Image classification method and device, computer device and storage medium

Abstract

The application relates to an image classification method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: obtaining a target brain functional connectivity graph of a measured object and a reference brain functional connectivity graph of a preset category, inputting the target brain functional connectivity graph and the reference brain functional connectivity graph into a space-based graph neural network, obtaining a similarity value between the target brain functional connectivity graph and the reference brain functional connectivity graph output by the graph neural network, and then determining the category of the target brain functional connectivity graph according to the similarity value. The space-based graph neural network has the advantage of accurately extracting structural attribute features, can accurately extract structural attribute features in the target brain functional connectivity graph and the reference brain functional connectivity graph, and accurately determine the similarity value between the target brain functional connectivity graph and the reference brain functional connectivity graph, thereby improving the accuracy of the determined category of the target brain functional connectivity graph, and further improving the accuracy of image classification.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to an image classification method, apparatus, computer device, and storage medium. Background Technology

[0002] Mild cognitive impairment (MCI) is an intermediate state between normal aging and dementia, representing a symptomatic pre-dementia stage with a high risk of developing Alzheimer's disease (AD). Therefore, identifying early MCI to correctly distinguish between normal aging individuals and early-stage MCI patients helps in providing early intervention and thus reducing the risk of MCI progressing to AD.

[0003] With the development of neuroimaging and artificial intelligence technologies, many machine learning algorithms based on magnetic resonance imaging (MRI) technology have been applied to the identification and classification of various diseases. Among them, thanks to the unique advantages of graph neural networks (GNNs) in brain network data analysis, related technologies use spectrum-based GNNs to classify patients' brain function maps and determine the patients' brain function diseases based on the classification results.

[0004] However, the spectrum-based GNNs used in related technologies cannot make good use of graph structure information and are not well applied to graph data with unique structural properties, such as the functional connectivity networks of brains in early-stage MCI patients, which reduces the accuracy of classifying such images. Summary of the Invention

[0005] Therefore, it is necessary to provide an image classification method, apparatus, computer device, computer-readable storage medium, and computer program product to address the aforementioned technical problems.

[0006] Firstly, this application provides an image classification method, including:

[0007] Obtain the target brain functional connectivity map and the reference brain functional connectivity map of the preset categories for the tested subjects;

[0008] The target brain function connectivity map and the reference brain function connectivity map are input into a spatial graph neural network to obtain the similarity value between the target brain function connectivity map and the reference brain function connectivity map output by the graph neural network.

[0009] The category of the target brain functional connectivity map is determined based on the similarity value.

[0010] In one embodiment, obtaining a target brain functional connectivity map of the subject includes:

[0011] Obtain brain blood oxygenation level-dependent signals from the tested subjects;

[0012] Based on brain functional area templates and brain blood oxygen level dependent signals, the brain spatial domain of the tested subjects was divided into regions to determine the blood oxygen level dependent signals of each brain functional area.

[0013] Based on the blood oxygenation level-dependent signals of each brain functional area, a target brain functional connectivity map is constructed.

[0014] In one embodiment, a target brain functional connectivity map is constructed based on blood oxygenation level-dependent signals from each brain functional region, including:

[0015] Using each brain functional area as a node, the blood oxygen level dependent signal of each node is determined based on the blood oxygen level dependent signal of each brain functional area.

[0016] Based on the blood oxygen level-dependent signals of each node, determine the correlation coefficient between any two nodes;

[0017] Determine whether there is a connecting edge between any two nodes based on the correlation coefficient;

[0018] A target brain function connectivity map is formed based on nodes and connecting edges.

[0019] In one embodiment, obtaining a similarity value between the target brain functional connectivity map and a reference brain functional connectivity map includes:

[0020] Extract the feature vectors of each node in the target brain functional connectivity graph; where the nodes in the target brain functional connectivity graph correspond to brain functional areas;

[0021] Extract the feature vectors of each node in the reference brain functional connectivity map;

[0022] Based on the feature vectors of each node in the target brain functional connectivity map and the feature vectors of each node in the reference brain functional connectivity map, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined.

[0023] In one embodiment, determining the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the feature vectors of each node in the target brain functional connectivity map and the feature vectors of each node in the reference brain functional connectivity map includes:

[0024] The target brain functional connectivity map is divided into multiple target sub-maps, and the reference brain functional connectivity map is divided into multiple reference sub-maps;

[0025] Based on the feature vectors of each node in the target brain functional connectivity graph, determine the first global feature vector of the target brain functional connectivity graph and the first global feature vector of the target subgraph.

[0026] Based on the feature vectors of each node in the reference brain functional connectivity graph, the second full-graph global feature vector of the reference brain functional connectivity graph and the second subgraph global feature vector of the reference subgraph are determined respectively.

[0027] Based on the global feature vectors of the first full map and the first sub-map, as well as the global feature vectors of the second full map and the second sub-map, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined.

[0028] In one embodiment, the first global feature vector of the target brain functional connectivity graph and the first global feature vector of the target subgraph are determined based on the feature vectors of each node in the target brain functional connectivity graph, including:

[0029] Based on the feature vectors of each node in the target brain functional connectivity graph, determine the global layer feature vector of the corresponding network layer structure, and based on the feature vectors of each node in the target subgraph, determine the subgraph global layer feature vector of the corresponding network layer structure.

[0030] The first global feature vector of the whole graph is obtained by concatenating the global feature vectors of each network layer structure with their corresponding weights. The first subgraph global feature vector is obtained by concatenating the subgraph global feature vectors of each network layer structure with their corresponding weights.

[0031] In one embodiment, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined based on the first full-map global feature vector and the first sub-map global feature vector, as well as the second full-map global feature vector and the second sub-map global feature vector, including:

[0032] The global eigenvectors of the first subgraph and the global eigenvectors of the second subgraph are correlated to obtain the target correlation matrix and the reference correlation matrix.

[0033] Generate the target local mask and the reference local mask based on the target correlation matrix and the reference correlation matrix, respectively;

[0034] Based on the global feature vectors of the first full-map, the global feature vectors of the second full-map, the target local mask, and the reference local mask, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined.

[0035] In one embodiment, correlation calculations are performed on the global feature vectors of the first subgraph and the global feature vectors of the second subgraph to obtain a target correlation matrix and a reference correlation matrix, including:

[0036] Concatenate all global feature vectors of the first subgraph to obtain the first concatenated vector;

[0037] Concatenate all global feature vectors of the second subgraph to obtain the second concatenated vector;

[0038] Perform pairwise bilinear pooling operations based on the first and second concatenation vectors to obtain the target correlation matrix and the reference correlation matrix.

[0039] In one embodiment, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined based on the first full-map global feature vector, the second full-map global feature vector, the target local mask, and the reference local mask, including:

[0040] Generate enhanced target feature vectors based on the global feature vectors of the first full-image and the target local mask;

[0041] Generate a reference enhanced feature vector based on the global feature vector of the second full image and the reference local mask;

[0042] The similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined based on the target augmented feature vector and the reference augmented feature vector.

[0043] In one embodiment, determining a similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the target augmentation feature vector and the reference augmentation feature vector includes:

[0044] Perform a dot product on the target enhanced feature vector and the reference enhanced feature vector, and obtain the combined similarity based on the dot product result;

[0045] Obtain the target autocorrelation matrix between local features in the target augmented feature vector, and the reference autocorrelation matrix between local features in the reference augmented feature vector;

[0046] Obtain the minimum cosine distance matrix between the target autocorrelation matrix and the reference autocorrelation matrix, and determine the alignment similarity based on the minimum cosine distance matrix;

[0047] The similarity value is determined based on the combination similarity and alignment similarity.

[0048] In one embodiment, determining the category of the target brain functional connectivity map based on a similarity value includes:

[0049] Obtain the set of similarity values ​​between the target brain functional connectivity map and all reference brain functional connectivity maps;

[0050] Based on the set of similarity values, the category of the target brain functional connectivity map is determined by weighted K-nearest neighbor classification.

[0051] Secondly, this application also provides an image classification apparatus, comprising:

[0052] The connectivity graph acquisition module is used to acquire the target brain function connectivity graph of the tested object and the reference brain function connectivity graph of the preset category;

[0053] The similarity determination module is used to input the target brain functional connectivity map and the reference brain functional connectivity map into a spatial graph neural network to obtain the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map output by the graph neural network.

[0054] The category determination module is used to determine the category of the target brain functional connectivity map based on the similarity value.

[0055] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned image classification methods.

[0056] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described image classification methods.

[0057] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described image classification methods.

[0058] The aforementioned image classification methods, apparatus, computer devices, storage media, and computer program products acquire a target brain function connectivity map of the tested object and a reference brain function connectivity map of a preset category. These maps are then input into a spatial graph neural network (GNN) to obtain a similarity value between the target and reference GNN outputs. The category of the target GNN is then determined based on this similarity value. Spatial graph neural networks have the advantage of accurately extracting structural attribute features, and the target and reference GNNs are images with significant structural attributes. Therefore, the spatial graph neural network used in the above method can accurately extract structural attribute features from the target and reference GNNs and accurately determine the similarity value between them, improving the accuracy of the determined category of the target GNN and thus enhancing the accuracy of image classification. Attached Figure Description

[0059] Figure 1 This is an internal structural diagram of a computer device in one embodiment;

[0060] Figure 2 This is a flowchart illustrating an image classification method in one embodiment;

[0061] Figure 3This is a flowchart illustrating the process of constructing a target brain functional connectivity map in one embodiment;

[0062] Figure 4 This is a flowchart illustrating the process of constructing a target brain functional connectivity map in another embodiment;

[0063] Figure 5 This is a flowchart illustrating the process of determining similarity values ​​in one embodiment;

[0064] Figure 6 This is a partial block diagram of a spatial-based graph neural network in one embodiment;

[0065] Figure 7 This is a flowchart illustrating the process of determining similarity values ​​in another embodiment;

[0066] Figure 8 This is a flowchart illustrating the process of determining similarity values ​​in another embodiment;

[0067] Figure 9 This is a diagram illustrating the process of determining similarity values ​​in one embodiment;

[0068] Figure 10 This is a flowchart illustrating the process of determining similarity values ​​in another embodiment;

[0069] Figure 11 This is a flowchart illustrating the process of determining similarity values ​​in another embodiment;

[0070] Figure 12 This is a flowchart illustrating the process of determining the category of a target brain functional connectivity map based on similarity values ​​in one embodiment.

[0071] Figure 13 This is a structural block diagram of an image classification device in one embodiment. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0073] The image classification method provided in this application embodiment can be applied to, for example... Figure 1The computer device shown can be a terminal, including a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an image classification method. The display screen can be an LCD screen or an e-ink display screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0074] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0075] In one embodiment, such as Figure 2 As shown, an image classification method is provided, which can be applied to... Figure 1 Taking a computer device as an example, the explanation includes the following steps:

[0076] S210. Obtain the target brain function connectivity map and the reference brain function connectivity map of the preset category for the tested object.

[0077] Among them, the brain functional connectivity map (i.e., the brain functional connectivity network map) is used to represent the connections between different regions of the brain. The reference brain functional connectivity map with preset categories is a pre-constructed brain functional connectivity map with known categories, stored in a computer device.

[0078] It should be noted that when applying brain functional connectivity maps to classify early-stage MCI patients and normal aging individuals, the preset categories can include both early-stage MCI and normal aging. There are only minor structural differences between the brain functional connectivity maps of early-stage MCI patients and those of normal aging individuals.

[0079] Optionally, the computer device can perform functional magnetic resonance imaging (fMRI) on the brain of the subject to be tested, so as to construct a target brain functional connectivity map of the subject to be tested based on the brain signals obtained during the fMRI imaging process, and receive a reference brain functional connectivity map of a preset category selected by the operator.

[0080] S220. Input the target brain function connectivity map and the reference brain function connectivity map into a spatial graph neural network to obtain the similarity value between the target brain function connectivity map and the reference brain function connectivity map output by the graph neural network.

[0081] Graph Neural Networks (GNNs) are connection models that extract dependencies in a network graph by observing the information exchange between nodes, thus representing the graph's features. Spatial-based GNNs are used to extract structural attribute features from the network graph. Similarity values ​​obtained using spatial-based GNNs are used to evaluate the similarity between a target brain functional connectivity map and a reference brain functional connectivity map. A higher similarity value indicates a higher similarity between the target and reference brain functional connectivity maps; conversely, a lower similarity value indicates a lower similarity.

[0082] Optionally, the computer device inputs the target brain functional connectivity map and the reference brain functional connectivity map into a spatial graph neural network, and uses the graph neural network to extract structural attribute enhancement features from the target brain functional connectivity map and the reference brain functional connectivity map respectively, so as to determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map.

[0083] S230. Determine the category of the target brain functional connectivity map based on the similarity value.

[0084] Optionally, if a reference brain functional connectivity map of a preset category is pre-stored in the computer device, the computer can obtain the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map, and compare the similarity value with a similarity threshold to determine the category of the target brain functional connectivity map based on the comparison result. For example, if the reference brain functional connectivity map of the preset category is a brain functional connectivity map of early MCI, and the similarity value is greater than the similarity threshold, the computer device determines that the category of the target brain functional connectivity map is early MCI; conversely, if the similarity value is less than or equal to the similarity threshold, the computer device determines that the category of the target brain functional connectivity map is not early MCI.

[0085] Optionally, if the computer device has multiple preset categories of reference brain functional connectivity maps pre-stored, the computer device can obtain the similarity values ​​between the target brain functional connectivity map and the multiple preset categories of reference brain functional connectivity maps, and determine the category of the target brain functional connectivity map based on the similarity values ​​corresponding to each preset category. For example, by comparing the similarity values ​​corresponding to each preset category, the preset category with the largest similarity value that is greater than a similarity threshold is determined as the category of the target brain functional connectivity map.

[0086] In this embodiment, a computer device acquires a target brain function connectivity map and a reference brain function connectivity map of a preset category for the tested object. These two maps are then input into a spatial graph neural network (GNN) to obtain a similarity value between the GNN outputs. The category of the target brain function connectivity map is then determined based on this similarity value. The spatial GNN has the advantage of accurately extracting structural attribute features from brain function connectivity maps, and the target and reference brain function connectivity maps are images with unique and significant structural attributes. Therefore, the spatial GNN used in the above method can accurately extract structural attribute features from the target and reference brain function connectivity maps and accurately determine the similarity value between them, improving the accuracy of the determined category of the target brain function connectivity map and thus improving the accuracy of image classification.

[0087] In practical applications, a target brain functional connectivity map of the tested subject can be constructed based on the blood oxygen level dependent (BOLD) signal obtained during fMRI imaging. Based on this, such as... Figure 3 As shown, obtaining the target brain functional connectivity map of the test subject in S210 above includes:

[0088] S310. Obtain the brain oxygenation level-dependent signal of the tested subject.

[0089] Among them, the brain oxygenation level-dependent signal is used to reflect changes in brain oxygen saturation and blood flow. In this embodiment, the brain oxygenation level-dependent signal acquired from the subject is a low-frequency brain oxygenation level-dependent signal.

[0090] Optionally, during the functional magnetic resonance imaging (fMRI) of the brain, the computer equipment can acquire the raw brain oxygenation level-dependent signal of the subject and perform preprocessing operations such as time point correction, spatial smoothing, head motion correction, noise removal, and spatial normalization on the raw brain oxygenation level-dependent signal. It can also perform bandpass filtering on the signal frequency in the range of 0.01 to 0.1 Hz to obtain a low-frequency brain oxygenation level-dependent signal as the brain oxygenation level-dependent signal of the subject.

[0091] S320. Based on the brain functional area template and the brain blood oxygen level dependent signal, the brain spatial domain of the test subject is divided into regions, and the blood oxygen level dependent signal of each brain functional area is determined.

[0092] Among them, the blood oxygen level dependent signal of the brain includes the blood oxygen level dependent signal of the entire brain spatial domain of the subject.

[0093] Optionally, the computer device uses a preset brain functional region template to divide the brain spatial domain of the test subject into multiple brain functional regions, also known as regions of interest (ROIs), and then obtains the blood oxygen level-dependent signal of each brain functional region.

[0094] S330. Construct a target brain function connectivity map based on the blood oxygenation level-dependent signals of each brain functional area.

[0095] Optionally, the computer device can use each brain functional area as a node in the target brain functional connectivity map, and determine the correlation between each brain functional area based on the blood oxygen level dependent signal of each brain functional area, so as to determine whether there are connecting edges between the corresponding nodes, and then form the target brain functional connectivity map of the tested object from each node and the connecting edges between nodes.

[0096] In an optional embodiment, the existence of a connection edge between corresponding nodes can be determined based on the correlation coefficient of the blood oxygenation level-dependent signal between any two nodes, and then a target brain functional connectivity map can be constructed based on the nodes and the connection edges. Based on this, as... Figure 4 As shown, in step S330 above, based on the blood oxygenation level-dependent signals of each brain functional area, a target brain functional connectivity map is constructed, including:

[0097] S410. Using each brain functional area as a node, determine the blood oxygen level dependent signal of each node based on the blood oxygen level dependent signal of each brain functional area.

[0098] Optionally, after obtaining the blood oxygen level-dependent signals of each brain functional area in the partition, the computer device uses each brain functional area as a node to obtain the average value of the blood oxygen level-dependent signals corresponding to each voxel in each brain functional area, and obtains the blood oxygen level-dependent signal S of each node.

[0099] in, N R N represents the number of functional areas. T Indicates the number of voxels.

[0100] S420. Determine the correlation coefficient between any two nodes based on the blood oxygen level dependent signal of each node.

[0101] The blood oxygen level-dependent signal is a time-series signal. In this embodiment, the Pearson partial correlation coefficient between nodes can be determined based on the time-series information of the blood oxygen level-dependent signal corresponding to each node, and used as the correlation coefficient between nodes.

[0102] Optionally, the computer device can use each brain functional region obtained from the partitioning as a node in the target brain functional connectivity map, and obtain the Pearson partial correlation coefficient between the blood oxygenation level-dependent signals of any two brain functional regions, i.e., any two nodes, as the correlation coefficient P between any two nodes. i (m,n).

[0103]

[0104] Where (m,n) represents any two nodes (i.e., brain functional areas) in the brain spatial domain of the tested subject i, and t m With t n Let m and n represent the time series of nodes respectively. and Let m and n represent the time series averages of nodes m and n, respectively.

[0105] S430. Determine whether there is a connecting edge between any two nodes based on the correlation coefficient.

[0106] Optionally, the computer device can determine whether a connection exists between any two nodes by using a threshold comparison method based on the obtained correlation coefficient. For example, if the correlation coefficient between any two nodes is greater than the threshold, then a connection exists between these two nodes; conversely, if the correlation coefficient between any two nodes is equal to or less than the threshold, then no connection exists between these two nodes. For multiple nodes, the computer device can also use the K-Nearest Neighbor (KNN) algorithm to ensure that each node is only connected to the node with the highest obtained correlation coefficient, thereby forming a connection between nodes.

[0107] Optionally, in the K-nearest neighbor algorithm, when K=2, after obtaining the correlation coefficients between any two nodes in the multiple nodes, the computer determines, in one iteration, whether there are connecting edges between the two neighboring nodes with the largest correlation coefficients for the same node, while there are no connecting edges between nodes with other correlation coefficients. That is, in each iteration, each node is connected to at most two neighboring nodes. For example, if the brain spatial domain of the tested subject includes four nodes (a, b, c, d), obtaining the correlation coefficients between any two nodes will yield six correlation coefficients (P1 between a and b, P2 between a and c, P3 between a and d, P4 between b and c, P5 between b and d, and P6 between c and d). Assuming P1 > P2 > P3 > P4 > P5 > P6, then among the correlation coefficients P1, P2, and P3 related to 'a', if P1 > P2 > P3, then there is a connection between 'a' and 'b', and between 'a' and 'c'; among the correlation coefficients P1, P4, and P5 related to 'b', if P1 > P4 > P5, then there is a connection between 'b' and 'a', and between 'b' and 'c'; among the correlation coefficients P2, P4, and P6 related to 'c', if P2 > P4 > P6, then there is a connection between 'c' and 'a', and between 'c' and 'b'; among the correlation coefficients P3, P5, and P6 related to 'd', if P3 > P5 > P6, then there is a connection between 'd' and 'a', and between 'd' and 'b'.

[0108] S440. Form a target brain function connectivity diagram based on nodes and connecting edges.

[0109] Optionally, after determining the nodes and connections between nodes in the brain spatial domain of the test subject, the computer device can form a target brain function connectivity map of the test subject from the nodes and the connections between them.

[0110] In this embodiment, the target brain functional connectivity map can accurately reflect the correlation between various brain functional areas in the brain of the subject under test, thereby improving the accuracy of image classification based on the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map.

[0111] The spatial-based graph neural network includes a convolutional embedding network for extracting time-dimensional feature vectors from the target brain functional connectivity map and the reference brain functional connectivity map, which are then used to determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map. Based on this, in one embodiment, such as Figure 5 As shown, in S220 above, the target brain functional connectivity map and the reference brain functional connectivity map are input into a spatial graph neural network to obtain the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map output by the graph neural network, including:

[0112] S510. Extract the feature vectors of each node in the target brain functional connectivity graph.

[0113] In this target brain function connectivity map, the nodes correspond to brain functional areas.

[0114] S520. Extract the feature vectors of each node in the reference brain functional connectivity graph.

[0115] Among them, the feature vector of each node... represents the initial representation of each node in the brain functional connectivity graph, which is determined by the node itself and the nodes connected to it.

[0116] Optionally, the computer device can acquire the blood oxygenation level-dependent signals of each node in the target brain functional connectivity map and the reference brain functional connectivity map, and perform convolutional embedding operations on the time axis of the blood oxygenation level-dependent signals of each node to extract the feature vector X, including the time dimension, of each node in the target brain functional connectivity map and the reference brain functional connectivity map, respectively. i .

[0117] in, V i N represents a node. F Indicates the size of the feature map.

[0118] Among them, convolutional embedding networks It is a convolutional embedding network consisting of three one-dimensional kernel convolutional layers (1D convolutions). In this embodiment, the convolutional embedding network The first convolutional layer has a kernel size of 65 and contains 5 filters; the second convolutional layer has a kernel size of 30 and contains 10 filters; the last convolutional layer has a kernel size of 37 and an output feature map size of 20.

[0119] S530. Based on the feature vectors of each node in the target brain functional connectivity map and the feature vectors of each node in the reference brain functional connectivity map, determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map.

[0120] Optionally, the computer device can determine the feature representation of the target brain functional connectivity map from the perspective of global and local cross-structure based on the feature vectors of each node in the target brain functional connectivity map, and determine the feature representation of the reference brain functional connectivity map from the perspective of global and local cross-structure based on the feature vectors of each node in the reference brain functional connectivity map, and then determine the similarity value between the two based on the feature representation of the target brain functional connectivity map and the reference brain functional connectivity map from the perspective of global and local cross-structure.

[0121] In this embodiment, a convolutional embedding network is used to extract the feature vectors, including the time dimension, of each node in the target brain functional connectivity map and the reference brain functional connectivity map. This provides an accurate data foundation for determining the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map, thereby improving the accuracy of the determined similarity value and ultimately improving the accuracy of image classification.

[0122] like Figure 6 As shown, the graph neural network also includes a subgraph extraction module and a global attention graph neural network. The subgraph extraction module is used to divide the target brain functional connectivity map and the reference brain functional connectivity map into subgraphs. The global attention graph neural network is used to capture the feature vectors of the target brain functional connectivity map and the reference brain functional connectivity map from a global structural perspective, for subsequent determination of the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map. Based on this, in one embodiment, as... Figure 7 As shown, S630 above, based on the feature vectors of each node in the target brain functional connectivity map and the feature vectors of each node in the reference brain functional connectivity map, determines the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map, including:

[0123] S710. Divide the target brain function connectivity map into multiple target sub-maps and divide the reference brain function connectivity map into multiple reference sub-maps.

[0124] Optionally, the computer device can use a subgraph extraction module to divide the target brain functional connectivity map into multiple subgraphs using a preset partitioning strategy, thereby obtaining multiple target subgraphs of the target brain functional connectivity map, and use the same partitioning strategy to divide the reference brain functional connectivity map into multiple subgraphs, thereby obtaining multiple reference subgraphs of the reference brain functional connectivity map.

[0125] Optionally, the computer device can input a target brain functional connectivity map and a reference brain functional connectivity map into a subgraph extraction module. The subgraph extraction module then divides the target brain functional connectivity map into multiple target subgraphs and the reference brain functional connectivity map into multiple reference subgraphs based on the average brain functional connectivity of the target and reference brain functional connectivity maps and using a non-parametric metric graph cutting algorithm (such as the Louvain algorithm). For example, the computer device can extract multiple reference subgraphs from the input brain functional connectivity map pair (G...). A G B ), which are divided into m subgraphs: {G A1 G A2 , ...G Am},{G B1 G B2 , ...G Bm}, where G A and G B These represent the target brain functional connectivity map and the reference brain functional connectivity map, respectively.

[0126] S720. Based on the feature vectors of each node in the target brain functional connectivity graph, determine the first global feature vector of the target brain functional connectivity graph and the first global feature vector of the target subgraph.

[0127] Among them, the first full-map global feature vector is the feature vector of the target brain functional connectivity map in the global structural perspective; the first sub-map global feature vector is the feature vector of the sub-map of the target brain functional connectivity map, i.e., the target sub-map, in the global structural perspective.

[0128] Optionally, the computer device can determine the global layer feature vector of the corresponding network layer structure based on the feature vectors of each node in the target brain functional connectivity graph, and determine the subgraph global layer feature vector of the corresponding network layer structure based on the feature vectors of each node in the target subgraph, and then combine the global layer feature vectors of the neural network layer structures of each graph. With the corresponding weight w k After weighting and concatenating, the first global feature vector of the entire graph is obtained. The subgraph global layer feature vectors of each graph neural network layer structure With the corresponding weight w k After weighting and concatenating, the global feature vector of the first subgraph is obtained.

[0129] Here, the network layer structure is the layer structure in a global attention graph neural network, and k represents the k-th layer. The weights w corresponding to different layer structures... k The differences are highlighted. Each layer in a global attention map neural network determines its own output response value based on the response value of the layer above it. The output of the final layer in the global attention map neural network is the feature vector of each node in the target brain functional connectivity map and the reference brain functional connectivity map.

[0130] Alternatively, the computer device can use the following formula to determine the global layer node features of the graph neural network layer structure corresponding to each node in different brain function connectivity graphs.

[0131]

[0132] in, ε represents the features of node v output from the k-th layer structure, where k ∈ [1, 2, ..., K]; N(v) is a group of nodes adjacent to node v. (k) It is a learnable parameter, and MLP is a multilayer perceptron with nonlinearity.

[0133] Alternatively, the computer device may use the following formula to determine the global feature vector of the first full-image. and the global feature vector of the first subgraph

[0134]

[0135]

[0136] in, G, representing the target brain functional connectivity map A The global layer features of the entire graph output by each node in the first to Kth layers of the structure; Characterize the target subgraph G respectively Am The subgraph global layer features output by each node in the structure of layers 1 to K.

[0137] S730. Based on the feature vectors of each node in the reference brain functional connectivity graph, determine the second global feature vector of the reference brain functional connectivity graph and the second global feature vector of the reference subgraph.

[0138] Among them, the global feature vector of the second full map is the feature vector of the full map of the reference brain functional connectivity map from the perspective of global structure; the global feature vector of the second sub-map is the feature vector of the sub-map of the reference brain functional connectivity map, i.e., the reference sub-map, from the perspective of global structure.

[0139] Optionally, the computer device can determine the global layer feature vector of the corresponding graph neural network layer structure based on the feature vector of each node in the reference brain functional connectivity graph, and determine the subgraph global layer feature vector of the corresponding graph neural network layer structure based on the feature vector of each node in the reference subgraph. Then, the global layer feature vector of each graph neural network layer structure is weighted and concatenated with the corresponding weight to obtain the second global feature vector of the whole graph, and the global layer feature vector of each graph neural network layer structure is weighted and concatenated with the corresponding weight to obtain the second global feature vector of the subgraph.

[0140] It should be noted that the specific process of determining the second global feature vector of the reference brain functional connectivity map and the second global feature vector of the reference subgraph can be found in the above-mentioned S720 process of determining the first global feature vector of the target brain functional connectivity map and the first global feature vector of the target subgraph, which will not be repeated here.

[0141] S740. Based on the global feature vector of the first full map and the global feature vector of the first sub-map, as well as the global feature vector of the second full map and the global feature vector of the second sub-map, determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map.

[0142] Optionally, the computer device can perform association calculations based on the global feature vectors of the first subgraph and the second subgraph respectively, so as to enhance the global feature vectors of the first and second subgraphs respectively based on the obtained association results, and then determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the enhanced global feature vectors of the first and second subgraphs.

[0143] In this embodiment, the feature vectors of the full map and sub-maps of the target brain functional connectivity map from the perspective of global structure can fully characterize the global and local features of the target brain functional connectivity map. Therefore, based on the global and local features of the target brain functional connectivity map and the reference brain functional connectivity map, the similarity value between the two can be accurately determined, thereby improving the accuracy of image classification.

[0144] like Figure 6 As shown, the graph neural network also includes a cross-attention model, which generates an attention mask to highlight discriminative and comparative regional features in the input graph, for subsequent determination of the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map. Based on this, in one embodiment, such as Figure 8 As shown, S740 above, based on the first full-map global feature vector and the first sub-map global feature vector, as well as the second full-map global feature vector and the second sub-map global feature vector, determines the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map, including:

[0145] S810. Perform correlation calculations on the global eigenvectors of the first subgraph and the global eigenvectors of the second subgraph respectively to obtain the target correlation matrix and the reference correlation matrix.

[0146] The target correlation matrix is ​​the correlation matrix obtained by performing correlation calculations on the global feature vectors of the first subgraph of the target brain functional connectivity map; the reference correlation matrix is ​​the correlation matrix obtained by performing correlation calculations on the global feature vectors of the second subgraph of the reference brain functional connectivity map.

[0147] Optionally, the computer device can concatenate all global feature vectors of the first subgraph to obtain a first concatenated vector, and concatenate all feature vectors of the second subgraph to obtain a second concatenated vector. Then, it can perform pairwise bilinear pooling operations on the first and second concatenated vectors to obtain the target correlation matrix and the reference correlation matrix, respectively.

[0148] The pairwise bilinear pooling operation PwB(·) is:

[0149]

[0150]

[0151] This represents the first concatenated vector obtained by concatenating all global feature vectors of the first subgraph. N represents the second concatenated vector obtained by concatenating the global feature vectors of all second subgraphs. F′ This represents the feature dimension.

[0152] The output of this layer can be abstracted as:

[0153]

[0154]

[0155] Among them, Corr A Represents the target correlation matrix, Corr B This represents the reference correlation matrix, and Vec(·) represents the vectorization operation.

[0156] S820. Generate the target local mask and the reference local mask based on the target correlation matrix and the reference correlation matrix, respectively.

[0157] The target local mask is the attention mask of the target brain functional connectivity map obtained based on the target correlation matrix; the reference local mask is the attention mask of the reference brain functional connectivity map obtained based on the reference correlation matrix.

[0158] Optionally, based on the target correlation matrix Corr A Generate target local mask Att A For example, combined with Figure 9 The computer device can obtain the target local mask Att through the following process. A :

[0159] Corr A As input, the spatial average pooling operation SAP(·) is first applied, followed by an embedded subnetwork E. θ (·) to capture the target brain functional connectivity map G A Comparative and discriminative characteristics of spatial perception:

[0160]

[0161]

[0162] in, and E θ The parameters in (·) are and The convolution function is used. B(·) is the batch normalization operation, and σ(·) is the ReLU function. Then, the attention mask, i.e., the target local mask Att, is generated using the following formula.A :

[0163]

[0164]

[0165] Here, Softmax(·) is the Softmax function used to normalize the output. This indicates an element-wise multiplication operation. and These are two embedding functions responsible for reducing the dimensionality of pooled pairs of features and reweighting the feature response values. In implementation, and E ρ The architecture of (·) is Conv(1×1)→BatchNormalization→ReLU, where 1×1 represents the kernel size.

[0166] Similarly, the above process can be used to obtain the reference brain functional connectivity map G. B Attention mask, i.e., reference local mask Att B .

[0167] S830. Based on the first full-map global feature vector, the second full-map global feature vector, the target local mask, and the reference local mask, determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map.

[0168] Optionally, after obtaining the first global feature vector and target local mask of the target brain functional connectivity map, and the second global feature vector and reference local mask of the reference brain functional connectivity map, the computer device can enhance the first global feature vector using the target local mask and enhance the second global feature vector using the reference local mask, and then determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the enhanced first global feature vector and second global feature vector.

[0169] In one embodiment, such as Figure 10 As shown, S830, based on the first full-map global feature vector, the second full-map global feature vector, the target local mask, and the reference local mask, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined, including:

[0170] S1010. Generate target enhancement feature vector based on the global feature vector of the first full image and the target local mask.

[0171] S1020. Generate a reference enhanced feature vector based on the global feature vector of the second full image and the reference local mask.

[0172] Among them, the target local mask and the reference local mask are feature masks for local substructure awareness.

[0173] Optionally, such as Figure 9 As shown, the computer device can use residual attention to generate a first global feature vector that enhances the first global feature vector of the entire image based on the target local mask, thus obtaining the target enhanced feature vector. It can also enhance the discriminative features in the second global feature vector of the entire image based on the reference local mask, thus obtaining the reference enhanced feature vector. Specifically, the enhancement involves the discriminative features in both the first and second global feature vectors of the entire image.

[0174] The strengthening process satisfies the following formula:

[0175]

[0176]

[0177] This represents the enhanced feature vector of the target obtained after enhancement. This represents the reference enhanced feature vector obtained after enhancement. and These represent element-wise multiplication and element-wise addition operations, respectively.

[0178] S1030. Determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the target enhanced feature vector and the reference enhanced feature vector.

[0179] Optionally, after obtaining the target enhanced feature vector of the target brain functional connectivity map and the reference enhanced feature vector of the reference brain functional connectivity map, the computer device can determine the similarity score between the target brain functional connectivity map and the reference brain functional connectivity map based on the target enhanced feature vector and the reference enhanced feature vector, and use the obtained similarity score as the similarity value between the two.

[0180] In this embodiment, correlation calculations can be used to determine the correlation between target subgraphs and generate corresponding attention masks (target local masks), and the correlation between reference subgraphs and generate corresponding attention masks (reference local masks). Then, based on the target local masks, comparative and discriminative local region features in the target brain functional connectivity map are enhanced, and based on the reference local masks, comparative and discriminative local region features in the reference brain functional connectivity map are enhanced, thus determining the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map. The comparative and discriminative local regions in the target brain functional connectivity map / reference brain functional connectivity map are target regions that fully reflect their features and serve as the basis for classification. Enhancing the features of these local regions enables more accurate identification and classification, and is more suitable for identifying target and reference brain functional connectivity maps with subtle differences.

[0181] In an alternative embodiment, such as Figure 5 As shown, the graph neural network also includes a metric learning module, which comprises a feature combination component and a feature alignment component to determine the similarity between the target brain functional connectivity map and the reference brain functional connectivity map in different dimensions, and then synthesize the two to obtain a similarity value between the target brain functional connectivity map and the reference brain functional connectivity map. Based on this, in one embodiment, as... Figure 11 As shown, S1030 above, determining the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the target enhanced feature vector and the reference enhanced feature vector, includes:

[0182] S1110. Perform a dot product on the target enhanced feature vector and the reference enhanced feature vector, and obtain the combined similarity based on the dot product result.

[0183] Among them, the combination similarity is used to characterize the similarity between the target brain functional connectivity map and the reference brain functional connectivity map in the dimension of direct feature association.

[0184] Optionally, the computer device can perform dot product on the target enhanced feature vector and the reference enhanced feature vector through the dot product layer in the feature combination component to combine the target enhanced feature vector and the reference enhanced feature vector. Then, a fully connected layer and a sigmoid function are used to perform feature weighting and single-value output on the dot product result to obtain the combined similarity between the target brain functional connectivity map and the reference brain functional connectivity map.

[0185] S1120. Obtain the target autocorrelation matrix between local features in the target enhanced feature vector, and the reference autocorrelation matrix between local features in the reference enhanced feature vector.

[0186] Alternatively, the computer device can obtain the target autocorrelation matrix SR(h) using the following formula. A) and reference autocorrelation matrix SR(h B ):

[0187]

[0188]

[0189] in, Each element in Local features representing the spatial location of features. for Each element e in i The normalized function is calculated using the following formula:

[0190]

[0191] μ i and σ i They represent e respectively i The sample mean and variance.

[0192] S1130. Obtain the minimum cosine distance matrix between the target autocorrelation matrix and the reference autocorrelation matrix, and determine the alignment similarity based on the minimum cosine distance matrix.

[0193] Among them, alignment similarity is used to characterize the similarity between the target brain functional connectivity map and the reference brain functional connectivity map in the feature alignment association dimension.

[0194] Optionally, after obtaining the target autocorrelation matrix and the reference autocorrelation matrix, the computer device obtains the minimum cosine distance matrix between the target autocorrelation matrix and the reference autocorrelation matrix according to the following formula, and then uses a fully connected layer and a sigmoid function to perform feature weighting and single-value output on the cosine distance matrix to obtain the alignment similarity between the target brain functional connectivity map and the reference brain functional connectivity map.

[0195] Wherein, the minimum cosine distance matrix D A,B Satisfy the following formula:

[0196]

[0197] S1140. Determine the similarity value based on the combination similarity and alignment similarity.

[0198] Optionally, the computer device can determine the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on a combination of combined similarity and alignment similarity. For example, the weighted sum of combined similarity and alignment similarity can be obtained as the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map. The weights are preset weights, which can be obtained by training a spatial graph neural network.

[0199] In this embodiment, by obtaining the combined similarity and aligned pixels between the target brain functional connectivity map and the reference brain functional connectivity map respectively, the similarity between the target brain functional connectivity map and the reference brain functional connectivity map under different dimensions is determined. Then, the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is obtained by combining the similarity under multiple dimensions, avoiding the absolute influence of the similarity under a single dimension on the classification result and improving the stability of image classification.

[0200] In practical applications, when multiple reference brain functional connectivity maps exist across different preset categories, the similarity values ​​between the target brain functional connectivity map and each of the reference brain functional connectivity maps are very close. To improve classification accuracy, in one embodiment, such as... Figure 12 As shown, S230 above, determining the category of the target brain functional connectivity map based on similarity values, includes:

[0201] S1210. Obtain the set of similarity values ​​between the target brain functional connectivity map and all reference brain functional connectivity maps.

[0202] Optionally, when there are multiple reference brain functional connectivity maps of different categories, the computer device obtains the similarity value between the target brain functional connectivity map and each reference brain functional connectivity map, that is, obtains the set of similarity values ​​between the target brain functional connectivity map and all reference brain functional connectivity maps.

[0203] S1220. Based on the set of similarity values, the category of the target brain functional connectivity map is determined using the weighted K-nearest neighbor classification algorithm.

[0204] Optionally, after obtaining the set of similarity values, the computer device can use a weighted K-nearest neighbor classification algorithm to classify all similarity values ​​in the set to determine the category of the target brain functional connectivity map based on the classification results. For example, for the target brain functional connectivity map n, the similarity values ​​between it and all six reference brain functional connectivity maps {(a,0),(b,1),(c,1),(d,0),(e,1),} (0 and 1 are category identifiers, referred to as "class labels") are (0.3, 0.2, 0.6, 0.1, 0.8). Taking K=3, the three nearest neighbors with the highest similarity to m are (a,c,e), with similarity values ​​of (0.3, 0.6, 0.8) and class labels of (0,1,1). The frequencies of the different labels belonging to these three nearest neighbor points (a,c,e) are calculated to be (0,1 times) and (1,2 times). The label 1 with the highest frequency of these three nearest neighbor points is taken as the label of the target brain functional connectivity map m. That is, the category corresponding to label 1 is the category of the target brain functional connectivity map n.

[0205] In this embodiment, when there are multiple reference brain functional connectivity maps of different preset categories, the weighted K-nearest neighbor classification algorithm can be used to classify all similarity values ​​between the obtained target brain functional connectivity map and all reference brain functional connectivity maps, so as to determine the category of the target brain functional connectivity map based on the classification results, thereby improving the accuracy of the classification results and being more suitable for situations where the obtained similarity values ​​are very close.

[0206] In one embodiment, the above method further includes a process of training a graph neural network, including:

[0207] During the model training phase, a multi-cross training method is used to divide the training dataset for multiple training sessions, resulting in a space-based graph neural network.

[0208] In a single training session, for all N training subjects with known class labels (normal aging / early MCI) and unknown class labels, they are randomly divided into two groups, and a global bipartite graph is constructed, forming a graph consisting of N... 2 The training dataset consists of 4 pairs of subjects and their brain functional connectivity networks.

[0209] During model training, the computer can use constrained variance loss to train the proposed network. The loss for different components is defined as follows:

[0210] L Com,Align =max(0, σ) 2+ -a)+max(0,σ 2- -a)+max(0,r-(μ) + -μ - ))

[0211] Among them, L Com L represents the combination similarity. Align Indicates alignment similarity. μ + μ represents the average similarity between feature maps belonging to the same group. - σ represents the average similarity between feature maps belonging to different groups. 2+ and σ 2- Let represent the pairwise similarity variances of the same group and different groups, respectively; r represents the difference between the mean values ​​of the similarity distributions of the same group and different groups; and a represents the variance threshold.

[0212] Then, using weighted fusion, the complete objective function for model training is obtained as follows:

[0213] L = L Com +βL Align

[0214] Here, β is an adjustment parameter that can be optimized and learned during training.

[0215] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0216] Based on the same inventive concept, this application also provides an image classification apparatus for implementing the image classification method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more image classification apparatus embodiments provided below can be found in the limitations of the image classification method described above, and will not be repeated here.

[0217] In one embodiment, such as Figure 13 As shown, an image classification device is provided, including: an acquisition module 1301, a determination module 1302, and a judgment module 1303, wherein:

[0218] The acquisition module 1301 is used to acquire the target brain functional connectivity map of the test subject and the reference brain functional connectivity map of the preset category;

[0219] The determination module 1302 is used to input the target brain functional connectivity map and the reference brain functional connectivity map into a spatial graph neural network to obtain the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map output by the graph neural network.

[0220] The determination module 1303 is used to determine the category of the target brain functional connectivity map based on the similarity value.

[0221] Each module in the aforementioned image classification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0222] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-described image classification methods.

[0223] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described image classification methods.

[0224] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of any of the above-described image classification methods.

[0225] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0226] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0227] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An image classification method, characterized in that, The method includes: Obtain the target brain functional connectivity map and the reference brain functional connectivity map of the preset categories for the tested subjects; The target brain functional connectivity map and the reference brain functional connectivity map are input into a spatial graph neural network. Feature vectors of each node in the target brain functional connectivity map and the reference brain functional connectivity map are extracted. The target brain functional connectivity map is divided into multiple target subgraphs, and the reference brain functional connectivity map is divided into multiple reference subgraphs. Based on the feature vectors of each node in the target brain functional connectivity map, the first global feature vector of the target brain functional connectivity map and the first subgraph global feature vector of each target subgraph are determined. Based on the feature vectors of each node in the reference brain functional connectivity map, the reference brain functional connectivity map is determined. Using the second global feature vector of the first full-map and the second global feature vector of the reference sub-map, correlation calculations are performed on the global feature vectors of the first and second sub-maps respectively to obtain a target correlation matrix and a reference correlation matrix. A target local mask and a reference local mask are generated based on the target correlation matrix and the reference correlation matrix respectively. The similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined based on the first global feature vector, the second global feature vector, the target local mask, and the reference local mask. Nodes in the target brain functional connectivity map correspond to brain functional regions. The category of the target brain functional connectivity map is determined based on the similarity value.

2. The method according to claim 1, characterized in that, The acquisition of the target brain functional connectivity map of the tested subject includes: Obtain the brain blood oxygenation level-dependent signal of the tested subject; Based on the brain functional region template and the brain blood oxygen level dependent signal, the brain spatial domain of the tested object is divided into regions to determine the blood oxygen level dependent signal of each brain functional region; Based on the blood oxygenation level-dependent signals of each brain functional area, the target brain functional connectivity map is constructed.

3. The method according to claim 2, characterized in that, The construction of the target brain functional connectivity map based on the blood oxygenation level-dependent signals of each of the brain functional regions includes: Using each brain functional area as a node, the blood oxygen level dependent signal of each node is determined based on the blood oxygen level dependent signal of each brain functional area. The correlation coefficient between any two nodes is determined based on the blood oxygen level dependence signal of each node. Determine whether there is a connecting edge between any two nodes based on the correlation coefficient; The target brain functional connectivity map is formed based on the nodes and the connecting edges.

4. The method according to claim 1, characterized in that, The step of determining the first global feature vector of the target brain functional connectivity graph and the first subgraph global feature vector of the target subgraph based on the feature vectors of each node in the target brain functional connectivity graph includes: Based on the feature vectors of each node in the target brain functional connectivity graph, determine the global layer feature vector of the corresponding network layer structure, and based on the feature vectors of each node in the target subgraph, determine the subgraph global layer feature vector of the corresponding network layer structure. The first full-graph global feature vector is obtained by weighting the global feature vectors of each network layer structure with their corresponding weights and then concatenating them. The first subgraph global feature vector is obtained by weighting the global feature vectors of each subgraph structure with their corresponding weights.

5. The method according to claim 1, characterized in that, The step of performing correlation calculations on the global feature vectors of the first subgraph and the global feature vectors of the second subgraph to obtain the target correlation matrix and the reference correlation matrix includes: Concatenate all global feature vectors of the first subgraph to obtain the first concatenated vector; Concatenate all global feature vectors of the second subgraph to obtain the second concatenated vector; Pairwise bilinear pooling is performed on the first concatenation vector and the second concatenation vector to obtain the target correlation matrix and the reference correlation matrix.

6. The method according to claim 1, characterized in that, The step of determining the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the first full-map global feature vector, the second full-map global feature vector, the target local mask, and the reference local mask includes: Generate an enhanced target feature vector based on the first full-image global feature vector and the target local mask; A reference enhanced feature vector is generated based on the second full-image global feature vector and the reference local mask; The similarity value between the target brain functional connectivity map and the reference brain functional connectivity map is determined based on the target enhanced feature vector and the reference enhanced feature vector.

7. The method according to claim 6, characterized in that, Determining the similarity value between the target brain functional connectivity map and the reference brain functional connectivity map based on the target augmented feature vector and the reference augmented feature vector includes: Perform a dot product on the target enhanced feature vector and the reference enhanced feature vector, and obtain the combined similarity based on the dot product result; Obtain the target autocorrelation matrix between local features in the target enhanced feature vector, and the reference autocorrelation matrix between local features in the reference enhanced feature vector; Obtain the minimum cosine distance matrix between the target autocorrelation matrix and the reference autocorrelation matrix, and determine the alignment similarity based on the minimum cosine distance matrix; The similarity value is determined based on the combined similarity and the alignment similarity.

8. The method according to any one of claims 1-7, characterized in that, Determining the category of the target brain functional connectivity map based on the similarity value includes: Obtain the set of similarity values ​​between the target brain functional connectivity map and all reference brain functional connectivity maps; Based on the set of similarity values, a weighted average is applied. The nearest neighbor classification method determines the category of the target brain functional connectivity map.

9. An image classification device, characterized in that, The device includes: The acquisition module is used to acquire the target brain functional connectivity map and the reference brain functional connectivity map of the preset categories of the tested object. The determination module is used to input the target brain functional connectivity map and the reference brain functional connectivity map into a spatial graph neural network, extract feature vectors of each node in the target brain functional connectivity map, and extract feature vectors of each node in the reference brain functional connectivity map; divide the target brain functional connectivity map into multiple target subgraphs and the reference brain functional connectivity map into multiple reference subgraphs; determine the first global feature vector of the target brain functional connectivity map and the first subgraph global feature vector of the target subgraph based on the feature vectors of each node in the target brain functional connectivity map; and determine the reference brain functional connectivity map based on the feature vectors of each node in the reference brain functional connectivity map. The second global feature vector of the functional connectivity graph and the second global feature vector of the reference subgraph are used to perform correlation calculations on the global feature vectors of the first subgraph and the second subgraph, respectively, to obtain a target correlation matrix and a reference correlation matrix. A target local mask and a reference local mask are generated based on the target correlation matrix and the reference correlation matrix, respectively. A similarity value between the target brain functional connectivity graph and the reference brain functional connectivity graph is determined based on the first global feature vector, the second global feature vector, the target local mask, and the reference local mask. Nodes in the target brain functional connectivity graph correspond to brain functional regions. The determination module is used to determine the category of the target brain functional connectivity map based on the similarity value.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images