An adaptive graph hinting strategy-based mental illness auxiliary diagnosis system

Abstract

The present application relates to the technical field of medical auxiliary diagnosis, and particularly relates to a mental illness auxiliary diagnosis system based on an adaptive graph prompt strategy, which comprises: a brain function network construction module, which is used for calculating the functional connection strength value of a brain region to construct an individualized brain function connection graph; a disease prototype prompt graph library module, which is used for storing a plurality of mental illness prototype prompt graphs; a graph isomorphism coding module, which is used for graph isomorphism coding to obtain an individual graph representation vector and a prompt graph representation vector; a graph prompt matching module, which is used for calculating a similarity score to obtain an adaptive prompt fusion weight; a prompt edge weight modulation module, which is used for generating an individualized modulation matrix to obtain a modulated brain function connection graph; and a graph convolution classification module, which is used for obtaining a discriminative node embedding matrix according to the modulated brain function connection graph to obtain a mental illness auxiliary diagnosis probability distribution. The present application significantly improves the accuracy and cross-center generalization capability of auxiliary diagnosis of a plurality of mental illnesses.

Description

Technical Field

[0001] This invention relates to the field of medical auxiliary diagnostic technology, and in particular to an auxiliary diagnostic system for mental illness based on an adaptive graph cueing strategy. Background Technology

[0002] For the auxiliary diagnosis of mental illnesses, existing technologies mainly employ two strategies: First, based on constructing a brain functional connectivity matrix, traditional machine learning methods flatten the functional connectivity matrix and input it into a support vector machine or random forest classifier. Second, deep learning methods utilize multilayer perceptrons or convolutional neural networks to classify functional connectivity features. Some studies have attempted to input brain functional connectivity maps into graph neural networks to directly learn graph-level classification results. Furthermore, multi-task learning frameworks are introduced to train classification models for multiple diseases, or transfer learning is used to transfer knowledge from the source domain to the target domain. Regarding the fusion of group information, a group average map is typically used as auxiliary information, directly concatenated with individual features and input into the classifier.

[0003] Existing auxiliary diagnostic technologies have shortcomings that limit the accuracy and generalization ability of clinical diagnosis. First, traditional models often use fixed feature extraction and fixed weight fusion methods, ignoring the individual differences in brain function among different patients and failing to adapt to individual differences in brain pathology. Second, existing technologies lack a standardized reference system for disease prototype features, making it difficult to accurately capture the specific brain network pathological features of various mental illnesses, resulting in weak feature discrimination. In addition, conventional graph coding matching methods are easily affected by the randomness of single feature calculations, leading to insufficient feature matching accuracy and an inability to achieve adaptive feature modulation optimization. This makes it difficult to effectively enhance abnormal disease features and weaken invalid interference information, resulting in poor model diagnostic accuracy and stability, and significantly limiting the adaptability and clinical applicability to multiple diseases. Summary of the Invention

[0004] To address the shortcomings of existing technologies and meet the needs of practical applications, this invention provides a mental illness auxiliary diagnostic system based on an adaptive graph cueing strategy. The system includes: a brain functional network construction module for preprocessing resting-state functional magnetic resonance imaging (fMRI) images of subjects and calculating functional connectivity strength values ​​of brain regions to construct an individualized brain functional connectivity map; a disease prototype cueing map library module for storing various mental illness prototype cueing maps, each composed of a node set, an edge set, and an edge weight matrix; and a graph isomorphism encoding module for encoding the individualized brain functional connectivity map and the mental illness prototype cueing maps. Graph isomorphic encoding yields individual graph representation vectors and cue graph representation vectors; a graph cue matching module calculates the similarity score between the individual graph representation vectors and the cue graph representation vectors, and obtains adaptive cue fusion weights based on the similarity score; a cue edge weight modulation module generates an individualized modulation matrix based on the adaptive cue fusion weights, and obtains a modulated brain function connectivity map by combining it with the individualized brain function connectivity map; a graph convolutional integration module obtains a discriminative node embedding matrix based on the modulated brain function connectivity map, and obtains a probability distribution for auxiliary diagnosis of mental illness based on the discriminative node embedding matrix. This invention effectively adapts to the individualized differences in brain function among different patients, significantly improving the accuracy and generalization ability of intelligent auxiliary diagnosis of mental illness.

[0005] Optionally, the preprocessing of the subject's resting-state functional magnetic resonance imaging (fMRI) images and the calculation of functional connectivity strength values ​​of brain regions to construct an individualized brain functional connectivity map includes: performing head movement correction, spatial standardization, bandpass filtering, and smoothing on the resting-state fMRI images to obtain standardized image data; extracting the mean oxygenation level dependent signal time series of the brain regions from the standardized image data based on a preset brain region template; selecting any two brain regions, calculating the Pearson correlation coefficient between the mean oxygenation level dependent signal time series, and obtaining the functional connectivity strength value through the Pearson correlation coefficient; obtaining the absolute value of the functional connectivity strength value, and filtering brain region connections exceeding a preset threshold to construct the individualized brain functional connectivity map. This invention relies on brain region temporal signal correlation analysis to construct brain connectivity relationships, restore the subject's true brain functional linkage characteristics, and provides basic data support.

[0006] Optionally, the personalized brain functional connectivity graph includes: the personalized brain functional connectivity graph uses the selected brain regions as nodes, the connections between the brain regions as edges, and the functional connectivity strength values ​​as edge weights. This invention, through explicit rules for constructing personalized brain functional connectivity graphs, achieves standardized definitions of brain region nodes, connections, and weights, thus unifying the data structure paradigm.

[0007] Optionally, the various mental illness prototype diagrams include: prototype diagrams for depression, bipolar disorder, and schizophrenia. This invention enables intelligent auxiliary diagnosis of multiple mental illnesses, broadening the system's clinical application scenarios and coverage.

[0008] Optionally, the mental illness prototype prompting map consists of a node set, an edge set, and an edge weight matrix, including: extracting brain functional connectivity matrices for the mental illness patient group and the healthy control group based on the resting-state functional magnetic resonance imaging; performing statistical difference analysis on the brain functional connectivity matrices of the two groups to screen brain functional connectivity edges; forming the node set using the brain regions corresponding to the brain functional connectivity edges; forming the edge set using the brain functional connectivity edges; and obtaining the inter-group difference effect size as the initial edge weight to construct the edge weight matrix. This invention highlights the core pathological brain network characteristics of the disease, providing a highly recognizable disease benchmark feature basis for subsequent adaptive prompting matching.

[0009] Optionally, the step of performing graph isomorphic encoding on the individualized brain function connectivity map and the mental illness prototype cue map to obtain individual graph representation vectors and cue map representation vectors includes: the graph isomorphic encoding module includes a graph isomorphic network encoder, which is composed of multiple layers of graph isomorphic convolutional layers and pooling layers; inputting the individualized brain function connectivity map into the graph isomorphic network encoder to obtain the individual graph representation vector; and inputting the mental illness prototype cue map into the same graph isomorphic network encoder to obtain the cue map representation vector. This invention ensures consistency of the representation space through the graph isomorphic network encoder, which helps improve the accuracy of subsequent similarity matching.

[0010] Optionally, calculating the similarity score between the individual graph representation vector and the suggestion graph representation vector, and obtaining the adaptive suggestion fusion weight based on the similarity score, includes: using the individual graph representation vector as a query vector and the suggestion graph representation vector as a key vector to calculate the similarity score; and normalizing the similarity score to obtain the adaptive suggestion fusion weight. This invention achieves adaptive differential fusion of suggestion features, avoiding feature adaptation bias caused by fixed weights and improving diagnostic targeting.

[0011] Optionally, the step of using the individual graph representation vector as the query vector and the hint graph representation vector as the key vector to calculate the similarity score includes: performing multiple random masking operations on the query vector and the key vector to obtain multiple sets of masked representation vector pairs; calculating the attention score of the representation vector pairs; and obtaining the average of the attention scores as the similarity score. This invention effectively avoids the randomness and accidental interference of single feature calculations, enhancing the reliability of system feature matching.

[0012] Optionally, the step of generating a personalized modulation matrix based on the adaptive cue fusion weights and obtaining a modulated brain function connectivity map by combining the personalized brain function connectivity map includes: using the adaptive cue fusion weights as weighting coefficients to perform a weighted summation of the edge weight matrix of the mental illness prototype cue map to obtain the personalized modulation matrix; traversing each connection edge of the personalized brain function connectivity map, multiplying the edge weight value of the connection edge element-wise with the corresponding element value of the personalized modulation matrix to obtain the modulation edge weight, and forming the modulated brain function connectivity map based on the modulation edge weight. This invention effectively enhances the abnormal disease characteristics of the subject, weakens ineffective interference characteristics, and significantly improves the discriminative effect of subsequent classification diagnosis.

[0013] Optionally, the step of obtaining a discriminant node embedding matrix based on the modulated brain functional connectivity graph and obtaining a probability distribution for auxiliary diagnosis of mental illness based on the discriminant node embedding matrix includes: the graph convolutional network module includes a graph convolutional network, an average pooling layer, and a fully connected layer; inputting the adjacency matrix and initial node feature matrix of the modulated brain functional connectivity graph into the graph convolutional network to obtain the discriminant node embedding matrix; performing global average pooling on the discriminant node embedding matrix through the average pooling layer to obtain a full-graph feature vector; and inputting the full-graph feature vector into the fully connected layer to obtain the probability distribution for auxiliary diagnosis of mental illness. This invention efficiently extracts refined pathological features, accurately outputs the probability distribution for disease diagnosis, improves the accuracy and stability of diagnostic results, and adapts to the needs of clinical auxiliary diagnosis.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention introduces cue learning into the field of brain functional connectivity analysis, constructs a disease prototype cue library containing typical pathological patterns of multiple mental illnesses, breaks through the technical limitations of static template comparison in the prior art, and realizes auxiliary diagnosis driven by prior pathological knowledge.

[0015] Second, this invention screens brain function connectivity edges with significant differences through inter-group statistical difference analysis, uses the inter-group difference effect size as the initial edge weight, and designs the edge weight matrix as a learnable parameter for end-to-end optimization, thereby achieving a deep integration of prior medical knowledge and data-driven learning.

[0016] Third, this invention is based on the attention mechanism, using the individual graph representation vector as the query vector and the cue graph representation vector as the key vector to calculate the adaptive cue fusion weight, generating an individual-specific modulation matrix for each subject, thus overcoming the adaptation bias of the traditional fixed-weight fusion method.

[0017] Fourth, this invention selectively amplifies pathologically relevant connections and suppresses irrelevant interfering connections by multiplying the individualized modulation matrix with the edge weights of the brain functional connectivity graph element by element, while preserving the topological structure completely, thus significantly improving the classification and discrimination effect.

[0018] Fifth, this invention can simultaneously assist in the diagnosis of multiple mental illnesses, output multi-class probability distributions, and effectively reduce the risk of overfitting by introducing common prior knowledge of diseases, making it suitable for applications in cross-center and cross-population scenarios. Attached Figure Description

[0019] Figure 1 This is a framework diagram of a mental illness auxiliary diagnosis system based on an adaptive graph prompting strategy according to an embodiment of the present invention; Figure labeling: Brain functional network construction module A1, Disease prototype prompt library module A2, Graph isomorphic encoding module A3, Graph prompt matching module A4, Prompt edge weight modulation module A5, Graph volume integral module A6. Detailed Implementation

[0020] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0021] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0022] Please see Figure 1 An embodiment of the present invention provides a mental illness auxiliary diagnosis system based on an adaptive graph cue strategy. The system includes: a brain functional network construction module A1, a disease prototype cue graph library module A2, a graph isomorphic encoding module A3, a graph cue matching module A4, a cue edge weight modulation module A5, and a graph volume integral class module A6.

[0023] In this embodiment, the brain functional network construction module A1 is used to preprocess the resting-state functional magnetic resonance imaging (fMRI) images of the subject and calculate the functional connectivity strength values ​​of brain regions to construct an individualized brain functional connectivity map; including: S11. Perform head motion correction, spatial normalization, bandpass filtering, and smoothing on the resting-state functional magnetic resonance imaging to obtain standardized image data.

[0024] First, head motion correction is performed on the resting-state functional magnetic resonance imaging (fMRI) images, including: acquiring the raw format image data of the subject's resting-state fMRI scans and converting it into four-dimensional time-series images, which contain... Each time point corresponds to a 3D whole-brain functional image. Using the 3D whole-brain functional image at the center of the 4D time series as the reference image, the 3D images at each of the other time points are rigidly registered with the reference image one by one. During the registration process, the mutual information between the two images is used as the similarity measure. The least squares method is used to iteratively optimize the six rigid body motion parameters, including the translation along the three coordinate axes and the rotation around the three coordinate axes, so as to minimize the spatial position difference between the image at each time point and the reference image. After the rigid body registration of each time point is completed, the spatial position of the image at that time point is updated. After all time points are processed, the head motion corrected 4D image is obtained.

[0025] Secondly, the head-motion-corrected 4D images are spatially standardized and registered to the standard brain template space defined by the Montreal Neuroscience Institute. This includes: First, calculating the average image of the head-motion-corrected 4D images in the time dimension, affine registering this average image with the standard brain template, and obtaining the affine transformation matrix by maximizing normalized mutual information to initially correct the global position, angle, and scale differences between the average image and the standard brain template; Second, using a nonlinear registration algorithm to finely fit the local deformations between the affine-registered average image and the standard brain template, generating a nonlinear deformation field; Third, combining the affine transformation matrix and the nonlinear deformation field into a unified coordinate mapping relationship, applying this coordinate mapping relationship to each time point of the head-motion-corrected 4D images, and resampling the 3D images at each time point to the standard brain template space using trilinear interpolation. The resampled voxels are isotropic cubes, resulting in spatially standardized 4D images.

[0026] Next, bandpass filtering is performed on the spatially normalized four-dimensional image, including: bandpass filtering of the time series corresponding to each voxel in the spatially normalized four-dimensional image, with the filter passband range set to 0.01Hz–0.08Hz to retain low-frequency signal components related to spontaneous fluctuations in neural activity, while filtering out high-frequency physiological noise and low-frequency drift noise; a finite impulse response bandpass filter is used, with the filter order set to 64 to ensure sufficient frequency selectivity. The time series of each voxel is first forward filtered and then backward filtered to achieve zero phase shift and avoid introducing additional phase distortion during the filtering process.

[0027] Finally, after bandpass filtering, smoothing is performed, including spatial smoothing of the time series of all voxels in the whole brain. Spatial smoothing uses isotropic Gaussian kernels for three-dimensional convolution to further improve the signal-to-noise ratio and finally obtain standardized image data.

[0028] S12. Based on the preset brain region template, extract the average blood oxygen level dependent signal time series of the brain region through the standardized image data.

[0029] In this embodiment, the preset brain region template is a standardized brain region division scheme based on anatomical or functional connectivity, which divides the whole brain gray matter into... Each brain region has a unique integer label value. The preset brain region template and the standardized image data are located in the same standard brain template space, and the voxel size and spatial dimension of the two are consistent.

[0030] Specifically, the preset brain region template uses an automatic anatomical labeling template, which divides the whole brain into 116 brain regions, with each brain region numbered sequentially from 1 to 116.

[0031] In this embodiment, the label values ​​of each brain region in the preset brain region template are traversed, and the following operations are performed on each brain region: a binary mask image corresponding to the label value of that brain region is generated. In the binary mask image, the voxel positions with the label value equal to the current brain region label value are set to 1, and the other voxel positions are set to 0; the binary mask image is multiplied voxel by voxel with each time point of the standardized image data and the sum is obtained, and then divided by the total number of voxels contained in that brain region to obtain the average signal intensity of that brain region at that time point, which satisfies the following relationship:

[0032] in, For the first The brain region in the first The average signal strength at each time point For the first The total number of voxels contained in each brain region. For voxels, For the first A set of voxels from each brain region voxels In the Signal strength values ​​at each time point.

[0033] right Perform the above operations one by one at each time point to obtain the length of the brain region. The average signal time series, for all After performing the above operations on each brain region, the following results are obtained: The length of the strip is The mean blood oxygen level depends on the signal time series, with each time series corresponding to a brain region. A time series constitutes a OK The original signal matrix of the column.

[0034] S13. Select any two brain regions, calculate the Pearson correlation coefficient between the time series of the mean blood oxygen level dependent signal, and obtain the functional connectivity strength value through the Pearson correlation coefficient.

[0035] In this embodiment, firstly, the first signal is selected from the original signal matrix. The brain regions and the first Two mean blood oxygenation level-dependent signal time series corresponding to each brain region are denoted as follows: and Next, the Pearson correlation coefficient between the two time series is calculated. This is used to measure the consistency of blood oxygenation level-dependent signal fluctuations between two brain regions; finally, the Pearson correlation coefficient is obtained. Then, the Pearson correlation coefficient was subjected to Fisher's z-transformation to obtain the functional connectivity strength value. This transformation maps the original correlation coefficient from the range of [-1,1] to the entire real number axis, making the transformed value approximately follow a normal distribution, which is beneficial for subsequent statistical analysis and threshold screening.

[0036] The above Fisher z-transform satisfies the following relationship:

[0037] in, This represents the functional connection strength value. This is the Pearson correlation coefficient.

[0038] Perform the above calculations on each brain region pair one by one, obtaining a total of N(N-1) / 2 functional connectivity strength values, which constitute a OK Column functional connection strength symmetric matrix The main diagonal elements of the matrix are set to 0, and the nth element in the matrix is... Line 1 The element value of the column is equal to , indicating the first The brain region and the first The strength of functional connectivity between brain regions.

[0039] S14. Obtain the absolute value of the functional connectivity strength value, and filter brain region connections that exceed a preset threshold to construct the individualized brain functional connectivity map.

[0040] First, obtain the functional connectivity strength symmetric matrix. Then, the absolute value of each element in the matrix is ​​taken to obtain the absolute functional connectivity strength matrix. As an adjacency matrix of a personalized brain functional connectivity map, .

[0041] Then, set a preset threshold. Used to distinguish between brain region pairs with significant functional connectivity and weak connectivity caused by noise; preset threshold Based on the sparsity principle, the threshold is determined by traversing a series of candidate thresholds and calculating the proportion of connections retained under each candidate threshold to all possible connections. The candidate threshold with a proportion between 10% and 30% is selected as the preset threshold to ensure that the brain functional connectivity map has appropriate sparsity, which retains sufficient topological information while eliminating most noise edges.

[0042] Next, for the matrix Each element in ,like Then determine the first The brain region and the first Significant functional connections exist between brain regions; these connections are preserved. If no significant functional connectivity is found, the connectivity of that brain region is removed.

[0043] Finally, after threshold screening, based on the retained brain region connections, the screened brain regions are used as nodes, the retained brain region connections are used as edges, and the functional connectivity strength values ​​corresponding to the brain region connections are used as edge weights to generate an undirected weighted graph as an individualized brain functional connectivity graph.

[0044] In this embodiment, the disease prototype hint graph module A2 is used to store various mental illness prototype hint graphs, which are composed of a node set, an edge set, and an edge weight matrix; including: S21. Based on the resting-state functional magnetic resonance imaging, extract the brain functional connectivity matrix of the mental illness patient group and the healthy control group respectively.

[0045] The mental illness patient group included patients diagnosed with depression, bipolar disorder, and schizophrenia; the healthy control group consisted of healthy volunteers of the same age and sex as the patient group who had no history of mental illness.

[0046] The same steps S11-S13 were used to process the resting-state functional magnetic resonance imaging (fMRI) images of patients with mental illness and healthy controls to obtain the functional connectivity strength values ​​of the subjects. The functional connectivity strength values ​​of all brain region pairs were then organized into a single... OK The brain functional connectivity matrix of columns, in which The total number of brain regions defined by the preset brain region template.

[0047] Following the above method, the following four sets of brain functional connectivity matrices were obtained: brain functional connectivity matrix of patients with depression, brain functional connectivity matrix of patients with bipolar disorder, brain functional connectivity matrix of patients with schizophrenia, and brain functional connectivity matrix of the healthy control group.

[0048] S22. Perform statistical difference analysis on the brain function connectivity matrices of the two groups to screen brain function connectivity edges.

[0049] After obtaining the brain functional connectivity matrix for each group, for each mental illness, a statistical difference analysis was performed between the brain functional connectivity matrix of the patient group and the brain functional connectivity matrix of the healthy control group to screen out brain functional connectivity edges with significant inter-group differences.

[0050] Taking depression as an example, the brain functional connectivity matrix of the depression patient group was compared with that of the healthy control group one by one. For the connection of the first... The brain region and the first Brain functional connectivity edges were extracted from each brain region. The functional connectivity strength values ​​of all subjects in the depression patient group and the healthy control group were extracted for each edge, forming two independent samples. An independent two-sample t-test was performed on both groups to calculate the t-statistic and corresponding p-value. To control for false positives from multiple comparisons, the p-values ​​of all brain region pairs were corrected for false discovery rate. Brain region pairs with corrected p-values ​​less than a preset significance level (0.05, which can be adjusted according to actual needs) were selected as brain functional connectivity edges with significant inter-group differences.

[0051] It should be noted that the p-value represents the probability of observing the current difference or a more extreme difference under the premise that the null hypothesis is true. Therefore, the smaller the p-value, the lower the probability that the difference between the two groups is caused by random error, that is, the more significant the difference between the two groups.

[0052] Following the same procedure described above, statistical difference analysis was performed on the brain functional connectivity matrices of the bipolar disorder patient group and the schizophrenia patient group compared with the healthy control group to obtain the brain functional connectivity edges with significant inter-group differences for each group.

[0053] S23. The brain regions corresponding to the brain function connection edges are used to form the node set.

[0054] The node set is constructed by traversing the brain function connection edges corresponding to the mental illness, extracting the brain regions corresponding to the two endpoints of each edge, merging and deduplicating all extracted brain regions to form the node set of the prototype prompt graph of the mental illness. Each brain region in the node set has a unique brain region label, which is consistent with the brain region label in the preset brain region template.

[0055] S24. The edge set is formed by connecting the brain functions.

[0056] The edge set is constructed as follows: the combination of all brain function connection edges corresponding to the mental illness is used as the edge set. Each edge in the edge set is defined by the connection relationship between two brain regions in the node set. Each edge corresponds to a pair of brain region connections with significant inter-group differences.

[0057] S25. Obtain the inter-group difference effect size as the initial edge weights to construct the edge weight matrix.

[0058] The edge weight matrix is ​​constructed as follows: Each edge in the edge set is assigned an initial edge weight, and the between-group difference effect size is obtained as the initial edge weight. The between-group difference effect size is measured using the Cohen d-value, which satisfies the following relationship:

[0059] in, This represents the effect size of the difference between groups. For this group of patients with mental illness, in the first The brain region and the first The mean value of functional connectivity strength between brain regions This represents the mean of the functional connectivity strength values ​​corresponding to the healthy control group. Let be the combined standard deviation of the two sets.

[0060] It should be noted that the larger the absolute value of the initial edge weight, the more significant the difference between the patient group and the healthy control group, and the greater the contribution of the edge to the disease discrimination.

[0061] Furthermore, the initial edge weights of all edges are organized into a matrix form according to the arrangement order of the brain regions in the node set, thus obtaining the edge weight matrix.

[0062] It should be noted that the edge weight matrix is ​​a learnable parameter matrix. During the pre-training phase, the backpropagation algorithm is used to optimize the values ​​of each element in the matrix end-to-end, so that the values ​​of the elements in the matrix more accurately characterize the contribution strength of the corresponding connection edges in the typical abnormal functional connection patterns of mental illness.

[0063] In this embodiment, for each mental illness, a corresponding mental illness prototype hint graph is constructed based on the selected brain functional connectivity edges, which consists of three elements: a node set, an edge set, and an edge weight matrix. According to the above construction method of the node set, edge set, and edge weight matrix, prototype hint graphs for depression, bipolar disorder, and schizophrenia are generated respectively and stored in the disease prototype hint graph library module A2.

[0064] Among them, the prototype hint diagram for depression depicts the typical abnormal brain functional connectivity patterns of patients with depression, the prototype hint diagram for bipolar disorder depicts the typical abnormal brain functional connectivity patterns of patients with bipolar disorder, and the prototype hint diagram for schizophrenia depicts the typical abnormal brain functional connectivity patterns of patients with schizophrenia. The prototype hint diagrams for the three diseases have different node sets, edge sets, and edge weight matrices, and are stored independently for retrieval.

[0065] In this embodiment, the graph isomorphic encoding module A3 is used to perform graph isomorphic encoding on the individualized brain function connectivity map and the mental illness prototype cue map to obtain individual graph representation vectors and cue graph representation vectors; including: S31. The graph isomorphic encoding module A3 includes a graph isomorphic network encoder, which is composed of multiple graph isomorphic convolutional layers and pooling layers connected together.

[0066] In this embodiment, the graph isomorphic network encoder includes three graph isomorphic convolutional layers and one global summing pooling layer, with each layer connected in series. The output of the previous layer serves as the input of the next layer. The input of each graph isomorphic convolutional layer is the node feature matrix of the previous layer, and the input of the first layer is the initial node feature matrix of the individualized brain functional connectivity map or the mental illness prototype cue map.

[0067] The following operations are performed inside each graph isomorphic convolutional layer: For each node in the graph, the feature vectors of its neighboring nodes are aggregated, the aggregated result is concatenated with the node's own features, and then a multilayer perceptron is used to perform a nonlinear transformation on the concatenated features to output the updated node feature vector.

[0068] The aforementioned multilayer perceptron contains two fully connected layers. The first fully connected layer is followed by batch normalization and a linear rectified activation function, while the second fully connected layer is followed only by batch normalization.

[0069] The number of output channels of the three graph isomorphic convolutional layers are set to 64, 128 and 256 respectively, that is, the feature vector dimension of the output node of the first layer is 64, the feature vector dimension of the output node of the second layer is 128, and the feature vector dimension of the output node of the third layer is 256.

[0070] After the third graph isomorphic convolutional layer, a global summing pooling layer is connected to sum the feature vectors of all nodes in the graph element by element along the node dimension, resulting in a graph-level representation vector of fixed dimension. The vector dimension is equal to the number of output channels of the third graph isomorphic convolutional layer, i.e., 256 dimensions. This graph-level representation vector is the graph representation vector output by the graph isomorphic network encoder.

[0071] S32. Input the individualized brain function connectivity map into the graph isomorphic network encoder to obtain the individual graph representation vector.

[0072] First, construct the initial feature matrix of the nodes in the individualized brain functional connectivity map. The initial feature matrix of the nodes is as follows: OK A real matrix of columns, where This represents the number of nodes in a personalized brain functional connectivity graph. The initial feature dimensions for each node are as follows: For the brain region corresponding to each node, calculate the mean, variance, and skewness of the mean blood oxygenation level dependent signal time series. Concatenate these three statistics with the node's own degree centrality into a feature vector. The degree centrality is the number of connection edges of the node in the individualized brain functional connectivity graph.

[0073] Secondly, the initial feature matrix of the nodes and the adjacency matrix of the individualized brain functional connectivity graph are input into the graph isomorphic network encoder. The first graph isomorphic convolutional layer of the graph isomorphic network encoder receives the initial feature matrix of the nodes and the adjacency matrix. After feature extraction and transformation by three graph isomorphic convolutional layers, the third graph isomorphic convolutional layer outputs a... A node embedding matrix with 256 rows and 256 columns.

[0074] Subsequently, the node embedding matrix is ​​input into the global summing pooling layer, and all node embedding vectors are summed element-wise along the node dimension to obtain a 256-dimensional vector, which is the individual graph representation vector. This vector integrates the whole-brain topology and node feature information of the individualized brain functional connectivity map, providing an individualized graph-level feature representation for subsequent graph cue matching.

[0075] S33. Input the mental illness prototype cue graph into the same graph isomorphic network encoder to obtain the cue graph representation vector.

[0076] For each mental illness prototype hint graph stored in the disease prototype hint graph library module A2, construct its initial node feature matrix. For the mental illness prototype hint graph, the initial node features are constructed as follows: use the brain region labels of each brain region in the node set in the preset brain region template as node identifiers, and use the average value of the edge weights corresponding to the node in the edge weight matrix as the initial feature value of the node. If a node has only one edge connected to other nodes in the graph, then use the edge weight of that edge directly as the initial feature value. Input the initial node feature matrix and the adjacency matrix of the mental illness prototype hint graph into the graph isomorphic network encoder.

[0077] It should be noted that the graph isomorphic network encoder used is the same as the graph isomorphic network encoder used in step S32 to encode the individualized brain functional connectivity map.

[0078] The graph isomorphic network encoder performs the same processing flow as step S32 on the initial feature matrix and adjacency matrix of the input mental illness prototype cue graph: after layer-by-layer feature extraction and transformation through three layers of graph isomorphic convolutional layers, the third layer of graph isomorphic convolutional layer outputs the node embedding matrix of the mental illness prototype cue graph, and then the global summing pooling layer sums the node embedding vector element by element to finally obtain a 256-dimensional vector, which is the cue graph representation vector corresponding to the mental illness prototype cue graph.

[0079] Following the above method, prototype cue maps for depression, bipolar disorder, and schizophrenia are encoded to obtain three cue map representation vectors, which are then used by the subsequent graph cue matching module A4.

[0080] It should be noted that since the individualized brain function connectivity map and the mental illness prototype cue map are encoded into the same 256-dimensional representation space through the same graph isomorphic network encoder, the two are comparable in this space, which provides a technical basis for subsequent attention similarity calculation.

[0081] In this embodiment, the graph cue matching module A4 is used to calculate the similarity score between the individual graph representation vector and the cue graph representation vector, and to obtain an adaptive cue fusion weight based on the similarity score; including: S41. Use the individual graph representation vector as the query vector and the hint graph representation vector as the key vector to calculate the similarity score.

[0082] In this embodiment, the individual graph representation vector and three cue graph representation vectors output by the graph isomorphism encoding module A3 are obtained; the individual graph representation vector is denoted as... It has 256 dimensions and serves as the query vector. It is used to actively retrieve and match various disease prototype patterns; the three cue graph representation vectors are denoted as follows: , , Corresponding to prototype clue diagrams for depression, bipolar disorder, and schizophrenia, each with a dimension of 256, serving as key vectors. , is the object being queried and matched.

[0083] Calculating the query vector With a certain key vector When scoring similarity, an asymmetric random masking strategy is used to reinforce the role difference between the query and the key: for the query vector Randomly generate a binary mask vector The mask retention rate is set to 0.8. and The query vector after the mask is obtained by multiplying element by element. ; for key vectors Randomly generate another independent binary mask vector The mask retention rate is also set to 0.8. and The key vector after the mask is obtained by element-wise multiplication. .because and It is generated independently and randomly. and What is preserved is the information of different subsets in the representation vector. The above asymmetric mask allows the query vector and the key vector to play different roles in the calculation, which enhances the robustness of the matching.

[0084] Furthermore, this masking operation is repeated 5 times, generating different random mask vector pairs each time, resulting in 5 sets of masked representation vector pairs. For each set of masked representation vector pairs, a scaled dot product attention score, dominated by the query vector, is calculated as the attention score. This calculation is based on the query vector... Actively match key vectors The information requirements of the query vectors dominated the similarity evaluation process.

[0085] After obtaining 5 attention scores, their arithmetic mean is taken as the query vector. With the key vector The similarity score is obtained by performing the above-mentioned asymmetric random masking and attention score mean calculation process on the three key vectors respectively, and obtaining three similarity scores, which represent the degree of matching similarity between the subject's individualized brain functional connectivity map and the prototype cue maps of three mental illnesses: depression, bipolar disorder and schizophrenia.

[0086] S42. Normalize the similarity score to obtain the adaptive prompt fusion weight.

[0087] In this embodiment, after obtaining three similarity scores, the three similarity scores are combined into a score vector, and the score vector is normalized using the softmax function to obtain three adaptive cue fusion weights. , , The constraints are satisfied that the sum is 1, and the value of each weight is in the open interval (0,1).

[0088] It should be noted that the three adaptive cue fusion weights represent the degree of matching between the subject's individualized brain functional connectivity map and the prototype cue maps of depression, bipolar disorder, and schizophrenia, respectively. The larger the weight value, the more similar the subject's brain functional connectivity pattern is to the typical abnormal connectivity pattern of the disease.

[0089] In this embodiment, the cue edge weight modulation module A5 is used to generate an individualized modulation matrix based on the adaptive cue fusion weights, and to obtain a modulated brain functional connectivity map by combining the individualized brain functional connectivity map; including: S51. Using the adaptive cue fusion weight as a weighting coefficient, the edge weight matrix of the mental illness prototype cue graph is weighted and summed to obtain the individualized modulation matrix.

[0090] In this embodiment, the three adaptive cue fusion weights output by the graph cue matching module A4 are obtained and denoted as follows: , , This corresponds to depression, bipolar disorder, and schizophrenia.

[0091] Simultaneously, the edge weight matrix of the depression prototype cue graph is extracted from the disease prototype cue graph library module A2. Edge weight matrix of the prototype cue graph of bipolar disorder Edge weight matrix of the schizophrenia prototype cue graph The edge weight matrices are all OK A square matrix of columns, in which The total number of brain regions defined for the preset brain region template.

[0092] It should be noted that the node sets and edge sets of different disease prototype cue maps differ. For brain region pairs that are not included in the edge set of a certain disease prototype cue map, the corresponding edge weight matrix element value is zero.

[0093] Furthermore, using the adaptive cue fusion weights as weighting coefficients, the three edge weight matrices are weighted and summed to obtain the individualized modulation matrix, satisfying the following relationship:

[0094] in, For individualized modulation matrices, For adaptive cue fusion weights corresponding to depression, This is the edge weight matrix of the depression prototype cue graph. For adaptive cue fusion weights corresponding to bipolar disorder, The edge weight matrix is ​​the prototype cue graph for bipolar disorder. For the adaptive cue fusion weights corresponding to schizophrenia, This is the edge weight matrix of the prototype cue graph for schizophrenia.

[0095] It should be noted that the individualized modulation matrix is ​​also... OK The matrix is ​​a square matrix, where each element represents the modulation intensity of the corresponding brain region connection in the subject's individualized disease pattern. Since the adaptive cue fusion weight is calculated based on the similarity between the graph representation vector of the subject's individualized brain functional connectivity map and the graph representation vector of each disease prototype cue map, the adaptive cue fusion weight is different for different subjects, and therefore the generated individualized modulation matrix is ​​also different, thus realizing the individualized modulation of the edge weight.

[0096] S52. Traverse each connection edge of the individualized brain functional connectivity graph, multiply the edge weight value of the connection edge with the corresponding element value of the individualized modulation matrix element by element to obtain the modulation edge weight, and form the modulation brain functional connectivity graph based on the modulation edge weight.

[0097] In this embodiment, the adjacency matrix of the individualized brain functional connectivity map is obtained, and the adjacency matrix is... OK A square matrix of columns, where the first column is... Line 1 Column elements Indicates the first The brain region and the first The functional connectivity strength value between brain regions is set as follows: if there is no connection between two brain regions, the corresponding element value is zero.

[0098] Iterate through each connection edge in the individualized brain functional connectivity graph. For each connection edge, use the individualized modulation matrix... Extract the corresponding position element value As the modulation coefficient of this connection edge, the edge weight value of the connection edge is... With modulation coefficient Perform element-wise multiplication, and use the product as the modulation edge weight, denoted as . After traversing and multiplying each element of the adjacency matrix, the set of modulation edge weights is obtained, forming the adjacency matrix of the modulation brain functional connectivity graph. .

[0099] Furthermore, the modulated brain functional connectivity map uses the same brain regions as nodes and modulated edge weights as edge weights. Its topology is completely consistent with the individualized brain functional connectivity map, except that the edge weight values ​​are adjusted according to the individualized modulation matrix.

[0100] During modulation, if the modulation coefficient corresponding to a certain connection edge If the modulation coefficient is greater than 1, the edge weight of the connection is amplified, indicating that the connection is enhanced in the subject's disease pattern; if the modulation coefficient is greater than 1, the edge weight of the connection is amplified. If the value is less than 1, the edge weight of the connection is reduced, indicating that the connection is suppressed; if the modulation coefficient is less than 1, the edge weight of the connection is reduced, indicating that the connection is suppressed. If the value is 0, then the connecting edge is completely suppressed.

[0101] In this embodiment, the graph convolution module A6 is used to obtain a discriminant node embedding matrix based on the modulated brain functional connectivity graph, and to obtain a probability distribution for auxiliary diagnosis of mental illness based on the discriminant node embedding matrix; including: S61, the graph convolutional integration module A6 includes a graph convolutional network, an average pooling layer, and a fully connected layer.

[0102] In this embodiment, the graph convolutional network consists of three graph convolutional layers connected in series. Each graph convolutional layer is used to perform neighborhood message passing and nonlinear transformation on the input node features. The number of input channels of the first graph convolutional layer is the dimension of the initial node features, and the number of output channels is set to 64. The number of input channels of the second graph convolutional layer is 64, and the number of output channels is set to 128. The number of input channels of the third graph convolutional layer is 128, and the number of output channels is set to 256.

[0103] Each graph convolutional layer performs the following operations: First, matrix multiplication is performed between the input node feature matrix and the adjacency matrix of the modulated brain functional connectivity graph to achieve weighted aggregation of the features of each node's neighboring nodes; then, the aggregation result is multiplied by the learnable weight matrix of that layer for linear transformation; finally, a nonlinear mapping is performed on the linear transformation result through a linear rectified activation function to obtain the node embedding matrix output by that layer. The output node embedding vectors of the three graph convolutional layers have dimensions of 64, 128, and 256, respectively.

[0104] The average pooling layer is connected after the graph convolutional network, and its input is the discriminative node embedding matrix output from the third graph convolutional layer, with dimension O(log n). 256 rows and 256 columns, of which The number of nodes; the average pooling layer performs global average pooling on the discriminative node embedding matrix along the node dimension, and... The 256-dimensional embedding vectors of each node are compressed into a 256-dimensional full-graph feature vector.

[0105] The fully connected layer is connected after the average pooling layer. Its input is the feature vector of the whole image, and its output is a 4-dimensional vector. The four dimensions correspond to the four categories of healthy control, depression, bipolar disorder and schizophrenia, respectively. The weight matrix of the fully connected layer is a 256-row, 4-column real number matrix, the bias vector is a 4-dimensional real number vector, and the output is mapped by the softmax function to obtain the predicted probabilities of the four categories.

[0106] S62. Input the adjacency matrix and initial feature matrix of the modulated brain functional connectivity graph into the graph convolutional network to obtain the discriminative node embedding matrix.

[0107] First, obtain the adjacency matrix of the modulated brain functional connectivity map. and node initial feature matrix Adjacency matrix for OK A real matrix of columns, where Let be the number of nodes in the brain region, and be the in the matrix. Line 1 Column elements Modulation edge weights; initial feature matrix of nodes for OK The columns are real-valued matrices, and the initial features of each node are the mean, variance, and skewness of the mean blood oxygen level of the brain region corresponding to that node in the signal time series, as well as the degree centrality of that node in the modulated brain functional connectivity map. This represents the initial feature dimension.

[0108] Secondly, the adjacency matrix and node initial feature matrix The input is the first graph convolutional layer. For each node, the first graph convolutional layer performs operations based on the adjacency matrix. The corresponding edge weights are used to sum the initial features of all their neighboring nodes using a weighted sum. This weighted sum is then linearly transformed through the learnable weight matrix of the first layer, and finally, after passing through a linear rectified activation function, the first-layer node embedding matrix is ​​output. , dimension Rows 64 columns.

[0109] Next, the first-layer node embedding matrix and adjacency matrix The second graph convolutional layer takes a common input, performs neighborhood message passing and feature transformation in the same way, and outputs the second-layer node embedding matrix. , dimension Rows 128 columns.

[0110] Finally, the second-layer node embedding matrix and adjacency matrix The common input is the third graph convolutional layer, which outputs the third-layer node embedding matrix in the same manner as described above. , dimension 256 rows, third-level node embedding matrix This is the discriminative node embedding matrix. Each row of this matrix corresponds to a 256-dimensional embedding vector of a brain region node. The embedding vector integrates the initial statistical features of the node with the global topological information of the brain after being aggregated layer by layer through multi-level graph convolution.

[0111] S63. The discriminative node embedding matrix is ​​subjected to global average pooling through the average pooling layer to obtain the full-image feature vector.

[0112] The average pooling layer performs a global average pooling operation along the node dimensions of the discriminative node embedding matrix. The operation involves iterating through all nodes in each of the 256 feature dimensions. The embedding value of each node in this dimension is calculated. The arithmetic mean of the embedded values ​​is used as the value of the full image feature vector in that dimension. After calculating the average value of each of the 256 feature dimensions, the 256 average values ​​are concatenated in the original dimensional order to form a 256-dimensional full image feature vector.

[0113] S64. Input the full-image feature vector into the fully connected layer to obtain the probability distribution of the auxiliary diagnosis of mental illness.

[0114] The fully connected layer receives the feature vector of the entire image as input; the fully connected layer includes a learnable weight matrix and a learnable bias vector, where the number of rows of the weight matrix is ​​equal to the dimension of the input features (256), the number of columns is equal to the number of target classification categories (4), and the bias vector has a length of 4. The four classification categories correspond to healthy controls, depression, bipolar disorder, and schizophrenia, respectively.

[0115] It should be noted that the weight matrix and bias vector of the fully connected layer are optimized through end-to-end supervised learning during the system training phase using the backpropagation algorithm and the cross-entropy loss function, which enables the system to learn the mapping relationship from the feature vector of the whole image to the correct disease category based on the training data.

[0116] Specifically, the fully connected layer performs a linear transformation on the input full-image feature vector, multiplies the full-image feature vector with the weight matrix, and adds a bias vector to obtain a 4-dimensional original output vector. Each element in this vector represents the original score of the subject belonging to the corresponding category. The higher the score, the greater the probability that the subject belongs to that category.

[0117] Furthermore, a normalized exponential function is used to perform probability mapping on the 4-dimensional original output vector, converting the original scores into an interpretable probability distribution. The process is as follows: First, the exponential values ​​of the four original scores are calculated with the natural constant e as the base; then, the four exponential values ​​are summed to obtain the total exponential sum; finally, the exponential value of each category is divided by the total exponential sum to obtain four probability values ​​ranging from 0 to 1 with a sum of 1, representing the probability that the subject belongs to a healthy control, depression, bipolar disorder, and schizophrenia, respectively.

[0118] The above four probability values ​​constitute the probability distribution for auxiliary diagnosis of mental illness. The category corresponding to the maximum value among the four probability values ​​is taken as the auxiliary diagnosis result of the subject's mental illness.

[0119] In summary, the present invention provides a mental illness auxiliary diagnostic system based on an adaptive graph cueing strategy. This system constructs an individualized brain functional connectivity graph and maintains multiple prototype cue graphs for mental illnesses, including node sets, edge sets, and learnable edge weight matrices. The individual graph and each cue graph are input into a graph isomorphic network encoder to obtain graph representation vectors. An attention mechanism is used to calculate adaptive cue fusion weights, and the edge weight matrices of each cue graph are weighted and summed to generate an individualized modulation matrix. This individualized modulation matrix is ​​then multiplied element-wise with the edge weights of the individual graph to obtain a modulated brain functional connectivity graph. Finally, a graph convolutional integral network outputs the auxiliary diagnostic probability. The method of this invention is easy to understand, computationally simple, requires minimal workload, and is convenient for practical application, providing a theoretical foundation and technical support for the further development of the field of medical auxiliary diagnostic technology.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A mental illness auxiliary diagnostic system based on an adaptive graph cueing strategy, characterized in that, include: The brain functional network construction module is used to preprocess the resting-state functional magnetic resonance imaging of subjects and calculate the functional connectivity strength values ​​of brain regions in order to construct an individualized brain functional connectivity map. The Disease Prototype Hint Library Module is used to store various mental illness prototype hint graphs, which are composed of a node set, an edge set, and an edge weight matrix. The graph isomorphic encoding module is used to perform graph isomorphic encoding on the individualized brain function connectivity map and the mental illness prototype cue map to obtain individual graph representation vectors and cue graph representation vectors; The graph suggestion matching module is used to calculate the similarity score between the individual graph representation vector and the suggestion graph representation vector, and to obtain adaptive suggestion fusion weights based on the similarity score; The cue edge weight modulation module is used to generate an individualized modulation matrix based on the adaptive cue fusion weight, and to obtain a modulated brain functional connectivity map by combining the individualized brain functional connectivity map. The graph integral module is used to obtain a discriminant node embedding matrix based on the modulated brain functional connectivity graph, and to obtain a probability distribution for auxiliary diagnosis of mental illness based on the discriminant node embedding matrix. The various mental illness prototype illustrations include: The prototype diagrams of mental illnesses include prototype diagrams of depression, bipolar disorder, and schizophrenia. The mental illness prototype cue graph consists of a node set, an edge set, and an edge weight matrix, including: Based on the resting-state functional magnetic resonance imaging, brain functional connectivity matrices were extracted from the mental illness patient group and the healthy control group, respectively. Statistical difference analysis was performed on the brain functional connectivity matrices of the two groups to screen brain functional connectivity edges; The node set is formed by using the brain regions corresponding to the brain function connection edges; The edge set is formed by connecting the brain functions; The inter-group difference effect size is obtained as the initial edge weight to construct the edge weight matrix; The step of generating a personalized modulation matrix based on the adaptive cue fusion weights and obtaining a modulation brain functional connectivity map by combining the personalized brain functional connectivity map includes: The adaptive cue fusion weights are used as weighting coefficients to sum the edge weight matrix of the mental illness prototype cue graph to obtain the individualized modulation matrix. Each edge of the individualized brain functional connectivity graph is traversed, and the edge weight value of the edge is multiplied element by element by the corresponding element value of the individualized modulation matrix to obtain the modulation edge weight. The modulation brain functional connectivity graph is formed based on the modulation edge weight.

2. The mental illness auxiliary diagnostic system based on adaptive graph cueing strategy according to claim 1, characterized in that, The process of preprocessing the resting-state functional magnetic resonance imaging (fMRI) images of the subjects and calculating the functional connectivity strength values ​​of brain regions to construct an individualized brain functional connectivity map includes: The resting-state functional magnetic resonance imaging is subjected to head motion correction, spatial normalization, bandpass filtering, and smoothing to obtain standardized image data. Based on a preset brain region template, the average blood oxygen level dependent signal time series of the brain region is extracted from the standardized image data; Select any two brain regions, calculate the Pearson correlation coefficient between the time series of the mean blood oxygen level dependent signal, and obtain the functional connectivity strength value through the Pearson correlation coefficient. Obtain the absolute value of the functional connectivity strength value, and filter brain region connections that exceed a preset threshold to construct the individualized brain functional connectivity map.

3. The mental illness auxiliary diagnostic system based on adaptive graph cueing strategy according to claim 2, characterized in that, The individualized brain connectivity map includes: The individualized brain functional connectivity graph uses the selected brain regions as nodes, the connections between the brain regions as edges, and the functional connectivity strength values ​​as edge weights.

4. The mental illness auxiliary diagnostic system based on adaptive graph cueing strategy according to claim 1, characterized in that, The step of performing graph isomorphic encoding on the individualized brain function connectivity map and the mental illness prototype cue map to obtain individual graph representation vectors and cue graph representation vectors includes: The graph isomorphic encoding module includes a graph isomorphic network encoder, which is composed of multiple graph isomorphic convolutional layers and pooling layers connected together. The individualized brain function connectivity map is input into the graph isomorphic network encoder to obtain the individual graph representation vector; The mental illness prototype cue graph is input into the same graph isomorphic network encoder to obtain the cue graph representation vector.

5. The mental illness auxiliary diagnostic system based on adaptive graph cueing strategy according to claim 1, characterized in that, The step of calculating the similarity score between the individual graph representation vector and the cue graph representation vector, and obtaining the adaptive cue fusion weight based on the similarity score, includes: The individual graph representation vector is used as the query vector, and the hint graph representation vector is used as the key vector to calculate the similarity score; The similarity scores are normalized to obtain the adaptive cue fusion weights.

6. The mental illness auxiliary diagnostic system based on adaptive graph cueing strategy according to claim 5, characterized in that, The step of using the individual graph representation vector as the query vector and the hint graph representation vector as the key vector to calculate the similarity score includes: The query vector and the key vector are randomly masked multiple times to obtain multiple sets of masked representation vector pairs. Calculate the attention score for the representation vector pair, and obtain the mean of the attention scores as the similarity score.

7. The mental illness auxiliary diagnostic system based on adaptive graph cueing strategy according to claim 1, characterized in that, The step of obtaining a discriminative node embedding matrix based on the modulated brain functional connectivity map, and obtaining a probability distribution for auxiliary diagnosis of mental illness based on the discriminative node embedding matrix, includes: The graph convolutional integration module includes a graph convolutional network, an average pooling layer, and a fully connected layer. The adjacency matrix and initial node feature matrix of the modulated brain functional connectivity graph are input into the graph convolutional network to obtain the discriminative node embedding matrix; The discriminative node embedding matrix is ​​subjected to global average pooling through the average pooling layer to obtain the full-image feature vector; The full-image feature vector is input into the fully connected layer to obtain the probability distribution for the auxiliary diagnosis of mental illness.

Images