Automatic extraction method of brain functional network by integrating clustering and independent component analysis
Through the method of fusion clustering and independent component analysis, the problem of inaccurate estimation of ICA components is solved, stable extraction of brain functional networks and individual-specific analysis are achieved, and the accuracy of mental illness diagnosis is improved.
Patent Information
- Application Number
- CN202110825113.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-07-21
- Publication Date
- 2025-08-26
- Estimated Expiration
- 2041-07-21
AI Technical Summary
The existing independent component analysis methods are insufficiently accurate when estimating the number of components, and cannot integrate the results of different components at the same time, resulting in unstable brain functional network analysis.
The fusion clustering and independent component analysis method was used to construct multi-level PCA and ICA decomposition, and cluster it in combination with community discovery algorithms. GIG-ICA was used to estimate individual-level brain functional networks, and stable group-level and individual-specific brain functional networks were obtained through noise removal.
It realizes accurate and automatic extraction of brain functional networks, improves the stability and repeatability of results, and provides reliable brain functional network indicators for the research and diagnosis of mental illnesses.
Smart Images
Figure CN113610780B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of brain image analysis, and is particularly suitable for extracting brain functional networks. Background Art
[0002] Functional magnetic resonance imaging (fMRI) provides a powerful means for extracting and analyzing brain functional networks, greatly promoting the progress of brain science research and the diagnosis and treatment of mental illness. Currently, using brain functional imaging data to analyze brain functional networks has become a very important research topic.
[0003] Independent component analysis (ICA) is a data-driven method for analyzing brain functional networks with a wide range of applications. However, in applications where ICA is used to extract brain functional networks, the number of components has a significant impact on the results of ICA. To estimate the number of ICA components, some researchers have used information theory methods. However, due to the complex noise structure, it is impossible to accurately estimate the number of ICA components, and different estimation methods also produce different results. In addition, some work has set a certain range for the number of components, conducted multiple experiments within this range, and then determined the number of components based on indicators such as the stability and repeatability of the results on different datasets. However, to date, no ICA method can accurately estimate the number of components, nor can any ICA method simultaneously fuse components with different numbers of components. Summary of the Invention
[0004] In order to solve the above-mentioned problems, the present invention provides a method for automatically extracting brain functional networks by integrating clustering and independent component analysis.
[0005] The technical solution adopted by the present invention to solve the technical problem is to construct a method for automatically extracting brain functional networks by integrating clustering and independent component analysis, including:
[0006] The pre-processed brain functional imaging data of multiple subjects were concatenated based on the time point direction. Under different component number settings, two-level principal component analysis (PCA) was performed at the individual level and the group level to reduce the data dimension. Then, independent component analysis (ICA) was performed on the reduced data to obtain all the initial group-level independent components.
[0007] Based on the split-merge method combined with the community discovery algorithm, all initial group-level independent components are automatically clustered and reliable group-level independent components are extracted;
[0008] Based on the reliable group-level independent components, the noise components at the group level are removed to obtain the group-level brain functional network;
[0009] Using the group-level brain functional network as reference information, group information guided independent component analysis (GIG-ICA) was performed on the brain functional imaging data of individual subjects to estimate the brain functional network at the individual level.
[0010] The fusion of similar brain functional networks at the individual level and the removal of noise components are performed to obtain brain functional networks with individual specificity for different subjects.
[0011] The steps of preprocessing the subject's brain image data include:
[0012] The brain functional imaging data of the first few time points were removed, and the brain functional imaging data of the remaining time points were subjected to time layer correction, head motion correction, spatial normalization, and spatial smoothing.
[0013] The steps of using PCA dimensionality reduction and ICA decomposition to analyze multi-subject brain functional imaging data with different numbers of components to obtain initial group-level independent components include:
[0014] For each subject, the three-dimensional brain functional imaging data at each time point of the individual subject was converted into a row vector, and the row vectors obtained at different time points were concatenated in the time dimension to obtain a brain functional imaging data matrix of the number of time points × the number of effective voxels in the brain;
[0015] When the number of components is set to k, the brain functional imaging data matrix of each subject is subjected to individual level PCA dimensionality reduction in the time dimension, and then the individual level PCA dimensionality reduction results of each subject are concatenated in the time dimension, and group level PCA dimensionality reduction with the number of components being k is performed in the time dimension;
[0016] Perform ICA decomposition with k components on the result of group-level PCA dimensionality reduction to obtain a set of independent component sets;
[0017] Based on multiple different k, multiple individual-level PCA dimensionality reduction, group-level PCA dimensionality reduction and ICA decomposition are performed to obtain multiple sets of independent components, and all the obtained independent components are used as the initial group-level independent components.
[0018] Among them, based on the split-merge method combined with the community discovery algorithm, all initial group-level independent components are automatically clustered, and the steps of extracting reliable group-level independent components include:
[0019] Perform initial clustering on all initial group-level independent components using a community discovery algorithm to obtain category labels of the initial group-level independent components, where initial group-level independent components with the same category label are divided into the same cluster;
[0020] The cluster is split until the number of clusters does not change. The splitting rule is: if the sum of the distances between the closest and farthest components to the cluster center is greater than twice the average inter-cluster distance, the cluster is split using the community discovery algorithm. The average inter-cluster distance is the average of the differences in the absolute values of the Pearson correlation coefficients between the unit distance 1 and the cluster centers of all current clusters.
[0021] Perform cluster merging until the number of clusters no longer changes. The merging rule is: if the distance between the centers of two clusters is less than half of the average distance between clusters, then merge the two clusters;
[0022] Cluster splitting and cluster merging are repeated until the category labels of all initial group-level independent components no longer change.
[0023] The steps of removing noise components based on reliable group-level independent components to obtain a group-level brain functional network include:
[0024] Noise removal methods were used to remove noise components and obtain group-level brain functional networks. Noise removal can be achieved by using the noisecloud method to establish a classification model to distinguish between noise components and components corresponding to meaningful brain functional networks, or by observing the spatial information of the components.
[0025] The steps of performing GIG-ICA to estimate the individual-level brain functional network using the brain functional imaging data of the individual subjects and the group-level brain functional network include:
[0026] Taking each group-level brain functional network as a reference and based on the brain functional imaging data of individual subjects, a multi-objective optimization strategy is adopted to simultaneously optimize the independence of the individual subject's brain functional network and the correspondence between the individual subject's brain functional network and the group-level brain functional network to obtain the individual subject's brain functional network.
[0027] The steps of performing the fusion of similar brain functional networks at the individual level and removing noise components to obtain brain functional networks with individual specificity for different subjects include:
[0028] For each sample's individual-level brain functional network, the Pearson correlation coefficient was used to measure the correlation between independent components at the individual level. Two or more brain functional networks with large correlation coefficients were fused, and then noise removal methods were used to remove the noise components to obtain the individual-level brain functional network. The noise removal method can use NoiseCloud to establish a classification model to distinguish between noise components and effective components, or it can identify components by observing their spatial and time-frequency information.
[0029] Different from the existing technology, the automatic extraction method of brain functional network by integrating clustering and independent component analysis provided by the present invention mainly includes the following processes: performing ICA decomposition under different numbers of components to obtain initial group-level independent components. Performing a split-merge method combined with a community discovery algorithm to cluster the initial group-level independent components to obtain reliable group-level independent components. Based on the reliable group-level independent components, the noise components are removed to obtain a group-level brain functional network. The individual-level brain functional network is estimated using the brain functional imaging data of individual subjects and the group-level brain functional network. The individual-level brain functional network is subjected to fusion of similar brain functional networks and removal of noise components, and finally a brain functional network with individual specificity for different subjects is obtained. Through the present invention, the problem that the previous ICA method requires artificially given number of components is solved, and a sample-specific brain functional network can also be obtained, providing brain functional network indicators for the research and diagnosis of mental illness, which is of great significance to the diagnosis and treatment of mental illness. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0031] Figure 1 It is a flow chart of a method for automatically extracting brain functional networks by integrating clustering and independent component analysis provided by the present invention.
[0032] Figure 2 It is a schematic diagram of the results obtained by analyzing two different sample data on simulated data using an automatic brain functional network extraction method that integrates clustering and independent component analysis provided by the present invention.
[0033] Figure 3 It is a box plot of the Pearson correlation between the individual level components and the simulated components estimated in two groups of simulated data samples by an automatic brain function network extraction method that integrates clustering and independent component analysis provided by the present invention.
[0034] Figure 4 This is a schematic diagram of the results of a brain functional network estimated based on real data in an automatic brain functional network extraction method that integrates clustering and independent component analysis provided by the present invention. DETAILED DESCRIPTION
[0035] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described are only used to explain the present invention, but are not used to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without making creative work will fall within the scope of protection of the present invention.
[0036] Refer to the attached Figure 1 The present invention provides a method for automatically extracting brain functional networks by integrating clustering and independent component analysis, comprising:
[0037] 1. The preprocessed brain functional imaging data of multiple subjects were concatenated based on the time point direction. Under different component number settings, two-level PCA at the individual level and group level was performed to reduce the data dimension. Then, ICA decomposition was performed on the reduced data to obtain all the initial group-level independent components.
[0038] 2. Based on the split-merge method combined with the community discovery algorithm, all initial group-level independent components are automatically clustered and reliable group-level independent components are extracted.
[0039] 3. Based on the reliable group-level independent components, the noise components at the group level are removed to obtain the group-level brain functional network.
[0040] 4. Using the group-level brain functional network as reference information, GIG-ICA was performed on the brain functional imaging data of individual subjects to estimate the individual-level brain functional network.
[0041] 5. Perform fusion of similar brain functional networks at the individual level and remove noise components to obtain brain functional networks with individual specificity for different subjects.
[0042] Among them, the process of preprocessing the subject's brain image data mainly includes: removing the first few time points, time layer correction, head movement correction, spatial normalization, spatial smoothing, etc.
[0043] The step of performing PCA dimensionality reduction and ICA decomposition on the brain imaging data of the subjects with different numbers of components to obtain initial group-level independent components includes: for each subject, converting the three-dimensional brain functional imaging data at each time point into a row vector, and concatenating the row vectors obtained at different time points along the time dimension to obtain a brain functional imaging data matrix with the number of time points × the number of effective voxels in the brain. With the number of components set to k, individual-level PCA dimensionality reduction is performed on each subject's brain functional imaging data matrix along the time dimension. The individual-level PCA dimensionality reduction results for each subject are then concatenated along the time dimension, and group-level PCA dimensionality reduction is performed along the time dimension with k components. The results of the group-level PCA dimensionality reduction are then subjected to ICA decomposition with k components to obtain a set of independent components. Multiple individual-level PCA dimensionality reduction, group-level PCA dimensionality reduction, and ICA decomposition are performed based on different values of k to obtain multiple sets of independent components, and all obtained independent components are used as the initial group-level independent components.
[0044] Among them, based on the split-merge rule proposed in combination with the community discovery algorithm, the automatic clustering of the initial group-level independent components is completed, and the steps of extracting reliable group-level independent components include: calculating the Pearson correlation coefficient between the initial group-level independent components, constructing a graph structure, using the community discovery algorithm to obtain the initial category labels of the initial group-level independent components, and then updating the initial category labels based on the proposed split-merge rule to obtain stable group-level independent components.
[0045] All initial group-level independent components are initially clustered using a community discovery algorithm to obtain their category labels. Initial group-level independent components with the same category label are grouped into the same cluster. Cluster splitting is performed until the number of clusters no longer changes. The splitting rule is: if the sum of the distances from the cluster center to the components closest to and farthest from the cluster center in a cluster is greater than twice the average inter-cluster distance, the cluster is split using the community discovery algorithm, where the average inter-cluster distance is the average of the differences in the absolute values of the Pearson correlations between the unit distance 1 and the cluster centers of all current clusters. Cluster merging is performed until the number of clusters no longer changes. The merging rule is: if the distance between the cluster centers of two clusters is less than half the average inter-cluster distance, the two clusters are merged. Cluster splitting and merging are repeated until the category labels of all initial group-level independent components no longer change.
[0046] Because the number of independent components (the number of components) is often determined based on experience rather than theory, the resulting stable group-level independent components are unreliable. This paper combines community discovery algorithms with split-merge rules to perform stable clustering. This allows independent component extraction to overcome the subjectivity of parameter setting and improves the repeatability of stable independent component extraction.
[0047] The present invention uses the difference between the standard unit 1 and the absolute value of the Pearson correlation of the two components as the distance measurement method between the two components. Among them, the splitting rule is:
[0048] Step s1: Calculate the average distance between all current clusters (d mean ).
[0049] Step s2: For each cluster, calculate the sum of the distances between the closest and farthest components (excluding the cluster center itself) within the cluster and the cluster center (d intra ).
[0050] Step s3: For a cluster, if d intra >d mean / 2, then we use the community discovery algorithm to split the cluster.
[0051] Step s4: Update the cluster center and the number of current clusters.
[0052] Step s5: If the number of clusters has changed, return to step s1; if not, the splitting operation ends.
[0053] The merging rules are:
[0054] Step m1: Calculate the average distance between all current clusters (d mean ).
[0055] Step m2: Calculate the distance between the cluster centers of any two clusters in all current clusters (d inter ).
[0056] Step m3: For any two clusters, if the distance between their cluster centers satisfies d inter <d mean / 2, then merge the two clusters.
[0057] Step m4: Update the cluster center and the number of current clusters.
[0058] Step m5: If the number of clusters has changed, return to step m1; if not, the merge operation ends.
[0059] Repeat the above splitting and merging rules until the category labels at the initial group level no longer change, thereby obtaining reliable group-level independent components.
[0060] The steps of removing group-level noise components based on reliable group-level independent components to obtain a group-level brain functional network include:
[0061] The noise removal method removes the noise components and obtains the brain functional network at the group level. The noise removal method can use the noisecloud method to distinguish between noise and effective components by establishing a classification model.
[0062] The steps of performing GIG-ICA to estimate the individual-level brain functional network using the brain functional imaging data of the individual subjects and the group-level brain functional network include:
[0063] Taking each group-level brain functional network as a reference and based on the brain functional imaging data of individual subjects, a multi-objective optimization strategy is adopted to simultaneously optimize the independence of the individual subject's brain functional network and the correspondence between the individual subject's brain functional network and the group-level brain functional network to obtain the individual subject's brain functional network.
[0064] The steps of performing the fusion of similar brain functional networks at the individual level and removing noise components to obtain brain functional networks with individual specificity for different subjects include:
[0065] For each sample's individual-level brain functional network, the Pearson correlation coefficient was used to measure the correlation between independent components at the individual level. Two or more brain functional networks with large correlation coefficients were fused, and then noise removal methods were used to remove the noise components to obtain the individual-level brain functional network. The noise removal method can use NoiseCloud to establish a classification model to distinguish between noise components and effective components, or it can identify components by observing their spatial and time-frequency information.
[0066] Data Generation in This Example: Simulated data was generated using the SimTB toolbox. This experiment simulated two groups of samples, each containing 15 samples. The simulated data for each sample consisted of eight components. Six of these eight components shared common spatial patterns between the two groups, while two exhibited distinct spatial patterns within each group. The real data for this example came from functional magnetic resonance imaging (fMRI) data from 25 healthy subjects.
[0067] Figure 2 The results of the present invention's analysis of two different sample data on simulated data show the estimated individual-level independent components and the simulated independent components, as well as the Pearson correlation (r value) between the estimated and simulated components. (A) and (B) represent two samples from different groups.
[0068] Figure 3The following is a box plot of the Pearson correlation between the estimated individual-level components and the simulated components in two groups of simulated data. (A) shows the results for the common components between the two groups; (B) shows the results for the specific components of each group.
[0069] Figure 4 Schematic diagram of the results of brain functional networks estimated based on real data. The results of brain functional networks estimated using the present invention based on real data are shown. (A) shows two brain functional networks from two samples with the maximum spatial similarity (spatial similarity is represented by r) estimated based on the same group-level independent component. (B) shows two brain functional networks from two samples with the minimum spatial similarity estimated based on the same group-level independent component.
[0070] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.
Claims
1. A method for automatically extracting brain functional networks by integrating clustering and independent component analysis, characterized in that: The steps include: The pre-processed brain functional imaging data of multiple subjects were concatenated based on the time point direction. Under different component number settings, two-level principal component analysis (PCA) was performed at the individual level and group level to reduce the data dimension. Then, independent component analysis (ICA) was performed on the reduced data to obtain all the initial group-level independent components. The steps of using PCA dimensionality reduction and ICA decomposition to analyze multi-subject brain functional imaging data with different numbers of components to obtain initial group-level independent components include: For each subject, the three-dimensional brain functional imaging data at each time point of the individual subject was converted into a row vector, and the row vectors obtained at different time points were concatenated in the time dimension to obtain a brain functional imaging data matrix of the number of time points × the number of effective voxels in the brain; When the number of components is set to k, the brain function imaging data matrix of each subject is subjected to individual level PCA dimensionality reduction in the time dimension, and then the individual level PCA dimensionality reduction results of each subject are concatenated in the time dimension, and the number of components is set to k in the time dimension. Group-level PCA dimensionality reduction; Perform ICA decomposition with k components on the result of group-level PCA dimensionality reduction to obtain a set of independent component sets; Based on multiple different k, multiple individual-level PCA dimensionality reduction, group-level PCA dimensionality reduction and ICA decomposition are performed to obtain multiple sets of independent components, and all the obtained independent components are used as the initial group-level independent components; Based on the split-merge method combined with the community discovery algorithm, all initial group-level independent components are automatically clustered and reliable group-level independent components are extracted; Based on the reliable group-level independent components, the noise components at the group level are removed to obtain the group-level brain functional network; Using the group-level brain functional network as reference information, group-information guided independent component analysis (GIG-ICA) was performed on the brain functional imaging data of individual subjects to estimate the brain functional network at the individual level. The fusion of similar brain functional networks at the individual level and the removal of noise components are performed to obtain brain functional networks with individual specificity for different subjects.
2. The method for automatically extracting brain functional networks by integrating clustering and independent component analysis according to claim 1, characterized in that: The steps for preprocessing the subject's brain image data include: The brain functional imaging data of the first few time points were removed, and the brain functional imaging data of the remaining time points were subjected to time layer correction, head motion correction, spatial normalization, and spatial smoothing.
3. The method for automatically extracting brain functional networks by integrating clustering and independent component analysis according to claim 1, characterized in that: Based on the split-merge method combined with the community discovery algorithm, all initial group-level independent components are automatically clustered. The steps to extract reliable group-level independent components include: Perform initial clustering on all initial group-level independent components using a community discovery algorithm to obtain category labels of the initial group-level independent components, where initial group-level independent components with the same category label are divided into the same cluster; The cluster is split until the number of clusters does not change. The splitting rule is: if the sum of the distances between the components closest to and farthest from the cluster center is greater than twice the average inter-cluster distance, the cluster is split using the community discovery algorithm, where the average inter-cluster distance is the average of the differences in the absolute values of the Pearson correlation coefficients between the unit distance 1 and the cluster centers of all current clusters. The clusters are merged until the number of clusters does not change. The merging rule is: if the distance between the centers of two clusters is less than half of the average distance between clusters, the two clusters are merged. Cluster splitting and cluster merging are repeated until the category labels of all initial group-level independent components no longer change.
4. The method for automatically extracting brain functional networks by integrating clustering and independent component analysis according to claim 1, characterized in that: Based on reliable group-level independent components, the noise components are removed to obtain the group-level brain functional network. The steps include: The noise removal method is used to remove the noise components and obtain the brain functional network at the group level. Noise removal can be achieved by using the noisecloud method to establish a classification model to distinguish between noise components and components corresponding to meaningful brain functional networks, or by observing the spatial information of the components.
5. The method for automatically extracting brain functional networks by integrating clustering and independent component analysis according to claim 1, characterized in that: Using the brain functional imaging data of individual subjects and the brain functional network at the group level, the steps of performing GIG-ICA to estimate the individual-level brain functional network include: Taking each group-level brain functional network as a reference and based on the brain functional imaging data of individual subjects, a multi-objective optimization strategy is adopted to simultaneously optimize the independence of the individual subject's brain functional network and the correspondence between the individual subject's brain functional network and the group-level brain functional network to obtain the individual subject's brain functional network.
6. The method for automatically extracting brain functional networks by integrating clustering and independent component analysis according to claim 1, characterized in that: The steps for performing the fusion of similar brain functional networks at the individual level and removing noise components to obtain brain functional networks with individual specificity for different subjects include: For the individual-level brain functional network of each sample, the Pearson correlation coefficient is used to measure the correlation between independent components at the individual level. Two or more brain functional networks with larger correlation coefficients are fused, and then the noise components are removed using the noise removal method to obtain the individual-level brain functional network. Among them, the noise removal method can use noisecloud to distinguish between noise components and effective components by establishing a classification model, or it can be identified by observing the spatial information and time-frequency information of the components.
Citation Information
Patent Citations
Method for recognizing function response signal under function nuclear magnetic resonance scan
CN101788656A
Detection method for causal connection strength of magnetic resonance brain imaging based on PCA (Principal component analysis) and GCA (Granger causality analysis)
CN102366323A