Cognitive disorder auxiliary analysis method based on multi-view brain network feature fusion

Abstract

A cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion, belonging to the field of medical information processing technology, solves the technical problem that existing technologies struggle to fully exploit the correlations and complementarities between multi-level structural features of patients' brain functional connectivity data, leading to low accuracy in assessing patients' brain functional status. This method uses rs-fMRI as input data, extracts multi-level brain connectivity information by constructing functional connectivity networks and higher-order functional connectivity networks, and builds a multi-view local-global graph neural network model (MVLG-GNN) to achieve collaborative modeling, adaptive fusion, and auxiliary analysis combining group relationships of subjects from multiple views of brain connectivity features. This invention is also applicable to the field of Alzheimer's disease brain data processing.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to a method for assisting in the analysis of cognitive impairment based on multi-view brain network feature fusion. Background Technology

[0002] Alzheimer's disease (AD) is a progressive neurodegenerative disease characterized by cognitive impairment. Its clinical symptoms primarily include memory loss, cognitive decline, and behavioral abnormalities. The disease is characterized by its insidious onset, slow progression, and irreversible nature, significantly reducing patients' quality of life and placing additional pressure on family care and the social healthcare system.

[0003] Functional magnetic resonance imaging (fMRI) data, especially resting-state fMRI data, can reflect the state of neural activity in brain regions and the functional connectivity between brain regions, and has important application value in the auxiliary diagnosis of Alzheimer's disease. Current Alzheimer's disease auxiliary analysis methods based on fMRI data mostly rely on single-level functional connectivity information, making it difficult to fully explore the correlations and complementarities between the multi-level structural features of the patient's brain functional connectivity data, resulting in low accuracy in assessing the patient's brain functional state. Summary of the Invention

[0004] This invention addresses the technical problem that existing technologies struggle to fully exploit the correlations and complementarities among the multi-level structural features of patients' brain functional connectivity data, resulting in low accuracy in assessing patients' brain functional status.

[0005] The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion described in this invention includes the following steps: Step 1: Obtain rs-fMRI data of the subjects and preprocess them to obtain time series data; construct a multi-view dataset based on the time series data. Step 2: Based on the multi-view dataset, construct a functional connectivity network to represent the low-level functional connectivity matrix between brain regions; Step 3: Based on the functional connectivity network, construct a higher-order functional connectivity network to represent the higher-order functional connectivity matrix; Step 4: Construct a multi-view local feature extraction module, taking the low-level functional connectivity matrix and high-level brain connectivity relationship between brain regions as input, to generate the corresponding functional connectivity feature representation and high-order functional connectivity feature representation; Step 5: Adaptively fuse the functional connectivity feature representation and the higher-order functional connectivity feature representation to generate the brain network features of the subject. Step 6: Obtain sociodemographic characteristics and construct a group relationship graph module for the subjects. This module uses sociodemographic characteristics and the subjects' brain network characteristics as inputs to generate a group relationship graph. Step 7: Construct a global correlation analysis module. Input the brain network characteristics and group relationship diagram of the subjects into this module for training and evaluation to obtain a cognitive impairment auxiliary analysis model and complete the cognitive impairment auxiliary analysis.

[0006] Furthermore, in one embodiment of the present invention, the preprocessing in step 1 includes at least time point removal, time layer correction, head motion correction, structural image registration, structural image segmentation, normalization, smoothing, filtering and linear regression processing.

[0007] Furthermore, in one embodiment of the present invention, the construction of the functional connectivity network in step 2 specifically includes: The correlation between any two brain regions in the multi-view dataset is calculated using the Pearson correlation coefficient. A low-level functional connectivity matrix between brain regions is constructed, and the low-level functional connectivity matrix between brain regions is standardized to obtain the functional connectivity network.

[0008] Furthermore, in one embodiment of the present invention, the construction of a higher-order functional connectivity network in step 3 specifically includes: The connection vector of each brain region in the low-level functional connectivity matrix between brain regions is used as the connectivity representation feature of the corresponding brain region. The correlation between the connectivity representation features of any two brain regions is calculated to construct a high-order functional connectivity matrix and obtain a high-order functional connectivity network.

[0009] Furthermore, in one embodiment of the present invention, the multi-view local feature extraction module in step 4 specifically comprises: Individual brain network features are extracted from input data using a shared local graph neural network. Key nodes are selected during the feature extraction process using shared mutual information constraint pooling to obtain output features.

[0010] Furthermore, in one embodiment of the present invention, the shared local graph neural network includes a first graph convolutional layer, a second graph convolutional layer, and a third graph convolutional layer connected in sequence; The first convolutional layer is used to map the original feature dimension of each node in the input data to the first hidden layer dimension, extract the initial local connection features of the nodes, and perform nonlinear transformation using the ReLU activation function; The second graph convolutional layer is used to map the first hidden layer dimension to the second hidden layer dimension, extracting the local structural features of the nodes; The third convolutional layer is used to extract features from key nodes after they have been filtered by shared mutual information constraint pooling, and to obtain output features.

[0011] Furthermore, in one embodiment of the present invention, the shared mutual information constraint pooling specifically includes: The local structural features of the current node are encoded by graph convolution to obtain the local representation of the node. Positive and negative sample pairs are constructed. The mutual information between the local representation and the input features of the shared mutual information constraint pooling is estimated by a fully connected layer. Based on the mutual information, each node is scored for importance, and the top k nodes with the highest importance scores are selected as the key nodes after screening.

[0012] Furthermore, in one embodiment of the present invention, the subject group relationship diagram module in step 6 specifically comprises: By combining the brain network features of the subjects with age and gender information from sociodemographic features, phenotypic affinity relationships among the subjects are constructed, generating edge indices and edge input features of the group graph. The edge input features are then input into a fully connected network with shared parameters to extract the corresponding feature representations. Cosine similarity is used to calculate the connection weight of each edge, and a group relationship graph is generated based on the connection weight and edge index of each edge.

[0013] Furthermore, in one embodiment of the present invention, the global correlation analysis module in step 7 specifically comprises: Multi-layer graph convolutional layers are used to extract features from the input data. After each graph convolution, batch normalization and ReLU activation are performed sequentially. The final group graph feature representation is obtained by combining inter-layer residual connections and weighted fusion. The auxiliary analysis results are output through fully connected layers.

[0014] Furthermore, in one embodiment of the present invention, the feature extraction of the input data using multi-layer graph convolutional layers specifically includes: The first graph convolutional layer maps the feature dimensions of the input data, and subsequent graph convolutional layers keep the feature dimensions unchanged.

[0015] This invention addresses the technical problem in existing technologies where it is difficult to fully exploit the correlations and complementarities between the multi-level structural features of patients' brain functional connectivity data, resulting in low accuracy in assessing patients' brain functional status. Specific beneficial effects include: 1. This invention proposes a cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion. By simultaneously constructing functional connectivity networks and higher-order functional connectivity networks, it can not only describe low-level functional connectivity relationships between brain regions but also uncover higher-level brain connectivity pattern information. Based on this, it can achieve comprehensive capture of brain functional connectivity features related to Alzheimer's disease, thereby improving the model's ability to effectively represent abnormal brain functional patterns. 2. This invention proposes a cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion. In this method, the constructed MV-Local-GNN (multi-view local feature extraction module) synergistically combines a shared local graph neural network with a shared mutual information constraint pooling mechanism. On the one hand, this module utilizes shared multi-layer graph convolution operations to extract localized topological features from both the functional connectivity network and higher-order functional connectivity networks; on the other hand, by screening key brain regions, it strengthens the information representation of core brain regions while effectively suppressing the negative impact of redundant information. 3. This invention proposes a cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion. The proposed WGAFM (Adaptive Feature Aggregation Unit) learns the gating weight values ​​corresponding to the brain connectivity features of each view, and performs a weighted aggregation operation on functional connectivity features and higher-order functional connectivity features accordingly. This mechanism enables efficient and adaptive collaborative integration of information between multi-view features; 4. This invention proposes a cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion. The SRGM (Subject Group Relationship Graph Module) in this method takes sociodemographic features as input to generate a group relationship graph representing the relationships between subjects. Based on this, it is combined with the Global-GNN (Global Association Analysis Module) to jointly learn the interaction patterns between subjects. This combined strategy achieves the organic integration of feature information at different levels, significantly improving the ability to characterize complex brain functional abnormalities and dependencies between cross-level features. Ultimately, it improves the accuracy of Alzheimer's disease auxiliary identification on the ADNI dataset (Alzheimer's Disease Neuroimaging Dataset).

[0016] This invention is applicable to the fields of early auxiliary diagnosis of Alzheimer's disease and prediction of conversion to mild cognitive impairment. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of the cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion as described in Implementation Method 8; Figure 2This is a flowchart of the preprocessing described in Implementation Method 8; Figure 3 (a) is a visualization of the construction of the functional connection matrix as described in Implementation Method 8; Figure 3 (b) is a visualization of the Fisher-Z transform described in Implementation Method 8; Figure 3 (c) is a visualization of the construction of the high-order functional connectivity matrix as described in Implementation Method 8. Detailed Implementation

[0018] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] Implementation Method 1: Current techniques for auxiliary assessment of Alzheimer's disease using resting-state functional magnetic resonance imaging (fMRI) data largely limit the construction of brain functional networks to extracting functional connectivity features at a single level. This approach struggles to effectively capture complementary information and intrinsic connections between different levels of brain connectivity patterns, thus failing to comprehensively characterize the abnormal changes in brain networks caused by the disease. Specifically, existing technologies struggle to fully exploit the correlations and complementary relationships between the multi-level structural features of a patient's brain functional connectivity data, resulting in low accuracy in assessing the patient's brain functional state.

[0020] To address the aforementioned technical problems, this embodiment proposes a cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion, the method comprising the following steps: Step 1: Obtain rs-fMRI data and sociodemographic data from the Alzheimer's disease neuroimaging dataset ADNI, and preprocess the obtained rs-fMRI data to obtain time series data. Based on the time series data, construct the multi-view dataset required for the cognitive impairment auxiliary analysis model. The Alzheimer's disease neuroimaging dataset includes the ADNI dataset. Step 2: Divide the time series data obtained in Step 1 into a training set and a test set; Step 3: Based on the training and test sets defined in Step 2, construct a functional connectivity network to represent the low-order functional connectivity relationships between brain regions.

[0021] Step 4: Based on the functional connectivity network obtained in Step 3, a higher-order functional connectivity network is further constructed to represent the higher-level brain connectivity relationships formed by the lower-order functional connectivity patterns between brain regions.

[0022] Step 5: Based on the functional connectivity network constructed in Step 3 and the higher-order functional connectivity network constructed in Step 4, construct the MV-Local-GNN module to extract the local connectivity features of the individual brain network of the subject.

[0023] Step 6: Construct WGAFM. Based on the functional connectivity graph feature representation and high-order functional connectivity graph feature representation obtained in Step 5, perform adaptive weighted fusion to obtain the final brain network features of the subject.

[0024] In this embodiment, gating weights are learned through feature splicing, fully connected layers, and nonlinear activation functions, and fusion weights are generated using the Sigmoid function. The fusion weights are then used to adaptively weight and fuse the feature representations of the functional connectivity graph and the higher-order functional connectivity graph to obtain the final brain network features of the subject, thereby enhancing the complementarity of information between different views and improving feature representation ability and discrimination performance.

[0025] Step 7: Based on the sociodemographic features obtained in Step 1 and the fused postbrain network features obtained in Step 4, construct the subject group relationship graph module SRGM to establish the association between subjects and generate the subject group relationship graph and its connection weights for subsequent analysis.

[0026] Step 8: Construct the Global-GNN module, combine the brain network features of the subjects and the subject group relationship graph module SRGM, and complete the binary assessment of AD vs. MCI through a multi-layer graph convolution discriminator to obtain the cognitive impairment auxiliary analysis model (MVLG-GNN model), thus completing the construction of the cognitive impairment auxiliary analysis model.

[0027] Step 9: Train and validate the MVLG-GNN model on the Alzheimer's disease neuroimaging dataset to obtain the trained cognitive impairment auxiliary analysis model.

[0028] In this embodiment, the time series data obtained in step 1 is divided into a training set and a test set. To process the sample data of the ADNI dataset using ten-fold cross-validation, the ratio of the training set to the test set is 9:1.

[0029] This implementation uses an Alzheimer's disease neuroimaging dataset as input data. It acquires multi-view brain network features of subjects by constructing functional connectivity networks and higher-order functional connectivity networks. It extracts local connectivity information using MV-Local-GNN, combines WGAFM to achieve adaptive weighted fusion of dual-view features, establishes group relationships among subjects through SRGM, and finally uses Global-GNN to complete the auxiliary diagnosis of AD and MCI, thereby improving the accuracy of the model's auxiliary diagnosis.

[0030] Implementation Method 2: The difference between this implementation method and Implementation Method 1 is that the preprocessing in step 1 includes converting the data format before preprocessing to NIfTI format (Technical Initiative Format), removing the first 5 time points, time-layer correction, head motion correction, registration of fMRI and T1-weighted magnetic resonance imaging, structural image segmentation, spatial normalization to MNI152, spatial smoothing with a full width at half maximum (FWHM) of 5 mm, filtering, and linear regression processing to obtain 135×116 time-series data based on AAL (Automatic Anatomical Marker) atlas.

[0031] Implementation Method 3: The difference between this implementation method and Implementation Method 1 is that the functional connection network is constructed in step 2, specifically as follows: The preprocessed 135×116 time series data is input into the functional connectivity network construction process, where rows represent time points and columns represent brain region signals. Using 116 brain regions as nodes, the time series of each brain region is extracted, and the correlation between any two brain regions is calculated using the Pearson correlation coefficient. This correlation is used as the functional connectivity strength, thereby constructing a 116×116 functional connectivity matrix and generating the functional connectivity network of the subjects.

[0032] Implementation Method Four: The difference between this implementation method and Implementation Method One is that in step 3, a high-order functional connection network is constructed, specifically as follows: Based on the functional connectivity matrix, the connectivity patterns of each brain region are extracted, the correlation between any two brain region connectivity patterns is calculated, a 116×116 high-order functional connectivity matrix is ​​constructed, and a high-order functional connectivity network is generated.

[0033] Implementation Method 5: The difference between this implementation method and Implementation Method 1 is that the multi-view local feature extraction module in step 4 is specifically as follows: The functional connectivity network and the higher-order functional connectivity network are respectively input into the local feature extraction module with shared parameters. Local features are extracted through graph convolution, and the key node selection, mutual information constraint and subgraph structure update are completed by combining the SMICP (Shared Mutual Information Constrained Pooling) module. Finally, the local graph embedding representations corresponding to the two views are obtained, in which SMICP enhances the expression of key brain region information.

[0034] The shared local graph neural network includes a first graph convolutional layer, a second graph convolutional layer, and a third graph convolutional layer connected in sequence; The first convolutional layer is used to map the original feature dimension of each node in the input data to the first hidden layer dimension, extract the initial local connection features of the nodes, and perform nonlinear transformation using the ReLU activation function; The second graph convolutional layer is used to map the first hidden layer dimension to the second hidden layer dimension, extracting the local structural features of the nodes; The third convolutional layer is used to extract features from key nodes after they have been filtered by shared mutual information constraint pooling, and to obtain output features.

[0035] The shared mutual information constraint pooling is specifically as follows: The local structural features of the current node are encoded by graph convolution to obtain the local representation of the node. Positive and negative sample pairs are constructed. The mutual information between the local representation and the input features of the shared mutual information constraint pooling is estimated by a fully connected layer. Based on the mutual information, each node is scored for importance, and the top k nodes with the highest importance scores are selected as the key nodes after screening.

[0036] In this embodiment, the MV-Local-GNN consists of a shared local graph neural network and a shared mutual information constraint pooling, used to receive functional connectivity networks and higher-order functional connectivity networks, respectively. The shared local graph neural network uses a shared three-layer GCN convolution (graph convolutional layer) to extract features from both types of brain connectivity networks, learning the local connectivity features of the individual subject's brain network. The shared mutual information constraint pooling retains key brain region features by scoring and filtering brain region nodes, and then fuses pooling features with subsequent graph convolutional features through skip connections to generate corresponding functional connectivity graph feature representations and higher-order functional connectivity graph feature representations.

[0037] Implementation Method Six: The difference between this implementation method and Implementation Method One is that the subject group relationship diagram module in step 6 is specifically as follows: By combining the brain network features of the subjects with age and gender information from sociodemographic features, phenotypic affinity relationships among the subjects are constructed, generating edge indices and edge input features of the group graph. The edge input features are then input into a fully connected network with shared parameters to extract the corresponding feature representations. Cosine similarity is used to calculate the connection weight of each edge, and a group relationship graph is generated based on the connection weight and edge index of each edge.

[0038] In this embodiment, a group relationship graph of the subjects is constructed based on the subjects' brain network characteristics and sociodemographic characteristics; the subjects' brain network characteristics are used as node features, and the connection relationships between the subjects are determined by sociodemographic characteristics and feature similarity. The corresponding connection weights are obtained through the edge weight learning module to form a weighted group relationship graph of the subjects.

[0039] Implementation Method Seven: The difference between this implementation method and Implementation Method One is that the global correlation analysis module in step 7 is specifically as follows: Multi-layer graph convolutional layers are used to extract features from the input data. After each graph convolution, batch normalization and ReLU (rectified linear function) activation are performed sequentially. The final group graph feature representation is obtained by combining inter-layer residual connections and weighted fusion. The auxiliary analysis results are output through fully connected layers.

[0040] The feature extraction of the input data using multi-layer graph convolutional layers is specifically as follows: The first graph convolutional layer maps the feature dimensions of the input data, and subsequent graph convolutional layers keep the feature dimensions unchanged.

[0041] In this embodiment, based on the weighted subject group relationship graph, a global graph neural network is used to aggregate and update the node neighborhood information through multi-layer graph convolution to learn the association patterns between subjects and the structural information at the group level, thereby completing the auxiliary analysis of each subject sample and outputting the corresponding auxiliary analysis results.

[0042] Implementation Method Eight: This implementation method is a specific embodiment based on the methods described in Implementation Methods One to Seven.

[0043] like Figure 1 As shown, this embodiment provides a cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion, and the specific steps are as follows: Step 1: Obtain rs-fMRI data and sociodemographic data from the Alzheimer's disease neuroimaging dataset ADNI, perform preprocessing, and construct a multi-view dataset.

[0044] Step 1.1: Download rs-fMRI data of AD patients and MCI (mild cognitive impairment) subjects, as well as corresponding clinical scale data such as age and gender, from the ADNI dataset.

[0045] Step 1.2, as follows Figure 2 As shown, rs-fMRI data preprocessing includes the following steps: Step 1.2.1: Preprocessing Part 1: Use the SPM12 (Statistical Parameter Mapping Version 12) toolbox to convert DICOM format (Medical Digital Imaging and Communication Standard) to NIfTI format; Step 1.2.2, the second part of preprocessing, involves removing the first 5 time points. The first 5 time points are manually removed from the 140 time points to reduce the impact of signal fluctuations in the initial scanning stage on data quality and ensure that the acquired functional magnetic resonance time series has good stability. Step 1.2.3, Preprocessing Part 3: Using the CONN toolbox (Functional Connection Toolbox) for time-level correction and head motion correction. Time-level correction is used to eliminate the influence of differences in acquisition time at different levels, while head motion correction is used to reduce artifact interference caused by head movements of the subject. Step 1.2.4, Preprocessing Part 4: Use the CONN toolbox to perform fMRI and T1-weighted magnetic resonance imaging registration to improve the spatial consistency between functional and structural data. Step 1.2.5, the fifth part of preprocessing, uses the CONN toolbox to segment the structural image, obtaining tissue components such as gray matter, white matter, and cerebrospinal fluid, which provides a basis for subsequent spatial normalization and noise regression processing; Step 1.2.6, Preprocessing Part Six: Spatial Normalization. The data is registered to the MNI152 standard template (Montreal Neuroscience Institute 152 standard template) using the CONN toolbox, and unified to the same spatial coordinates. Step 1.2.7, Preprocessing Part 7: Spatial Smoothing. The data is spatially smoothed using a Gaussian kernel with a half-height and full width of 5mm using the CONN toolbox to reduce noise interference and improve the spatial continuity of the signal. Step 1.2.8, Preprocessing Part 8, uses the CONN toolbox to filter and perform linear regression on the data to reduce the impact of noise and interference on the functional magnetic resonance time series. Step 2: Based on the functional connectivity matrix data obtained after calculating the Pearson correlation coefficient in Step 1, divide the ADNI dataset into 10-fold cross-validation sets. Use 9 datasets as the training set (90%) and 1 dataset as the validation set (10%). Step 3: Construct a functional connectivity network based on the training and test sets divided in Step 2 to represent low-order functional connectivity relationships between brain regions. This step includes the following: Step 3.1: Import sub-units to build the basic libraries required for the functional connectivity network, including os (operating system interface module), numpy (numerical computing library), scipy.io (input / output sub-module of scientific computing toolkit), torch (core library of PyTorch framework), torch_geometric.data.Data (data structure of graph geometry deep learning library), and networkx (network analysis library). Also, import graph processing sub-modules such as remove_self_loops (remove self-loop function) and coalesce (edge ​​index merging and deduplication function) for subsequent connection matrix reading, undirected graph structure construction, and edge information normalization processing.

[0046] Step 3.2 defines a functional connectivity preprocessing module, used to read the data file corresponding to each subject according to the subject number and load the functional connectivity matrix, such as... Figure 3 As shown in (a): Applying the Fisher-Z transform (hyperbolic tangent inverse transform) to the functional connectivity matrix can make the distribution of connectivity values ​​more stable, such as... Figure 3 As shown in (b). During processing, outliers and values ​​that are too large or too small in the matrix are first processed, then the inverse hyperbolic tangent function is used to complete the transformation, and finally the result is symmetricized and the diagonal is set to zero.

[0047] Step 3.3: Define a functional connectivity graph construction module to construct an undirected weighted brain functional connectivity graph from the functional connectivity matrix after Fisher-Z transformation. Each brain region is considered a node, and the integer row vector corresponding to that brain region in the functional connectivity matrix zfc (the matrix obtained by Fisher-Z transformation of the original FC matrix) is used as the node feature. The zfc matrix is ​​a 116×116 matrix, and the i-th row zfc[i,:] is used to represent the functional connectivity pattern between the i-th brain region and all other brain regions. Since the zfc matrix satisfies zfc(i,j)=zfc(j,i) after symmetry processing, an undirected graph is constructed based on this symmetric matrix using networkx.from_numpy_array. The absolute values ​​of the connectivity matrix elements are used to represent the connectivity strength between brain regions and are used as edge weights and edge attributes to input the graph structure. Then, graph processing submodules such as remove_self_loops and coalesce are used to perform self-loop removal, edge index merging, and duplicate edge normalization, finally obtaining undirected weighted brain functional connectivity graph data that can be used as input to graph neural networks.

[0048] Step 4: Based on the functional connectivity network obtained in Step 3, further construct a higher-order functional connectivity network to represent higher-level functional connectivity relationships between brain regions. This step includes the following: Step 4.1: Import sub-units, which are the basic libraries required to build high-order functional connectivity networks, including os, numpy, scipy.io, torch, torch_geometric.data.Data, and networkx. They also import graph processing sub-modules such as remove_self_loops and coalesce for subsequent generation of high-order functional connectivity matrices, construction of undirected graph structures, and normalization of edge information.

[0049] Step 4.2 defines a higher-order functional connectivity preprocessing module to construct a higher-order functional connectivity matrix based on the functional connectivity matrix obtained in Step 3. Taking the functional connectivity matrix after Fisher-Z transform as input, the entire row connectivity vector zfc[i, :] of brain region i and the entire row connectivity vector zfc[j, :] of brain region j are taken, representing the functional connectivity patterns between the corresponding brain region and all other brain regions, respectively. The correlation between these two connectivity vectors is calculated to characterize the similarity between brain regions i and j in their overall connectivity patterns, thus generating the higher-order functional connectivity matrix. Subsequently, outlier handling, symmetry conversion, and setting the diagonal to 1 are performed on the result to ensure the stability of the matrix, such as... Figure 3 As shown in (c).

[0050] Step 4.3: Define a higher-order functional connectivity graph (HFM) construction module to convert the HFM matrix into undirected weighted graph structure data that can be processed by graph neural networks. Each brain region is considered a node, and the integer row vector corresponding to that brain region in the HFM matrix is ​​used as the node feature to represent the higher-order connection pattern between that brain region and all other brain regions. After symmetry processing, the HFM matrix satisfies hfc(i,j)=hfc(j,i). An undirected graph is constructed based on this symmetric matrix using networkx.from_numpy_array. The absolute values ​​of the HFM matrix elements represent the higher-order connection strength between brain regions and are used as edge weights and edge attributes input to the graph structure. Then, graph processing submodules such as remove_self_loops and coalesce are used to perform self-loop removal, edge index merging, and duplicate edge normalization on the graph structure, ultimately obtaining undirected weighted HFM data that can be used as input to a graph neural network.

[0051] Step 5: Based on the functional connectivity network constructed in Step 3 and the higher-order functional connectivity network constructed in Step 4, construct the MV-Local-GNN module to extract local connectivity features of the individual brain network of the subject, where SMICP is used to enhance the expression of information in key brain regions.

[0052] The steps to build this module are as follows: Step 5.1: Import sub-units to build the basic libraries required for MV-Local-GNN, including torch, torch.nn, and torch.nn.functional, and introduce graph processing sub-modules such as GCNConv, topk, and filter_adj; at the same time, introduce the custom SMICP module for subsequent key node selection, subgraph update, and local feature enhancement.

[0053] Step 5.2: Define the input graph representation, treating each brain region as a node, using its connection patterns with other brain regions as node features, and combining edge index and edge weight information as input to the graph convolutional network.

[0054] Step 5.3: Define the first GCNConv (graph convolutional layer) operation to extract initial local connectivity features: The original 116-dimensional features of each node are mapped to a 64-dimensional representation using GCNConv(116, 64). The initial local connectivity features of the nodes are extracted by aggregating information from neighboring nodes, and the nonlinear expressive power of the features is further enhanced by the ReLU activation function.

[0055] Step 5.4: Define the second-level GCNConv operation to further refine the local join pattern: GCNConv(64, 20) is used to map 64-dimensional node features to a 20-dimensional representation, and local structural features are further refined by aggregating neighbor information.

[0056] Step 5.5 defines the SMICP submodule, which is used to select key nodes and enhance feature representation during the local graph feature learning process: This module first encodes node features through graph convolution, then constructs positive and negative sample pairs, and estimates the mutual information between the local representation and the original features through a fully connected layer. At the same time, it sorts all nodes using a node scoring mechanism and retains nodes with higher scores through a Top-k strategy (a strategy that retains the top k best items), thus obtaining a local sub-graph that better reflects the information of key brain regions.

[0057] Step 5.6: Define the third-layer GCNConv operation to further extract features from the local sub-images filtered by SMICP. GCNConv(20, 20) is used to further aggregate neighborhood information on the pooled subgraph, thereby strengthening the local discriminative features of key nodes.

[0058] Step 5.7: In the forward propagation, the functional connectivity graph and the higher-order functional connectivity graph are input into the graph convolutional network. The node features are mapped from 116 dimensions to 64 dimensions and 20 dimensions respectively, resulting in node representations of 116×64 and 116×20. After SMICP filtering, 105×20 sub-graph features are obtained, and then an updated 105×20 feature is obtained through graph convolution. Finally, the features are fused and expanded into a 1×2100 subject-level representation, which is used as the input for subsequent multi-view aggregation units.

[0059] Step 6: Based on the functional connectivity graph feature representation and higher-order functional connectivity graph feature representation obtained in Step 5, construct a weighted guided adaptive feature aggregation unit (WGAFM) to perform weighted fusion of dual-view local brain network features, obtaining the final brain network representation of the subject. The construction steps of this module are as follows: Step 6.1: Import sub-units, which are used to build the basic libraries required for WGAFM, including torch and torch.nn.

[0060] Step 6.2 defines the WGAFM submodule, which is used to jointly model the functional connectivity graph feature x_fc and the higher-order functional connectivity graph feature x_hfc. The size of both x_fc and x_hfc is 1×2^100.

[0061] Step 6.3: Define the gating weight generation process to represent the relative importance of the two view features in different dimensions: concatenate x_fc and x_hfc in the feature dimension to obtain a joint feature representation of size 1×4200, and generate a gating weight g of size 1×2100 through linear mapping and nonlinear activation.

[0062] Step 6.4: Define the weighted fusion process: In the forward propagation, gating weights g and 1-g are used to weight x_fc and x_hfc dimension by dimension to obtain a fused subject-level brain network representation of size 1×2100.

[0063] Step 7: Construct the SRGM based on the sociodemographic features obtained in Step 1 and the fused postbrain network features obtained in Step 6, which is used to establish associations among subjects. The construction steps of this module are as follows: Step 7.1: Import sub-units to build the basic libraries required for SRGM, including numpy and scipy.spatial.distance, and import the custom EDGE module (edge ​​weight learning module).

[0064] Step 7.2: Define the subject relationship graph construction process, used to establish group graph connections based on phenotypic information and brain network feature similarities among subjects: The fused embeddings (brain network features of the subjects) and nonimg (sociodemographic features) are used as inputs. Phenotypic affinity is constructed by combining age and gender information, and the edge index edge_index and edge input edge_input of the group graph are further generated.

[0065] Step 7.3 defines the EDGE submodule, which is used to calculate the connection weights of each edge in the group graph.

[0066] This module inputs the phenotypic features of the subjects at both ends of the edge into a fully connected network with shared parameters, extracts the corresponding feature representations, and calculates the edge weights using cosine similarity.

[0067] Step 7.4: In the forward propagation, input edge_input into the EDGE module to obtain the corresponding edge weight edge_weight, which together with edge_index forms the subject group relationship graph.

[0068] Step 8: Construct the Global-GNN module, combining the fused subject brain network features obtained in Step 6 and the subject group relationship graph constructed in Step 7, to complete the binary assessment of AD vs. MCI. The construction steps of this module are as follows: Step 8.1: Import sub-units to build the basic libraries required for Global-GNN, including torch, torch.nn, and torch.nn.functional, and import sub-modules such as ChebConv (multi-layer graph convolutional layer) and BatchNorm1d (batch normalization).

[0069] Step 8.2 defines the ChebConv submodule, which is used to extract graph convolutional features from the group relationship graph of the subjects: the first layer uses ChebConv(2100, 20) to map the fused subject features from 2100 dimensions to 20 dimensions; the remaining three layers use ChebConv(20, 20) to further extract high-level structural features from the group graph.

[0070] Step 8.3 defines the BatchNorm1d submodule, which is used to normalize the features after each layer of graph convolution and combine it with the ReLU activation function to enhance the feature representation capability.

[0071] Step 8.4 defines interlayer residual connections and weighted fusion operations to fuse group graph features from different layers: the features of the next layer are residually superimposed with the features of the previous layer, and the outputs of each layer are then weighted and fused using learnable weights.

[0072] Step 8.5: In the forward propagation, the fused feature embeddings obtained in step 6 and the edge_index and edge_weight of the population graph obtained in step 7 are input into the Global-GNN module. They are then processed through four layers of graph convolution, normalization, and activation operations. The final population graph feature representation is obtained by combining the inter-layer residual connections and weighted fusion. The auxiliary analysis results of size N×2 are output through the fully connected layer to complete the binary evaluation prediction of the subjects.

[0073] Step 9: Train and validate the MVLG-GNN model on the Alzheimer's disease neuroimaging dataset to obtain a cognitive impairment auxiliary analysis model. This step includes the following: Step 9.1: Import sub-units, which are the basic libraries required for model training and validation, including torch, torch.nn, numpy, and data loading modules, and introduce the MVLG-GNN class as the core model for training.

[0074] Step 9.2: Define the data loading and partitioning process for reading functional connectivity graphs, higher-order functional connectivity graphs, label information, and sociodemographic features: Load the multi-view graph data obtained in Steps 3 and 4 together with the labels, and partition the training set and test set using hierarchical cross-validation.

[0075] Step 9.3: Define the model initialization process for building the MVLG-GNN network: Use the sociodemographic feature nonimg and the phenotypic information phonetic_score as model input parameters to complete the overall construction of the local graph feature extraction, multi-view feature fusion, subject relationship graph construction, and group graph evaluation module.

[0076] Step 9.4: Define the loss function and optimization process for updating model parameters: During training, input the functional connectivity graph and higher-order functional connectivity graph into the MVLG-GNN model, and output the auxiliary analysis results predictions and mutual information constraint term mi_loss; based on this, construct the total loss function by combining the discriminant loss and mutual information loss, and update the model parameters through backpropagation.

[0077] Step 9.5: Define the model training and validation process: In each round of training, the model is forward propagated and parameters are updated using training set samples; in the validation phase, the auxiliary analysis results of the model are calculated using test set samples, and metrics such as accuracy, sensitivity, specificity, and AUC (area under the curve) are statistically analyzed to evaluate model performance.

[0078] Step 9.6: Define the model saving and result output process: When the model achieves better auxiliary analysis results on the validation set, save the corresponding model parameters; finally, obtain the cognitive impairment auxiliary analysis model for AD vs. MCI dichotomous assessment.

[0079] The technical solution and its effects of this embodiment will be further explained below with reference to specific experimental data.

[0080] To verify the effectiveness of each component module in the proposed MVLG-GNN model of this embodiment, ablation experiments were conducted on the Alzheimer's disease neuroimaging dataset ADNI. The ablation experiments included the following models: a model using only FC (functional connectivity matrix) as input, a model using only HFC (higher-order functional connectivity matrix) as input, an FC+HFC+Fixed Fusion model that fuses functional connectivity and higher-order functional connectivity matrices using a feature aggregation unit based on fixed weights, and the FC+HFC+WGAFM model proposed in this embodiment that uses an adaptive feature aggregation unit based on weighted guidance. The experimental results are shown in Table 1. The FC+HFC+WGAFM model based on a local-global structure proposed in this embodiment achieved an accuracy of 86.43% in the ADvs.MCI assisted diagnosis task, and its auxiliary analysis performance was superior to the above ablation models, indicating the effectiveness of the joint modeling method of functional connectivity and higher-order functional connectivity adopted in this embodiment. It also verified the effectiveness and superiority of the WGAFM module in multi-view feature fusion.

[0081] Table 1 Ablation Experiment

[0082] Implementation Method Nine: This implementation method provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in any one of Implementation Methods One to Seven.

[0083] Implementation Method 10: This implementation method provides a computer device, including a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the method described in any one of Implementation Methods 1 to 7.

[0084] The above provides a detailed description of the cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion, characterized in that, Includes the following steps: Step 1: Obtain rs-fMRI data of the subjects and preprocess them to obtain time series data; construct a multi-view dataset based on the time series data. Step 2: Based on the multi-view dataset, construct a functional connectivity network to represent the low-level functional connectivity matrix between brain regions; Step 3: Based on the functional connectivity network, construct a higher-order functional connectivity network to represent the higher-order functional connectivity matrix; Step 4: Construct a multi-view local feature extraction module, taking the low-level functional connectivity matrix and high-level brain connectivity relationship between brain regions as input, to generate the corresponding functional connectivity feature representation and high-order functional connectivity feature representation; Step 5: Adaptively fuse the functional connectivity feature representation and the higher-order functional connectivity feature representation to generate the brain network features of the subject. Step 6: Obtain sociodemographic characteristics and construct a group relationship graph module for the subjects. This module uses sociodemographic characteristics and the subjects' brain network characteristics as inputs to generate a group relationship graph. Step 7: Construct a global correlation analysis module. Input the brain network characteristics and group relationship diagram of the subjects into this module for training and evaluation to obtain a cognitive impairment auxiliary analysis model and complete the cognitive impairment auxiliary analysis.

2. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 1, characterized in that, The preprocessing in step 1 includes at least time point removal, time layer correction, head motion correction, structural image registration, structural image segmentation, normalization, smoothing, filtering, and linear regression processing.

3. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 1, characterized in that, Step 2, which involves constructing a functional connectivity network, specifically includes: The correlation between any two brain regions in the multi-view dataset is calculated using the Pearson correlation coefficient. A low-level functional connectivity matrix between brain regions is constructed, and the low-level functional connectivity matrix between brain regions is standardized to obtain the functional connectivity network.

4. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 1, characterized in that, Step 3, which involves constructing a high-order functional connectivity network, specifically includes: The connection vector of each brain region in the low-level functional connectivity matrix between brain regions is used as the connectivity representation feature of the corresponding brain region. The correlation between the connectivity representation features of any two brain regions is calculated to construct a high-order functional connectivity matrix and obtain a high-order functional connectivity network.

5. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 1, characterized in that, The multi-view local feature extraction module in step 4 specifically includes: Individual brain network features are extracted from input data using a shared local graph neural network. Key nodes are selected during the feature extraction process using shared mutual information constraint pooling to obtain output features.

6. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 5, characterized in that, The shared local graph neural network includes a first graph convolutional layer, a second graph convolutional layer, and a third graph convolutional layer connected in sequence; The first convolutional layer is used to map the original feature dimension of each node in the input data to the first hidden layer dimension, extract the initial local connection features of the nodes, and perform nonlinear transformation using the ReLU activation function; The second graph convolutional layer is used to map the first hidden layer dimension to the second hidden layer dimension, extracting the local structural features of the nodes; The third convolutional layer is used to extract features from key nodes after they have been filtered by shared mutual information constraint pooling, and to obtain output features.

7. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 6, characterized in that, The shared mutual information constraint pooling is specifically as follows: The local structural features of the current node are encoded by graph convolution to obtain the local representation of the node. Positive and negative sample pairs are constructed. The mutual information between the local representation and the input features of the shared mutual information constraint pooling is estimated by a fully connected layer. Based on the mutual information, each node is scored for importance, and the top k nodes with the highest importance scores are selected as the key nodes after screening.

8. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 1, characterized in that, The subject group relationship diagram module in step 6 is specifically as follows: By combining the brain network features of the subjects with age and gender information from sociodemographic features, phenotypic affinity relationships among the subjects are constructed, generating edge indices and edge input features of the group graph. The edge input features are then input into a fully connected network with shared parameters to extract the corresponding feature representations. Cosine similarity is used to calculate the connection weight of each edge, and a group relationship graph is generated based on the connection weight and edge index of each edge.

9. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 1, characterized in that, The global correlation analysis module in step 7 specifically includes: Multi-layer graph convolutional layers are used to extract features from the input data. After each graph convolution, batch normalization and ReLU activation are performed sequentially. The final group graph feature representation is obtained by combining inter-layer residual connections and weighted fusion. The auxiliary analysis results are output through fully connected layers.

10. The cognitive impairment auxiliary analysis method based on multi-view brain network feature fusion according to claim 9, characterized in that, The feature extraction of the input data using multi-layer graph convolutional layers is specifically as follows: The first graph convolutional layer maps the feature dimensions of the input data, and subsequent graph convolutional layers keep the feature dimensions unchanged.

Images