A feature classification method and system based on individual difference brain network

Abstract

The application discloses a feature classification method based on individual difference brain networks, reference brain networks of normal data and abnormal data are respectively constructed based on normal brain region data and abnormal brain region data, data in the abnormal brain region data is respectively added into the reference brain network of the normal data to obtain a perturbed brain network of each data; data in the normal brain region data is respectively added into the reference brain network of the abnormal data to obtain a perturbed brain network of each data; the perturbed brain network of each data is subtracted from the corresponding reference brain network to obtain an individual difference brain network; a feature matrix is obtained according to the individual difference brain network of the data; feature selection and classification are performed according to the obtained feature matrix. Starting from multiple groups of individual difference brain networks, not only can the individualized features of individuals be paid attention to, but also the common difference between the individuals and the reference groups can be mined, so that the abnormal brain region discrimination precision is significantly improved.

Description

Technical Field

[0001] This invention relates to medical image processing, specifically to a feature classification method and system based on individual differences in brain networks. Background Technology

[0002] Brain network research has received increasing attention in recent years to better study, diagnose, and treat brain diseases. This research aims to model interconnected brain tissue to describe and explain complex brain topology and the interactions between different brain regions. At a macroscopic scale, brain networks describe how the brain's structure and function are organized, and the interconnections between different brain regions. Brain network connectivity is crucial for the study of many diseases. Studies have shown that in the functional connectivity networks of schizophrenic patients, the connection strength of certain edges is significantly lower than in normal controls. Research has also found that in schizophrenic patients, functional connections between two brain regions often exhibit greater physical distances than in normal subjects. These studies demonstrate that analyzing brain networks can provide a powerful tool for assisting in the diagnosis and treatment of brain diseases.

[0003] Existing methods for studying brain diseases using brain networks largely focus on population statistical analysis, primarily relying on significance tests to identify differentially characterized variables between groups. For example, Leroux et al. used analysis of covariance (ANCOVA) to study the functional and structural brain networks of schizophrenia patients, verifying the hypothesis that language networks in schizophrenia patients exhibit abnormalities. Tomson and his team extracted the weights of each connection from the brain network and used a two-sample binomial Wald test for significance analysis. Their results showed that the brain networks of patients with type I neurofibromatosis exhibited a reduced number of anterior-posterior connections and weakened bilateral connections.

[0004] However, due to the interaction of multiple factors, including genetic factors, environmental factors, biological diversity, disease progression, and individual differences in neuroanatomical structure and function, brain diseases often exhibit widespread heterogeneity in clinical practice. This heterogeneity manifests as diverse symptoms, pathological features, physiological changes, and clinical presentations of the same brain disease across different individuals and even within the same patient population. This heterogeneity poses challenges to treatment outcomes, health policies, and resource allocation, highlighting the urgent need for individualized research and precision medicine interventions. Currently, existing imaging techniques cannot accurately and reliably reflect the individual brain networks of a single patient. Furthermore, traditional case-control studies of brain networks are not specifically designed to identify individual differences between patients; their brain networks are primarily established at the group level, only capable of revealing differences between groups, and are insufficient to meet the needs of individualized clinical research. Summary of the Invention

[0005] Purpose of the invention: To address the above-mentioned shortcomings, this invention provides a more accurate feature classification method and system based on individual differences in brain networks.

[0006] Technical Solution: To solve the above problems, this invention employs a feature classification method based on individual-difference brain networks, comprising the following steps:

[0007] (1) Obtain data on normal brain regions and abnormal brain regions;

[0008] (2) Construct a reference brain network for normal brain region data based on normal brain region data, wherein the edges of the normal reference brain network are the correlation coefficients between pairs of normal brain region vectors; construct a reference brain network for abnormal brain region data based on abnormal brain region data, wherein the edges of the abnormal reference brain network are the correlation coefficients between pairs of abnormal brain region vectors.

[0009] (3) Add the data from the abnormal brain region data to the reference brain network of the normal data to obtain the perturbation brain network of each data in the abnormal brain region data; add the data from the normal brain region data to the reference brain network of the abnormal data to obtain the perturbation brain network of each data in the normal brain region data.

[0010] (4) Subtract the perturbation brain network of each data from the corresponding reference brain network to obtain the individual difference brain network of each data.

[0011] (5) The upper triangular elements of the individual difference brain network of the data are concatenated into a connection vector, and the connection vectors of all data are aggregated into a feature matrix;

[0012] (6) Use the feature selection algorithm to select features from the obtained feature matrix, select discriminant features, classify the selected discriminant features, and obtain the classification results.

[0013] Furthermore, the construction formula for the reference brain network is as follows:

[0014]

[0015] in, X represents the connection value between the i-th and j-th reference brain network. i and X j Let X represent the vectors corresponding to the i-th and j-th brain regions, respectively. i =(x i,1 ,...,x i,l ,...,x i,m ), brain region vector X j =(x j,1 ,...,x j,l ,...,x j,m ), where m represents the number of data points. and Representing brain region vectors X i and brain region vector X j The sample mean.

[0016] Furthermore, the formula for constructing the perturbation brain network is as follows:

[0017]

[0018] in, This represents the connection value between the i-th and j-th perturbation brain network.

[0019] Furthermore, the individual-differenced brain network is ΔFC. p ΔFC p =FC n+1 -FC n , among which, FC n+1 For the perturbation brain network corresponding to data p, FC n Let be the reference brain network corresponding to data p, m be the total number of data points in the reference brain network corresponding to data p, and the edge values ​​of the individual-difference brain network for data p be defined as the obtained Z values.

[0020]

[0021] Furthermore, the normal brain region data and abnormal brain region data are magnetic resonance imaging data from a publicly available data center. They are preprocessed to obtain preprocessed images of normal and abnormal brain regions. The obtained preprocessed images are then divided into brain regions, each divided into N brain regions, with each brain region corresponding to a node in the brain network. Parameters are extracted from each preprocessed image in the publicly available data center to construct a brain network. The mean gray matter volume of each brain region in each data set is calculated to obtain the brain region vector.

[0022] Furthermore, the preprocessed images are matched using standardized brain region maps, and the brain images are divided into N brain regions according to the standardized brain region maps, which include AAL maps, BNA maps, and Schaefer maps.

[0023] Furthermore, the feature selection algorithm includes a feature recursive elimination algorithm, and the classification of the selected discriminative features includes the use of support vector machine classification algorithm, random forest classification algorithm, and gradient boosting classification algorithm.

[0024] The present invention also provides a feature classification system based on individual difference networks, including a data acquisition module, a reference brain network construction module, a perturbation brain network construction module, an individual difference brain network construction module, and a classification module;

[0025] The data acquisition module is used to acquire data on normal brain regions and abnormal brain regions;

[0026] The reference brain network construction module is used to construct a reference brain network for normal data based on normal brain region data, wherein the edges of the normal reference brain network are the correlation coefficients between pairs of normal brain region vectors; and to construct a reference brain network for abnormal data based on abnormal brain region data, wherein the edges of the abnormal reference brain network are the correlation coefficients between pairs of abnormal brain region vectors.

[0027] The perturbation brain network construction module is used to add data from abnormal brain region data to the reference brain network of normal data to obtain perturbation brain networks for each data in abnormal brain region data; and to add data from normal brain region data to the reference brain network of abnormal data to obtain perturbation brain networks for each data in normal brain region data.

[0028] The individual-difference brain network construction module is used to subtract the perturbation brain network of each data from the corresponding reference brain network to obtain the individual-difference brain network of each data; to concatenate the upper triangular elements of the individual-difference brain network of the data into a connection vector; and to aggregate the connection vectors of all data into a feature matrix.

[0029] The classification module is used to select features from the obtained feature matrix using a feature selection algorithm, select discriminant features, classify the selected discriminant features, and obtain the classification result.

[0030] Beneficial Effects: Compared with existing technologies, the significant advantages of this invention are: by employing a perturbation difference method, it is possible to establish individualized brain networks for each abnormal and normal subject. Starting from the brain networks of multiple groups of individual differences, it is possible not only to focus on the individualized characteristics of each individual, but also to uncover the common differences between the individual and the reference group, thereby significantly improving the accuracy of abnormal brain region discrimination. Furthermore, experiments show that the embodiments of this invention have good robustness and can meet the clinical needs of personalized research. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the feature classification method of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the specific construction method of the individual difference brain network of the present invention. Detailed Implementation

[0033] like Figure 1 and Figure 2As shown in this embodiment, a feature classification method based on individual differences in brain networks is proposed. First, sMRI magnetic resonance imaging data is acquired from a public dataset. The acquired structural images are preprocessed using the SPM12 toolbox. In SPM12, voxel morphology (VBM) analysis is used to automatically perform operations such as examination, tissue segmentation, standardization, modulation, and smoothing to obtain the gray matter image of each subject's brain. This embodiment only uses schizophrenia (abnormal subjects) and normal subjects from the COBRE (The Center for Biomedical Research Excellence) dataset as examples for illustration.

[0034] A standardized brain region map (such as the AAL (Anatomical Automatic Labeling) map, BNA (Brainnetome Atlas) map, and Schaefer (Schaefer Atlas) map) was selected. Each preprocessed gray matter image of the subject's brain was matched against the map, and the brain image was divided into N brain regions according to the map, each corresponding one-to-one with N nodes in the constructed brain network. In this embodiment, the AAL map was selected to divide the brain into N = 90 brain regions, corresponding to 90 nodes in the brain network.

[0035] A reference brain network was established based on normal subjects.

[0036] In this step, the mean gray matter volume of each brain region is calculated for each normal subject, resulting in vector X corresponding to the i-th and j-th brain regions. i =(x i,1 ,...,x i,m ) and X j =(x j,1 ,...,x j,m ), where m represents the number of normal subjects. Calculate the Pearson correlation coefficients between all pairs of vectors:

[0037]

[0038] This constructs a reference brain network for m normal subjects, where the edges of the network represent the Pearson correlation coefficients between pairs of brain region vectors, i.e., the correlation coefficients of gray matter volume between pairs of brain regions (FC). m In this embodiment, the reference brain network is 90×90 in size and contains elements with values ​​of [0,1].

[0039] Each time a patient sample (abnormal sample) is added to the constructed reference brain network, the correlation is calculated, and a perturbation brain network for that patient (abnormal sample) is constructed.

[0040] Specifically, after adding a new patient (abnormal) sample k, the vector of the i-th brain region becomes X. i =(x i,1 ,...,x i,m ,x i,m+1 m+1 represents the number of subjects after adding sample k. For each new brain region vector, the Pearson correlation coefficient between each pair is calculated to obtain the perturbation brain network (FC) corresponding to the new sample k of the added patient (abnormal). m+1 ):

[0041]

[0042] In this embodiment, the perturbation brain network has a size of 90×90 and element values ​​of [0,1].

[0043] S3: Subtract the perturbation brain network of the patient (abnormal sample) from the reference brain network to obtain the individual-specific brain network of the patient (abnormal).

[0044] Specifically, ΔFC k =FC m+1 -FC m Patient (abnormal sample) k corresponds to the perturbation brain network (FC). m+1 ) and reference brain network (FC) m The difference result (ΔFC) k This measured the perturbation difference between patient (abnormal sample) k and the normal subject group;

[0045] Next, calculate ΔFC according to the following formula. k Z-value:

[0046]

[0047] The edge of the individual-discretionary brain network of patient (abnormal sample) k is defined as the obtained Z value, which reflects the change in gray matter volume correlation between paired brain regions of patient (abnormal sample) k relative to the normal subject group.

[0048] Obtain the individual-specific brain network for each patient (abnormal sample).

[0049] A reference brain network was established based on the patient (abnormal sample) group.

[0050] In this step, the mean gray matter volume of each brain region is calculated for each patient (abnormal sample), resulting in the vector Y corresponding to the i-th and j-th brain regions. i =(y i,1 ,...,y i,n ) and Y j =(y j,1 ,...,y j,n), where n represents the number of patients (abnormal samples). For all brain region vectors, calculate their Pearson correlation coefficients:

[0051]

[0052] A reference brain network was constructed for n patients (abnormal samples), where the edges of the brain network are the Pearson correlation coefficients between all pairwise brain region vectors, i.e., the gray matter volume correlation (FC) between paired brain regions. n In this embodiment, the reference brain network is 90×90 in size and has element values ​​of [0,1].

[0053] Each time a control sample (normal sample) is added to the constructed reference brain network, the correlation is calculated, and the perturbation brain network of the normal subject is constructed.

[0054] Specifically, after adding a new normal subject sample p, the brain region vector becomes Y. i =(y i,1 ,...,y i,n+1 n+1 represents the number of subjects after the addition of sample p. For each new brain region vector, the Pearson correlation coefficient between each pair is calculated to obtain the perturbed brain network (FC) corresponding to the new sample p added from the normal subjects. n+1 ):

[0055]

[0056] In this embodiment, the size of the perturbation brain network is 90×90, and the element value is [0,1].

[0057] The individual-differenced brain network of the normal subject was obtained by subtracting the perturbed brain network from the reference brain network of the patient (abnormal) group.

[0058] Specifically, ΔFC p =FC n+1 -FC n Normal subjects p corresponds to the perturbation brain network (FC) n+1 ) and reference brain network (FC) n The difference result (ΔFC) p The perturbation difference between the normal subjects (p) and the patients (abnormal) group was measured.

[0059] Next, calculate ΔFC according to the following formula. p Z-value:

[0060]

[0061] The edge definition of the individual difference brain network of normal subject p is the obtained Z value, which reflects the change in gray matter volume correlation between paired brain regions of normal subject p relative to the patient (abnormal) group.

[0062] We obtained the individual differences in the brain networks of each normal subject.

[0063] The upper triangle of the obtained individual-differenced brain network was used as the feature matrix for the patient (abnormal) group and the control (normal) group.

[0064] Specifically, the upper triangle of the individual-differenced brain network is extracted, arranged in rows, and pieced together to form a dimension of... The feature vector is N, which represents the number of regions into which the brain is divided by the prior map. In this embodiment, the dimension of the feature vector is 4005.

[0065] Next, the feature vectors of all normal subjects and patients (abnormal samples) are aggregated into a feature matrix, the dimension of which is the total number of subjects × the feature vector dimension.

[0066] Feature selection algorithms are used to select discriminative features from the obtained feature matrix. These discriminative features are then used for classification to obtain the classification results. In this embodiment, the feature selection employs a recursive feature elimination algorithm, and the diagnostic classification algorithm uses support vector machines, random forests, and gradient boosting algorithms. The classification results are validated using a 5-fold cross-validation method.

[0067] The method proposed in this embodiment was tested using the COBRE dataset (containing sMRI data of 93 schizophrenia patients (abnormal samples) and 82 normal subjects).

[0068] Tables 1-3 show the diagnostic performance of this embodiment on different datasets: COBRE, FBIRN (The Functional Biomedical Informatics Research Network), and ABIDE II (Autism Brain Imaging Data Exchange II), comparing the use of individual-difference brain networks with other feature construction methods for schizophrenia (abnormal data). The comparison methods are: directly using the gray matter volume values ​​of whole-brain voxels for feature dimensionality reduction and classification; and using the mean gray matter volume of brain regions segmented by prior atlases for feature selection and schizophrenia (abnormal data) diagnostic classification. Validation was also performed simultaneously on AAL, BNA, and Schaefer atlases.

[0069] Table 1. Comparison of diagnostic performance of different feature construction methods for schizophrenia (abnormal data) on the COBRE dataset.

[0070]

[0071] Table 2. Comparison of diagnostic performance of different feature construction methods for schizophrenia (abnormal data) on the FBIRN dataset.

[0072]

[0073] Table 3. Comparison of diagnostic performance of schizophrenia (abnormal data) using different feature construction methods in the ABIDEⅡ dataset.

[0074]

[0075] As shown in Tables 1-3, compared with other feature construction methods, the use of individual-difference brain networks achieved the best diagnostic performance for schizophrenia (abnormal data) and exhibited better robustness. This embodiment first establishes a reference brain network based on the normal subject group. Then, each time an abnormal sample is added, a perturbation brain network for that abnormal sample is constructed. By differencing the reference brain network, the individual-difference brain network for that abnormal sample is obtained. Next, a reference brain network is established based on the abnormal data, and then each time a normal subject sample is added, a perturbation brain network for that normal subject is constructed. By differencing, the individual-difference brain network for that normal subject is obtained. Unlike most current group-level brain networks, this embodiment uses a perturbation-difference method, which can establish individualized brain networks for each abnormal and normal subject. Starting from multi-group individual-difference brain networks, not only can individualized characteristics be considered, but common differences between individuals and the reference group can also be identified, thereby significantly improving the diagnostic accuracy of abnormal data.

Claims

1. A feature classification method based on individual-differenced brain networks, characterized in that, Includes the following steps: (1) Obtain data on normal brain regions and abnormal brain regions; (2) Construct a reference brain network for normal brain region data based on normal brain region data, wherein the edges of the normal reference brain network are the correlation coefficients between pairs of normal brain region vectors; construct a reference brain network for abnormal brain region data based on abnormal brain region data, wherein the edges of the abnormal reference brain network are the correlation coefficients between pairs of abnormal brain region vectors. (3) Add the data from the abnormal brain region data to the reference brain network of the normal data respectively to obtain the perturbation brain network of each data in the abnormal brain region data; By adding data from normal brain regions to the reference brain network of abnormal data, perturbation brain networks for each data point in normal brain regions are obtained. (4) Subtract the perturbation brain network of each data from the corresponding reference brain network to obtain the individual difference brain network of each data; (5) Concatenate the upper triangular elements of the individual difference brain network of the data into a connection vector, and aggregate the connection vectors of all data into a feature matrix; (6) Use the feature selection algorithm to select features from the obtained feature matrix, select discriminant features, classify the selected discriminant features, and obtain the classification results; The formula for constructing the reference brain network is as follows: in, Indicates the first The and the first A reference brain network connectivity value, and They represent the first The and the first The vector corresponding to each brain region, brain region vector brain region vectors , Indicates the number of data points. and Representing brain region vectors respectively and brain region vectors The sample mean; The formula for constructing the perturbation brain network is as follows: in, Indicates the first The and the first A perturbation of brain network connectivity values; The individual difference brain network is : ,in, For data The corresponding perturbation brain network, For data The corresponding reference brain network, For data The total number of data points in the corresponding reference brain network, data The edge definition of the individual differences in the brain network is obtained. value: 。 2. The feature classification method according to claim 1, characterized in that, The normal brain region data and abnormal brain region data are magnetic resonance imaging data from a public data center. They are preprocessed to obtain preprocessed images of normal brain region data and abnormal brain region data. The obtained preprocessed images are then divided into brain regions, each divided into N brain regions, and each brain region corresponds to a node in the brain network. Parameters were extracted from each preprocessed image in the public data center to construct a brain network; the mean gray matter volume of each brain region in each data was calculated to obtain the brain region vector.

3. The feature classification method according to claim 2, characterized in that, The preprocessed images are matched using a standardized brain region atlas, which divides the brain images into N brain regions. The standardized brain region atlas includes the AAL atlas, BNA atlas, and Schaefer atlas.

4. The feature classification method according to claim 2, characterized in that, The feature selection algorithm includes a feature recursive elimination algorithm, and the classification of the selected discriminative features includes the use of support vector machine classification algorithm, random forest classification algorithm, and gradient boosting classification algorithm.

5. A feature classification system according to the feature classification method as described in claim 1, characterized in that, It includes a data acquisition module, a reference brain network construction module, a perturbation brain network construction module, an individual difference brain network construction module, and a classification module; The data acquisition module is used to acquire data on normal brain regions and abnormal brain regions; The reference brain network construction module is used to construct a reference brain network for normal data based on normal brain region data, wherein the edges of the normal reference brain network are the correlation coefficients between pairs of normal brain region vectors; and to construct a reference brain network for abnormal data based on abnormal brain region data, wherein the edges of the abnormal reference brain network are the correlation coefficients between pairs of abnormal brain region vectors. The perturbation brain network construction module is used to add the data from the abnormal brain region data into the reference brain network of the normal data respectively, so as to obtain the perturbation brain network of each data in the abnormal brain region data. By adding data from normal brain regions to the reference brain network of abnormal data, perturbation brain networks for each data point in normal brain regions are obtained. The individual-difference brain network construction module is used to subtract the perturbation brain network of each data from the corresponding reference brain network to obtain the individual-difference brain network of each data; to concatenate the upper triangular elements of the individual-difference brain network of the data into a connection vector; and to aggregate the connection vectors of all data into a feature matrix. The classification module is used to select features from the obtained feature matrix using a feature selection algorithm, select discriminant features, classify the selected discriminant features, and obtain the classification result.

6. The feature classification system according to claim 5, characterized in that, The normal brain region data and abnormal brain region data are magnetic resonance imaging data from a public data center. They are preprocessed to obtain preprocessed images of normal brain region data and abnormal brain region data. The obtained preprocessed images are then divided into brain regions, each divided into N brain regions, and each brain region corresponds to a node in the brain network. Parameters were extracted from each preprocessed image in the public data center to construct a brain network; the mean gray matter volume of each brain region in each data was calculated to obtain the brain region vector.

7. The feature classification system according to claim 6, characterized in that, The preprocessed images are matched using a standardized brain region atlas, which divides the brain images into N brain regions. The standardized brain region atlas includes the AAL atlas, BNA atlas, and Schaefer atlas.

Images