A system for constructing individual and group brain maps based on multi-view clustering

Abstract

The present invention discloses a system for constructing individual and group brain maps based on multi-view clustering. The system comprises an experimental paradigm design module, an image processing module, a seed point acquisition module, a brain function map construction module, and a brain function map real-time display module. The experimental paradigm design module is used to determine the experimental content and process for multiple subjects and is designed according to the researcher's experimental objectives. The image processing module pre-processes and saves the collected structural magnetic resonance images and functional magnetic resonance images. The seed point acquisition module uses the SLIC method to resample the pre-processed functional magnetic resonance data of each subject to extract seed points. The brain function map construction module uses the multi-view clustering method to simultaneously obtain brain function maps at the individual and group levels. The brain function map real-time display module maps the obtained maps to the cortex and provides feedback to the researcher in the form of images. The present invention can simultaneously generate brain function maps at the individual and group levels.

Description

Technical Field

[0001] The present invention relates to the technical field of brain function map construction, and in particular to an individual and group brain map construction system based on multi-view clustering. Background Art

[0002] The brain is the human neural center, with complex structure and function. Studying its workings is a crucial area of ​​neuroscience research. To assist neuroscientists and clinicians in this cutting-edge field, constructing maps of brain structure and function is crucial. By normalizing images to a brain map space, statistical analysis of brain images at both the group and individual levels within a "common coordinate system" can be performed. The brain can be represented as a network. Using graph theory analysis methods, the brain can be modeled as a graph, with brain regions as nodes and functional regions' correlations or interregional white matter fiber connections as edges.

[0003] Since there is currently no standard definition for brain network nodes, the selection of brain network nodes remains an unresolved issue. Although voxel-based representations can be used, the resulting networks are noisy and high-dimensional, making them unsuitable for further network analysis. To control potential noise and compress network size, node definition methods based on brain parcellation maps are currently commonly used. Each node corresponds to a brain region in the atlas and is composed of multiple voxels with similar structure or function. The quality of the brain parcellation map directly affects the performance of the constructed brain network, and therefore high-quality brain maps are a prerequisite for effective brain network analysis.

[0004] In recent years, the rapid development of magnetic resonance imaging (MRI) technology has provided a new avenue for constructing brain maps. Compared to traditional landmarks (such as sulci and gyri), MRI technology can provide a more detailed depiction of the structural and functional properties of cerebral cortical cells. Combined with machine learning methods, higher-quality brain maps can be created. Compared to structural maps, functional maps developed later, but with greater functional consistency, they have gradually gained widespread attention and application in various brain function research. Currently, researchers have constructed two main types of brain functional maps: group-level brain functional maps and individual-level brain functional maps. Group maps are obtained by clustering the averaged individual similarity matrix or by group-clustering the individual clustering results. They can provide important observational metrics and statistical data on the macroscopic organization of the brain. However, group maps lack the ability to reflect inter-individual differences, making it difficult to accurately detect behavioral predictions or disease diagnoses that are related to these differences. Therefore, individual-level brain functional maps are gaining increasing importance in neuroscience and clinical research. They can not only predict human cognition, behavior, and emotion, but also capture individual differences in brain topology. However, current researchers mostly construct group-level and individual-level brain functional maps using classic clustering and image segmentation algorithms, such as k-means, region growing, hierarchical clustering, and spectral clustering. There is a lack of methods for simultaneously constructing both types of maps using rs-fMRI data from subjects. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an individual and group brain map construction system based on multi-view clustering, which can simultaneously generate brain function maps at the individual level and the group level.

[0006] To solve the above technical problems, the present invention provides an individual and group brain map construction system based on multi-view clustering, including: an experimental paradigm design module, an image processing module, a seed point acquisition module, a brain function map construction module and a brain function map real-time display module; the experimental paradigm design module is used to determine the experimental content and process in which multiple subjects participate, and is designed according to the experimental purpose of the researcher; the image processing module preprocesses and saves the collected structural magnetic resonance images and functional magnetic resonance images; the seed point acquisition module uses the SLIC method to resample the preprocessed functional magnetic resonance data of each subject to extract seed points; the brain function map construction module uses the multi-view clustering method to simultaneously obtain brain function maps at the individual level and the group level; the brain function map real-time display module maps the obtained maps to the cortex and feeds them back to the researcher in the form of images.

[0007] Preferably, the experimental paradigm design module adopts a resting-state scanning method.

[0008] Preferably, the image processing module includes a parameter configuration unit, an image reading unit, a format conversion unit and an image preprocessing unit; the parameter configuration unit is used to determine the scanning parameters of the magnetic resonance image; the image reading unit is used to read the collected structural magnetic resonance data and functional magnetic resonance data; the format conversion unit is used to convert the magnetic resonance data from the DICOM data format to the NIFTY data format; the image preprocessing unit is used to preprocess the magnetic resonance data, including: head motion correction, registration of functional images and structural images, and filtering and smoothing operations of the functional images.

[0009] Preferably, the seed point acquisition module uses the SLIC method to resample the pre-processed functional magnetic resonance data of each subject to extract seed points; the SLIC algorithm divides the rs-fMRI data into sub-regions based on functional correlation and spatial proximity. This method limits the search space to the neighborhood of the cluster center, which can reduce the number of distance calculations; at the same time, by integrating spatial distance and functional correlation to create a unified distance metric function, the spatial continuity within the cluster can be guaranteed; the specific steps are as follows: first, initialize k cluster centers, set N represents the number of voxels. For each voxel in the 3S*3S*3S region around the cluster center, the distance D between the voxel and the cluster center is calculated:

[0010]

[0011] where d c and d g Corresponding to functional and spatial distance metrics, respectively, functional similarity is measured by Pearson correlation, and spatial distance is measured by the geometric distance along the cortical surface, which is approximately the shortest path length between nodes, N c and N g Represents the normalization factor. The distance is used as a measure to determine which cluster the voxel should belong to. If the distance decreases compared with the result of the previous iteration, the voxel is associated with the current cluster center. The above assignment and update steps are repeated until the change in the cluster center is lower than the threshold ε. The k cluster centers finally obtained are the seed points.

[0012] Preferably, the brain function map construction module includes a time series extraction module, an individual similarity matrix acquisition module and a consistent similarity matrix acquisition module; the time series extraction module is used to extract the time series of seed points; the individual similarity matrix acquisition module is used to construct an individual similarity matrix for each subject, by performing a Fisher transformation on the time series of the seed points to normalize it, and then calculating the Pearson correlation coefficient between the time series of the seed points to obtain the individual similarity matrix; the consistent similarity matrix acquisition module generates a group consistent similarity matrix based on the constructed individual similarity matrix using a multi-view clustering method.

[0013] Preferably, the consistent similarity matrix acquisition module generates a group consistent similarity matrix based on the constructed individual similarity matrix using a multi-view clustering method, specifically comprising the following steps:

[0014] Step 1: Construct the spectral embedding matrix of the individual network. Given the subject v, assume W (v) Denotes the similarity matrix, D (v) Is a diagonal matrix, and the corresponding spectral embedding matrix H is calculated using the following formula (v) :

[0015]

[0016] Among them, A (v) =D (v)-1 / 2 W (v) D (v)-1 / 2 , then, the spectrum is embedded into the matrix H (v) The row normalization of the normalized spectral embedding matrix is ​​obtained in,

[0017] Step 2: Construct the third-order tensor and optimize the individual spectrum embedding matrix; given the subject v, assuming the number of nodes is n, is the Gram matrix, which is reorganized into a third-order tensor The following formula is used to optimize the individual spectrum embedding matrix:

[0018]

[0019] Where λ is the balance parameter, τ is the singular value threshold, For tensors, tensors For optimizing tensors Able to better capture tensors The principal components of By tensor singular value decomposition, we can get:

[0020]

[0021] in, and are two orthogonal tensors, is a diagonal tensor, t + =max(t,0), Indicates The result of discrete Fourier transform (DFT) on

[0022] In equation (4), the singular values ​​are shrunk with the same threshold τ. However, excessive shrinkage of large singular values ​​will result in a biased solution, which is not conducive to mining the principal components of the tensor. Therefore, we introduce a weighted tensor nuclear norm to improve its flexibility:

[0023]

[0024] Among them, r is a tensor The rank of n is the number of nodes, is the weight of the singular value;

[0025] Step 3: Obtain a group-consistent similarity matrix; normalize the rows of the optimized individual spectral embedding matrix obtained in step 2 to obtain a normalized spectral embedding matrix. The individual Gram matrix is ​​obtained by calculating the inner product of the optimized normalized spectral embedding matrix. Since the Gram matrix is ​​obtained through the inner product, its diagonal elements can represent the characteristics of the node itself, and the off-diagonal elements can represent the characteristics between nodes. Therefore, the Gram matrix can be used to represent the characteristics of the individual. Therefore, the averaged Gram matrix is ​​the group-consistent similarity matrix.

[0026] Preferably, the real-time display module of the brain function map is used to map the obtained map to the cortex and feed back the image to the researchers.

[0027] The beneficial effects of the present invention are as follows: the present invention uses resting-state functional magnetic resonance imaging (rs-fMRI) data and a multi-view clustering method to simultaneously generate brain function maps at the individual and group levels, which not only ensures the consistency of the individual and group-level brain function maps, but also reflects the differences between individuals. The brain function maps generated by this method have good reproducibility and homogeneity. In addition, by obtaining seed points through the SLIC method and introducing a tensor representation method, the original data can be reduced in dimensionality, which greatly reduces the running time and improves efficiency. The individual-level brain maps and group-level brain maps constructed by the present invention are conducive to obtaining more accurate biometric information and provide assistance for clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the system structure of the present invention.

[0029] Figure 2 Schematic diagram of the structure of the image processing module of the present invention.

[0030] Figure 3 A schematic diagram of the process of generating a group consensus similarity matrix according to the present invention.

[0031] Figure 4 It is a schematic diagram of the evaluation results of the method of the present invention. DETAILED DESCRIPTION

[0032] like Figure 1 As shown, a system for constructing individual and group brain maps based on multi-view clustering includes: an experimental paradigm design module, an image processing module, a seed point acquisition module, a brain function map construction module and a brain function map real-time display module; the experimental paradigm design module is used to determine the experimental content and process in which multiple subjects participate, and is designed according to the experimental purpose of the researcher; the image processing module preprocesses and saves the collected structural magnetic resonance images and functional magnetic resonance images; the seed point acquisition module uses the SLIC method to resample the preprocessed functional magnetic resonance data of each subject to extract seed points; the brain function map construction module uses the multi-view clustering method to simultaneously obtain brain function maps at the individual level and the group level; the brain function map real-time display module maps the obtained maps to the cortex and feeds them back to the researcher in the form of images.

[0033] The experimental paradigm design module adopts the resting-state scanning method.

[0034] like Figure 2 As shown, the image processing module includes a parameter configuration unit, an image reading unit, a format conversion unit and an image preprocessing unit; the parameter configuration unit is used to determine the scanning parameters of the magnetic resonance image; the image reading unit is used to read the collected structural magnetic resonance data and functional magnetic resonance data; the format conversion unit is used to convert the magnetic resonance data from the DICOM data format to the NIFTY data format; the image preprocessing unit is used to preprocess the magnetic resonance data, including: head motion correction, registration of functional images and structural images, and filtering and smoothing operations of the functional images.

[0035] The seed point acquisition module uses the SLIC method to resample the pre-processed functional magnetic resonance data of each subject to extract seed points. The SLIC algorithm divides the rs-fMRI data into sub-regions based on functional correlation and spatial proximity. This method limits the search space to the neighborhood of the cluster center, which can reduce the number of distance calculations. At the same time, by integrating spatial distance and functional correlation to create a unified distance metric function, the spatial continuity within the cluster can be guaranteed. The specific steps are as follows: First, initialize k cluster centers, set N represents the number of voxels. For each voxel in the 3S*3S*3S region around the cluster center, the distance D between the voxel and the cluster center is calculated:

[0036]

[0037] where d c and d g Corresponding to functional and spatial distance metrics, respectively, functional similarity is measured by Pearson correlation, and spatial distance is measured by the geometric distance along the cortical surface, which is approximately the shortest path length between nodes, N c and Ng Represents the normalization factor. The distance is used as a measure to determine which cluster the voxel should belong to. If the distance decreases compared with the result of the previous iteration, the voxel is associated with the current cluster center. The above assignment and update steps are repeated until the change in the cluster center is lower than the threshold ε. The k cluster centers finally obtained are the seed points.

[0038] like Figure 3 As shown in the figure, the brain function map construction module includes a time series extraction module, an individual similarity matrix acquisition module and a consistent similarity matrix acquisition module; the time series extraction module is used to extract the time series of seed points; the individual similarity matrix acquisition module is used to construct the individual similarity matrix of each subject, by performing Fisher transformation on the time series of the seed points to normalize it, and then calculating the Pearson correlation coefficient between the time series of the seed points to obtain the individual similarity matrix; the consistent similarity matrix acquisition module generates a group consistent similarity matrix based on the constructed individual similarity matrix using a multi-view clustering method.

[0039] Preferably, the consistent similarity matrix acquisition module generates a group consistent similarity matrix based on the constructed individual similarity matrix using a multi-view clustering method, specifically comprising the following steps:

[0040] Step 1: Construct the spectral embedding matrix of the individual network. Given the subject v, assume W (v) Denotes the similarity matrix, D (v) Is a diagonal matrix, and the corresponding spectral embedding matrix H is calculated using the following formula (v) :

[0041]

[0042] Among them, A (v) =D (v)-1 / 2 W (v) D (v)-1 / 2 , then, the spectrum is embedded into the matrix H (v) The row normalization of the normalized spectral embedding matrix is ​​obtained in,

[0043] Step 2: Construct the third-order tensor and optimize the individual spectrum embedding matrix; given the subject v, assuming the number of nodes is n, is the Gram matrix, which is reorganized into a third-order tensor The following formula is used to optimize the individual spectrum embedding matrix:

[0044]

[0045] Where λ is the balance parameter, τ is the singular value threshold, For tensors, tensors For optimizing tensors Able to better capture tensors The principal components of By tensor singular value decomposition, we can get:

[0046]

[0047] in, and are two orthogonal tensors, is a diagonal tensor, t + =max(t,0), Indicates The result of discrete Fourier transform (DFT) on

[0048] In equation (4), the singular values ​​are shrunk with the same threshold τ. However, excessive shrinkage of large singular values ​​will result in a biased solution, which is not conducive to mining the principal components of the tensor. Therefore, we introduce a weighted tensor nuclear norm to improve its flexibility:

[0049]

[0050] Among them, r is a tensor The rank of n is the number of nodes, is the weight of the singular value;

[0051] Step 3: Obtain a group-consistent similarity matrix; normalize the rows of the optimized individual spectral embedding matrix obtained in step 2 to obtain a normalized spectral embedding matrix. The individual Gram matrix is ​​obtained by calculating the inner product of the optimized normalized spectral embedding matrix. Since the Gram matrix is ​​obtained through the inner product, its diagonal elements can represent the characteristics of the node itself, and the off-diagonal elements can represent the characteristics between nodes. Therefore, the Gram matrix can be used to represent the characteristics of the individual. Therefore, the averaged Gram matrix is ​​the group-consistent similarity matrix.

[0052] The real-time display module of brain function maps is used to map the obtained maps to the cortex and provide feedback to researchers in the form of images.

[0053] We evaluated the performance of our method by comparing two segmentations obtained from the HCP test-retest dataset, performing reproducibility and functional homogeneity analyses on individual and group atlases. The number of clusters, k, was selected as 50, 100, 150, and 200. We used the Dice coefficient and the Adjusted Rand Index (ARI) to assess the reproducibility of the segmentation results, and the Silhouette coefficient (SI) to assess the functional homogeneity of the segmentation results. As shown in Table 1, our method produced highly reproducible results at each resolution level, with average Dice coefficients of at least 0.56 and ARIs of 0.51 for individual atlases and at least 0.64 and ARIs of 0.61 for group atlases. Overall, group-level segmentations were more reproducible than individual-level segmentations at each resolution, likely due to individual variability affecting segmentation reproducibility. There is a general downward trend for all metrics with increasing segmentation resolution, which can be attributed to the fact that as k increases, the differences between subjects become more significant, thereby reducing the similarity between common features within groups and leading to irreproducible group segmentation.

[0054] Table 1 Dice coefficients and ARI scores of individual and group-level maps

[0055]

[0056] We used the silhouette coefficient (SI) to evaluate the functional homogeneity of our method. As shown in Table 2, regardless of the segmentation resolution, our method produced good functional homogeneity results, and the homogeneity results of group-level and individual-level segmentations obtained by our method were highly consistent at different resolution levels, with an average SI score of at least 0.96 for individual brain maps and at least 0.95 for group brain maps.

[0057] Table 2 SI scores of individual and group-level profiles

[0058]

[0059]

[0060] We computed individual and group cortical parcellations of 50, 100, 150, and 200 brain regions using the HCP dataset. Figure 4 The individual and group segmentation results obtained by our method at various resolution levels are shown.

[0061] This study uses resting-state functional magnetic resonance imaging (fMRI) data and a multi-view clustering method to simultaneously generate brain functional maps at both the population and individual levels, demonstrating good reproducibility and functional homogeneity. The individual and population-level brain maps constructed using this method facilitate the acquisition of more accurate biometric information, providing insights for clinical applications.

Claims

1. A system for constructing individual and group brain maps based on multi-view clustering, characterized by: include: Experimental paradigm design module, image processing module, seed point acquisition module, brain function map construction module and brain function map real-time display module; the experimental paradigm design module is used to determine the experimental content and process of multiple subjects and design it according to the researcher's experimental purpose; the image processing module pre-processes and saves the collected structural magnetic resonance images and functional magnetic resonance images; The seed point acquisition module uses the SLIC method to resample the preprocessed functional magnetic resonance data of each subject to extract seed points; the brain function map construction module uses the multi-view clustering method to simultaneously obtain brain function maps at the individual and group levels. The brain function map construction module includes a time series extraction module, an individual similarity matrix acquisition module, and a consistent similarity matrix acquisition module. The time series extraction module is used to extract the time series of seed points. The individual similarity matrix acquisition module is used to construct the individual similarity matrix of each subject. The time series of the seed points is normalized by performing a Fisher transformation, and then the Pearson correlation coefficient between the time series of the seed points is calculated to obtain the individual similarity matrix. The consistent similarity matrix acquisition module uses the multi-view clustering method to generate a group consistent similarity matrix based on the constructed individual similarity matrix. The real-time display module of brain function maps the obtained maps to the cortex and feeds them back to researchers in the form of images.

2. The system for constructing individual and group brain maps based on multi-view clustering according to claim 1, characterized in that: The experimental paradigm design module adopts the resting-state scanning method.

3. The system for constructing individual and group brain maps based on multi-view clustering according to claim 1, characterized in that: The image processing module includes a parameter configuration unit, an image reading unit, a format conversion unit and an image preprocessing unit; the parameter configuration unit is used to determine the scanning parameters of the magnetic resonance image; the image reading unit is used to read the collected structural magnetic resonance data and functional magnetic resonance data; The format conversion unit is used for converting the magnetic resonance data from the DICOM data format to the NIFTY data format; The image preprocessing unit is used to preprocess the magnetic resonance data, including: head motion correction, registration of functional images and structural images, and filtering and smoothing operations of functional images.

4. The system for constructing individual and group brain maps based on multi-view clustering according to claim 1, characterized in that: The seed point acquisition module uses the SLIC method to resample the pre-processed functional magnetic resonance data of each subject to extract seed points; the SLIC algorithm divides the rs-fMRI data into sub-regions based on functional correlation and spatial proximity. This method limits the search space to the neighborhood of the cluster center and creates a unified distance metric function by integrating spatial distance and functional correlation. The specific steps are as follows: First, initialize k cluster centers, set N represents the number of voxels. For each voxel in the 3S*3S*3S region around the cluster center, the distance D between the voxel and the cluster center is calculated: where d c and d g Corresponding to functional and spatial distance metrics, respectively, functional similarity is measured by Pearson correlation, and spatial distance is measured by the geometric distance along the cortical surface, which is approximately the shortest path length between nodes, N c and N g Represents the normalization factor. The distance is used as a measure to determine which cluster the voxel should belong to. If the distance decreases compared with the result of the previous iteration, the voxel is associated with the current cluster center. The above assignment and update steps are repeated until the change in the cluster center is lower than the threshold ε. The k cluster centers finally obtained are the seed points.

5. The system for constructing individual and group brain maps based on multi-view clustering according to claim 1, characterized in that: The consistent similarity matrix acquisition module generates a group consistent similarity matrix based on the constructed individual similarity matrix using the multi-view clustering method. The specific steps include: Step 1: Construct the spectral embedding matrix of the individual network. Given the subject v, assume W (v) Denotes the similarity matrix, D (v) Is a diagonal matrix, and the corresponding spectral embedding matrix H is calculated using the following formula (v) : Among them, A (v) =D (v)-1 / 2 W (v) D (v)-1 / 2 , then, the spectrum is embedded into the matrix H (v) The row normalization of the normalized spectral embedding matrix is ​​obtained in, Step 2: Construct the third-order tensor and optimize the individual spectrum embedding matrix; given the subject v, assuming the number of nodes is n, is the Gram matrix, which is reorganized into a third-order tensor The following formula is used to optimize the individual spectrum embedding matrix: Where λ is the balance parameter, τ is the singular value threshold, For tensors, tensors For optimizing tensors Able to better capture tensors The principal components of By tensor singular value decomposition, we can get: in, and are two orthogonal tensors, is a diagonal tensor, Indicates The result of discrete Fourier transform (DFT) on In equation (4), the singular values ​​are shrunk with the same threshold τ, and a weighted tensor kernel norm is introduced to improve its flexibility: Among them, r is a tensor The rank of n is the number of nodes, is the weight of the singular value; Step 3: Obtain a group-consistent similarity matrix. Normalize the rows of the optimized individual spectral embedding matrix obtained in step 2 to obtain a normalized spectral embedding matrix. The individual Gram matrix is ​​obtained by calculating the inner product of the optimized normalized spectral embedding matrix. Since the Gram matrix is ​​obtained through the inner product, its diagonal elements can represent the characteristics of the node itself, and the off-diagonal elements can represent the characteristics between nodes. Therefore, the Gram matrix is ​​used to represent the characteristics of the individual. The averaged Gram matrix is ​​the group-consistent similarity matrix.

6. The system for constructing individual and group brain maps based on multi-view clustering according to claim 1, characterized in that: The real-time display module of brain function maps is used to map the obtained maps to the cortex and provide feedback to researchers in the form of images.

Images