-distance projection fibers; the total length of the fiber is between 15 and 85: This excludes overly short or long fibers and ensures that the selected fibers are within the anatomical length range of the superficial white matter fibers.
[0083] Step 2.2, establish a deep learning model classifier to automatically judge whether a fiber cluster belongs to the superficial white matter, and calculate the probability that the fiber cluster belongs to the superficial white matter based on the deep learning method. The calculation method is as follows:
[0084] Use the fibers in the ORG whole-brain fiber atlas as the training data input, and the deep / superficial white matter labels in the ORG atlas as the data truth values, and establish a fully connected network to predict the category (deep / superficial white matter) of the fiber cluster. The network is a three-layer fully connected neural network. The input is a 4×3-dimensional fiber feature vector matrix (flattened into a 12-dimensional vector). After passing through two hidden layers (128 and 64 units) activated by ReLU, the final output is a 2-dimensional classification result. It is used for binary classification of fiber bundles, extracting features and performing classification prediction through linear transformation and non-linear activation.
[0085] Step 2.3, perform a joint analysis based on the probabilities of two types of fibers belonging to the superficial white matter calculated according to the fiber clustering based on geometric features and deep learning methods, and retain the fiber clusters that have a high probability of belonging to the superficial white matter under both calculation methods. The joint analysis of the geometric feature threshold and the deep learning model improves the consistency of SWM fiber identification in individuals to more than 98%. The calculation method is as follows:
[0086] Multiply the probabilities of the fibers obtained by the two methods belonging to the superficial white matter as the joint probability of the fibers belonging to the superficial white matter. Calculate the zscore for the three probabilities (based on geometric features, based on deep learning methods, joint probability), and retain the fiber clusters with a zscore greater than 1.5 under each probability as the superficial white matter clusters.
[0087] Step 2.4, in combination with the functional brain network cortical segmentation atlas mask, retain the fibers that only connect adjacent brain regions, such as Figure 2 all the finally selected superficial white matter fibers shown.
[0088] Step 2.5, perform secondary hierarchical clustering on the superficial white matter fibers that meet the conditions to adaptively generate the optimal number of clusters of superficial white matter fiber sub-atlases. The clustering process is as follows:
[0089] Select different distance thresholds to perform multiple hierarchical clusterings on the superficial white matter fibers connecting the same functional brain network in the whole brain that are selected, merge the sub-clusters with similar distances, and calculate the Davies-Bouldin index of the finally obtained clustering result (an index used to evaluate the clustering quality, calculating the ratio of the within-cluster compactness to the between-cluster separation, and the smaller the value, the better the clustering effect).
[0090] Adaptively select the distance threshold corresponding to the smallest Davies-Bouldin index, and use the clustering result under the corresponding threshold as the final clustering result of the superficial white matter fibers, and finally obtain 680 superficial white matter classes. Figure 3 Shows the distribution of the number of fibers connecting different functional brain networks under three different hierarchical clustering schemes.
[0091] The analysis process of the functional properties of the superficial white matter fibers in Step 3 is as follows:
[0092] Step 3.1, use a trajectory-weighted dynamic functional connectivity calculation method (TW-dFC) to generate the TW-dFC image of the sample based on the individual's superficial white matter fiber image and functional magnetic resonance (fMRI) image;
[0093] The TW-dFC images were generated using the tckdfc command in the MRtrix3 toolkit. The dynamic functional connectivity (dFC) was calculated using a Hamming window (55 time points), that is, the changes in fMRI signals were evaluated in the time dimension through a sliding window method; the spatial accuracy of the trajectories was improved by backtracking to prevent misjudgment of fiber termination points; the voxel resolution was 2 mm.
[0094] Step 3.2, perform independent component analysis on the whole-brain TW-dFC images of all samples to obtain a whole-brain voxel-level segmentation of the superficial white matter at the group level, where the voxels in the same component show stronger correlations in function and structure.
[0095] The CanICA method in the nilearn library was used for independent component analysis. Compared with traditional ICA, CanICA combines PCA for pre-dimension reduction to improve stability; it is suitable for multi-subject data analysis and automatically aggregates individual data to identify neurally active patterns that are independent in time and space.
[0096] Step 3.3, combine the group brain voxel-level segmentation results obtained by ICA, calculate the number of different independent components passed through by each cluster of superficial white matter in the group, and calculate the uncertainty of the superficial white matter cluster according to the degree of consistency of the independent components passed through by the fibers in the cluster.
[0097] The fibers in the superficial white matter fiber clusters with high certainty should be more unified in both structure and function, so the independent components of the superficial white matter passed through should also have high consistency. The fiber streamlines were resampled to 30 points, and class labels were assigned according to the independent components to which the voxels belonged. The uncertainty of the superficial white matter cluster was evaluated by the probability that the streamline points were assigned to the same or two most likely components. If the points of a cluster were distributed in multiple functional partitions, it indicated an increase in the uncertainty of the cluster. The specific calculation method of uncertainty is as follows:
[0098]
[0099] where is the uncertainty measure of the superficial white matter cluster k, represents the probability that the points within the superficial white matter cluster k belong to the same M independent components (IC). N points represents the total number of points in the superficial white matter cluster k, and N(M) Same IC is the number of fiber points belonging to the same M independent components. When M = 1 or M = 2, they correspond to two different schemes respectively.
[0100] Step 3.4, calculate the overlap degree between different independent components and the functional brain networks connected by white matter fibers, and reveal the correlation between different superficial white matter components and brain functions. The calculation method is as follows:
[0101] Using the independent component information of the voxels passed by these fibers, analyze the overlap degree between the functional brain regions and their corresponding independent components. By calculating the probability that the fibers of the functional network pass through different independent components, evaluate the overlap degree between the functional brain network and the independent components. A larger overlap degree indicates that there may be a stronger correlation between the functional network and the corresponding independent component.
[0102] The present invention also provides a system for constructing a human brain superficial white matter fiber structure map, including:
[0103] An unsupervised division unit, which unsupervised generates a fine-divided whole-brain functional fiber cluster map based on the whole-brain diffusion magnetic resonance imaging data;
[0104] A hierarchical clustering unit, which combines geometric feature thresholds and a deep learning classifier to identify superficial white matter fibers from the whole-brain functional fiber clusters and perform hierarchical clustering to generate a superficial white matter fiber sub-map with the optimal number of clusters;
[0105] An analysis unit, which analyzes the functional attributes of the superficial white matter fibers through a trajectory-weighted dynamic functional connection analysis method to obtain a human brain superficial white matter fiber structure diagram.
[0106] Embodiment
[0107] First is the input part. The input of the present invention includes the folder path where all samples are located, the T1w data path of the whole brain of the sample, the dMRI scan image of the whole brain of the sample, the bval path of the whole brain of the sample, the bvec path of the whole brain of the sample, and the output folder path of the file.
[0108] Next is the whole-brain fiber generation and clustering of the group. The T1w structural images are used for tissue segmentation with FreeSurfer to generate 5TT images. The fiber orientation distribution function (fODF) is calculated using the dMRI images that have been processed by the HCP preprocessing pipeline. Then, fiber tracking is performed using fODF, and then the fibers of all samples are merged to obtain the whole-brain fibers of the group, and the first removal of unreasonable fibers is carried out. Based on the Yeo2011 functional brain network cortical segmentation atlas, the fibers are classified according to the different functional brain networks they connect. Then, two consecutive clusterings are performed on the group fibers in different classifications respectively, and finally 33,256 smaller group fiber clusters are obtained. Through the mapping relationship between the group fibers and the individual fibers, the fibers corresponding to each cluster in the individual can be found from the individual, and the fiber clustering results of the individual are obtained, and the second removal of unreasonable fibers is carried out. At this time, the whole-brain white matter fiber atlas of the group and the corresponding fibers of each fiber cluster in each sample are obtained.
[0109] Then, the superficial white matter fibers (SWM) are identified from the whole-brain white matter fiber atlas and merged by hierarchical clustering. According to the geometric feature threshold and the trained shallow / deep white matter fiber classifier respectively, the candidate fiber clusters in the whole-brain fiber atlas are screened, and the probability that the fiber cluster belongs to the superficial white matter based on geometric features or deep learning results is calculated. Next, a joint analysis is carried out according to the two probabilities of belonging to the superficial white matter calculated by the fiber cluster based on geometric features and deep learning methods, and the fiber clusters with a high probability of belonging to the superficial white matter under both calculation methods are retained. Finally, combined with the functional brain network cortical segmentation atlas mask, only the fibers connecting adjacent brain regions are retained as the finally identified superficial white matter fiber clusters. Then, a secondary hierarchical clustering is carried out on the basis of all the obtained superficial white matter clusters to adaptively generate the optimal superficial white matter fiber sub-atlas with the number of clusters.
[0110] Finally, the functional properties of the identified superficial white matter fibers are analyzed. First, the trajectory-weighted dynamic functional connectivity (TW-dFC) images of the brain voxels corresponding to the superficial white matter fibers of each individual are calculated, and then the group independent component analysis is performed on the whole-brain TW-dFC images of all samples to obtain the whole-brain voxel-level segmentation of the superficial white matter at the group level. Then, by calculating the number of different independent components passed through by each group superficial white matter cluster and calculating the uncertainty measure of the superficial white matter cluster according to the consistency degree of the independent components passed through by the fibers in the cluster; in addition, calculate the overlap degree between different independent components and the functional brain networks connected by the white matter fibers to reveal the correlation between different superficial white matter components and brain functions, and obtain the functional properties of different superficial white matter fibers.
[0111] The above process is the entire process of the method proposed by the present invention. Those skilled in the art to which the present invention pertains can select appropriate parameters and adjust each step therein according to needs to meet the requirements, but will not deviate from the general idea of the present invention or exceed the scope defined by the appended claims. In addition, although some specific terms are used in this specification, these terms are only for the convenience of description and do not constitute any limitation to the present invention.
Claims
1. A method for constructing a human brain superficial white matter fiber structure atlas, characterized in that Including: Step 1: Based on whole-brain diffusion magnetic resonance imaging data, unsupervised generation of a finely divided whole-brain functional fiber cluster atlas; Step 2: Combining geometric feature thresholds with a deep learning classifier, identifying superficial white matter fibers from the whole-brain functional fiber clusters and performing hierarchical clustering to generate a sub-atlas of superficial white matter fibers with the optimal number of clusters; Step 3: Through a trajectory-weighted dynamic functional connectivity analysis method, analyzing the functional properties of superficial white matter fibers to obtain a structural diagram of human brain superficial white matter fibers.
2. The method for constructing a human brain superficial white matter fiber structure atlas according to claim 1, wherein, The specific steps of Step 1 include: Step 1.1: Obtaining whole-brain diffusion magnetic resonance imaging data and performing preprocessing; Step 1.2: Based on the preprocessed data, reconstructing the three-dimensional spatial distribution of whole-brain white matter fibers through a fiber tracking method; Step 1.3: Based on Step 1.2, merging the whole-brain white matter fibers of multiple samples into a group's whole-brain fiber collection and screening out unreasonable fibers; Step 1.4: Performing two consecutive clusterings on the group fibers, dividing the whole-brain fiber atlas into 33,256 fiber clusters to obtain the group's whole-brain fiber atlas; Step 1.5: Mapping the fiber clusters obtained in Step 1.4 to individuals, finding the corresponding fiber clusters in each individual, and screening out unreasonable fibers to obtain the individual's whole-brain fiber atlas.
3. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 2, characterized in that The specific steps of reconstructing the three-dimensional spatial distribution of whole-brain white matter fibers through a fiber tracking method in Step 1.2 include: Step 1.2.1: Using a probabilistic fiber tracking method, performing whole-brain fiber tracking on all individuals based on the first-order integration method of fiber direction distribution, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm; Step 1.2.2: Using a probabilistic fiber tracking method, performing whole-brain fiber tracking on all individuals based on the second-order integration method of fiber direction distribution, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm; Step 1.2.3: Using a deterministic fiber tracking method, performing whole-brain fiber tracking on all individuals based on streamline tractography of spherical deconvolution, generating 1,000,000 fibers, with a minimum length of 20 mm and a maximum length of 250 mm; Step 1.2.4: Merging the fibers obtained in Steps 1.2.1 - 1.2.3 as the individual's whole-brain fibers.
4. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 2, characterized in that, The specific steps of performing two consecutive clusterings on the group fibers in Step 1.4 include: First, performing cortical division on the whole-brain fibers according to the Yeo2011 functional brain network partition atlas, and dividing the fibers into different categories according to the different cortical functional networks they connect; The first clustering clusters different fiber categories, and the number of categories is where n is the total number of fibers in this fiber classification; The second clustering is performed on each fiber subclass obtained after the first clustering, and the number of categories is where m is the total number of fibers in this fiber subclass.
5. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 1, characterized in that, The specific steps of Step 2 include: Step 2.1: According to a preset geometric feature threshold, screening candidate fiber clusters in the whole-brain fiber atlas and calculating the probability that the fiber cluster belongs to the superficial white matter based on geometric features; Step 2.2: Establishing a deep learning model classifier to determine whether a fiber cluster belongs to the superficial white matter and calculating the probability that the fiber cluster belongs to the superficial white matter based on deep learning methods; Step 2.3: Jointly analyze the probabilities of the two types of fibers belonging to the superficial white matter obtained in Step 2.1 and Step 2.2, and retain the fiber clusters with probabilities of belonging to the superficial white matter greater than the probability threshold under both calculation methods. Step 2.4: Based on the fiber clusters retained in Step 2.3, in combination with the functional brain network cortical segmentation map mask, retain the fibers that only connect adjacent brain regions. Step 2.5: Perform secondary hierarchical clustering on the superficial white matter fibers obtained in Step 2.4 to adaptively generate the optimal sub-map of superficial white matter fibers with the optimal number of clusters.
6. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 5, characterized in that, The geometric features in Step 2.1 specifically include: Step 2.1.1: Bending degree threshold: The ratio of the middle segment distance of the fiber to the total length is less than 2. Step 2.1.2: Fiber length range: The ratio of the middle segment distance of the fiber to the end point distance is greater than 1.
1. Step 2.1.3: Endpoint spatial hemisphere position constraint: The ratio of the total length of the fiber to the end point distance is greater than 1.1, and the total length range of the fiber is between 15 mm and 85 mm.
7. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 5, characterized in that, The probability that the fiber cluster belongs to the superficial white matter calculated by the deep learning method in Step 2.2 specifically includes: Using the fibers in the ORG whole-brain fiber map as the training data input and the deep / superficial white matter labels in the ORG map as the data ground truth, establish a fully connected network to predict the category of the fiber cluster; the fully connected network is a three-layer fully connected neural network, the input is a 4×3-dimensional fiber feature vector matrix, and after passing through two hidden layers activated by ReLU, the final output is a 2-dimensional classification result.
8. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Use the trajectory-weighted dynamic functional connectivity calculation method to generate the whole-brain TW-dFC image of the sample based on the individual's superficial white matter fiber image and functional magnetic resonance image. Step 3.2: Perform independent component analysis on the whole-brain TW-dFC images of all samples to obtain the whole-brain voxel-level segmentation of the superficial white matter at the population level. Step 3.3: Calculate the number of different independent components passed through by each population's superficial white matter cluster, and calculate the uncertainty of the superficial white matter cluster according to the consistency degree of the independent components passed through by the fibers in the cluster. Step 3.4: Calculate the overlap degree between different independent components and the functional brain networks connected by the white matter fibers to reveal the correlation between different superficial white matter components and brain functions.
9. A method for constructing a human brain superficial white matter fiber structure atlas according to claim 8, characterized in that The calculation formula for the uncertainty is: wherein, is the uncertainty measure of the superficial white matter cluster kk, represents the probability that the points within the superficial white matter cluster k belong to the same M independent components, N points represents the total number of points in the superficial white matter cluster k, N(M) same IC is the number of fiber points belonging to the same M independent components.
10. A human brain superficial white matter fiber structure atlas construction system adopting the method according to any one of claims 1-9, characterized in that, Including: An unsupervised partitioning unit that unsupervised generates a fine-partitioned whole-brain functional fiber cluster map based on the whole-brain diffusion magnetic resonance imaging data. A hierarchical clustering unit that combines geometric feature thresholds and deep learning classifiers to identify superficial white matter fibers from the whole-brain functional fiber clusters and perform hierarchical clustering to generate the optimal sub-map of superficial white matter fibers with the optimal number of clusters. An analysis unit that analyzes the functional attributes of the superficial white matter fibers through the trajectory-weighted dynamic functional connectivity analysis method to obtain the human brain superficial white matter fiber structure diagram.