A multicenter diagnostic system and method for depression based on spatiotemporal heterogeneous feature fusion and domain adversarial graph networks

Abstract

This invention discloses a multi-center depression diagnostic system and method based on spatiotemporal heterogeneous feature fusion and domain adversarial graph networks, belonging to the field of depression identification. It solves the core technical problem of catastrophic collapse of model generalization ability caused by device heterogeneity in existing diagnostic methods. The system includes a spatiotemporal heterogeneous feature extraction module, a cross-modal deep fusion module, a prior static graph construction module, a high-order graph information aggregation and pooling module, and a domain adversarial and disease classification module. The method employs: lightweight extraction and latent space alignment of spatiotemporal heterogeneous features; cross-modal bidirectional cross-attention and nonlinear bottleneck fusion; static topological graph construction based on medical prior knowledge; graph convolutional information aggregation and global attention pooling; and domain adversarial decoupling and MDD classification. By establishing a zero-sum game mechanism for feature extraction and site discrimination, this invention eliminates the batch effect caused by multi-center data, thereby achieving high-precision diagnosis in cross-hospital and new site scenarios.

Description

Technical Field

[0001] This invention belongs to the field of depression identification, specifically relating to a multicenter depression diagnostic system and method based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network. Background Technology

[0002] Major Depressive Disorder (MDD) is a severe mental illness with a high relapse and disability rate, causing a huge health burden and economic loss worldwide. Traditional clinical diagnosis of depression relies heavily on interviews between doctors and patients and structured scale assessments. This diagnostic approach, based on subjective symptom reports, is easily affected by individual cognitive biases and environmental changes. With the introduction of computer-aided diagnostic technology, finding objective and quantitative biomarkers to assist clinicians in the accurate identification and early screening of depression has become a core demand that urgently needs to be addressed in the interdisciplinary field of psychiatry and medical artificial intelligence.

[0003] The development of multimodal magnetic resonance imaging (MRI) technology has provided a solid data foundation for the objective assessment of depression. Resting-state functional magnetic resonance imaging (rs-fMRI) can capture the dynamic temporal fluctuations of oxygen-dependent (BOLD) signals in the brain, reflecting abnormal rhythmic features of functional connectivity between different brain regions. Structural magnetic resonance imaging (sMRI, such as T1-weighted imaging) provides high-resolution brain anatomical information, used to quantify static physical indicators reflecting cortical morphological atrophy, such as gray matter volume. Fusing the dynamic temporal features provided by fMRI with the static spatial structural priors provided by sMRI can comprehensively reveal the potential "functional-structural disharmony" neuropathological mechanisms of depression, providing crucial underlying support for constructing high-precision computer-aided diagnostic models.

[0004] However, when using multimodal MRI data to train deep learning models for real-world clinical applications, the "batch effect" problem persists across multiple sites. Since the amount of MDD samples collected by a single hospital is typically very small, researchers must pool data from multiple hospitals for training to satisfy the data hunger of deep learning. However, there are significant differences in the brand of MRI equipment, static magnetic field strength, number of radiofrequency coil channels, and scanning sequence parameters among different medical institutions. This physical heterogeneity introduces strong non-biological noise into the high-dimensional feature space of neuroimaging. This non-independent, identically distributed offset caused by equipment differences often masks the weak pathological signals of the MDD itself. Summary of the Invention

[0005] The purpose of this invention is to provide a multicenter depression diagnostic system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network, in order to solve the core technical problem that existing computer-aided diagnosis models for MDD suffer catastrophic collapse of generalization ability due to equipment heterogeneity when conducting joint diagnosis across hospitals.

[0006] Another objective of this invention is to provide a multicenter depression diagnosis method based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network. On the basis of lightweight extraction and fusion of multimodal MRI features, a gradient inversion layer is innovatively introduced at the bottom layer of the network. By establishing a zero-sum game mechanism for feature extraction and site discrimination, the batch effect caused by multicenter data is eliminated, thereby achieving high-precision diagnosis in cross-hospital and new site scenarios.

[0007] The technical solution of this invention is: (one) A multicenter depression diagnostic system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network includes a spatiotemporal heterogeneous feature extraction module, a cross-modal deep fusion module, a prior static graph construction module, a high-order graph information aggregation and pooling module, and a domain adversarial and disease classification module. The spatiotemporal heterogeneous feature extraction module contains two parallel lightweight extraction branches. One branch consists of a temporal convolutional network (TCN) and a linear projection layer without activation functions, which receives the preprocessed dynamic temporal signals from fMRI. The other branch consists of only a linear projection layer without activation functions, which receives the static gray matter volume scalar from sMRI. This module maps spatiotemporal features with heterogeneous physical meaning to a latent space of uniform dimension. The cross-modal deep fusion module is connected after the spatiotemporal heterogeneous feature extraction module. It contains a bidirectional cross-attention unit, a channel splicing unit, a nonlinear bottleneck layer, and a single-layer cross-modal nonlinear fusion module. The working principle of this module is to generate complementary enhancement features through mutual query between fMRI and sMRI, and to force nonlinear algebraic coupling in the deep network to eliminate modal redundancy. The prior static graph construction module directly receives raw fMRI signals and contains a Pearson correlation calculator and a sparsity threshold pruner. Its working principle is to use medical priors to generate a static brain network topology framework with anatomical significance. The high-order graph information aggregation and pooling module receives node features output from the cross-modal deep fusion module and adjacency matrix (skeleton) output from the prior static graph construction module. Internally, it contains a two-layer graph convolutional network (GCN) and a global attention pooling layer (Global Attention Readout). The working principle of this module is to transmit pathological information on the whole brain topology and compress it into a highly concentrated global graph-level representation vector. The domain adversarial and disease classification module (zero-sum game core) consists of two parallel branches, one bright and one dark, connected to the output of the high-order graph information aggregation and pooling module (4). One branch (bright line) is the MDD master classifier (MLP, implemented by a multilayer perceptron consisting of two fully connected networks), used to output the disease probability; the other branch (dark line) contains a gradient inversion layer (GRL) and a site discriminator, used to identify the hospital and site from which the data originates. The working principle of this module is to generate a negative gradient through GRL in backpropagation, which forces the front-end feature extraction network to completely remove the "batch watermark" of the scanning device while optimizing the classification performance.

[0008] Based on lightweight extraction and fusion of multimodal MRI features, the novel architecture innovatively introduces a gradient inversion layer (GRL) at the bottom layer of the network. By establishing a zero-sum game mechanism between feature extraction and site discrimination, it eliminates the batch effect caused by multi-center data, thereby achieving high-precision diagnosis in cross-hospital and new site scenarios. (two) A multicenter depression diagnosis method based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network, employing the aforementioned multicenter depression diagnosis system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network, includes the following steps: S1. Lightweight extraction and latent space alignment of spatiotemporal heterogeneous features: S1.1, fMRI dynamic stream processing: The spatiotemporal heterogeneous feature extraction module inputs the BOLD temporal signal matrix of 116 brain regions divided by preprocessing and AAL atlas into a temporal convolutional network (TCN). After feature extraction, without going through a nonlinear activation function, the temporal features are directly mapped to the feature space through a linear projection layer without activation function and weight sharing, reducing the dimension and outputting a dynamic spatial feature matrix. S1.2 sMRI Static Stream Processing: The spatiotemporal heterogeneous feature extraction module directly inputs the highly condensed gray matter volume scalars of 116 brain regions, which are preprocessed by voxel-based morphological analysis (VBM), into a linear projection layer without activation functions. This linear projection layer without activation functions directly upscales the physical volume scalars and aligns them to the static spatial feature matrix through a learnable parameter matrix. S2, Fusion of cross-modal bidirectional cross-attention and nonlinear bottleneck: S2.1, Bidirectional Interactive Inspection: The cross-modal deep fusion module inputs the obtained fMRI and sMRI features into the bidirectional cross-attention unit. The fMRI features are used as queries to inspect the key and value pairs of the sMRI, generating dynamic augmented features Z containing structural constraints. f←sSimilarly, sMRI features are used as keys and values ​​to examine fMRI, generating structural enhancement features Z that incorporate temporal evolution. s←f ; S2.2, Nonlinear bottleneck dimensionality reduction: The channel splicing unit will generate Z... f←s With Z s←f Concatenation is performed along the channel dimension to form a feature redundancy concatenated matrix. This matrix is ​​then fed into a single-layer cross-modal nonlinear fusion module with a fixed parameter structure. This module sequentially includes: a fully connected linear layer, a batch normalization layer (BatchNorm) to prevent internal covariate shift, and a leaky ReLU with nonlinear manifold distortion, thus creating a nonlinear bottleneck layer. Through the physical compression of this bottleneck layer, deep algebraic coupling of heterogeneous features is achieved, outputting the final unified node feature matrix X. fusion ; S3. Static topological graph based on prior medical knowledge: The prior static graph construction module directly reuses the initial input fMRI raw time-series signal matrix to calculate the Pearson correlation coefficient between pairwise time series of 116 brain regions. In order to filter out high-frequency noise and generate a sparse skeleton with anatomical interpretability, sparsity pruning is performed using a preset fixed empirical threshold. After pruning, a global static prior functional connectivity adjacency matrix A is generated. This adjacency matrix A constitutes the static brain network topological skeleton, providing an interpretable topological structure with medical prior constraints for subsequent high-order graph convolution information aggregation. S4. Graph Convolutional Information Aggregation and Global Attention Pooling: The high-order graph information aggregation and pooling module will refine the unified node feature matrix X. fusion As a node attribute, the static prior function connects the adjacency matrix A as the topology and is synchronously input into a two-layer graph convolutional network (GCN). The pathological features of each brain region are strictly aggregated into second-order neighborhoods along the topological skeleton of the static brain network. The brain region features updated by the graph convolutional network are connected to the global attention pooling layer for global pooling (Readout). The global attention pooling layer learns an attention weight vector and dynamically weights and sums the 116 brain regions, compressing the local node features into a global representation with extremely high information density, and outputting the final graph-level representation vector z representing the patient's overall brain pathological state. S5, Domain Adversarial Decoupling and MDD Classification: S5.1 Probability Prediction: The domain adversarial and disease classification module inputs the ultimate graph-level representation vector z into the MDD main classifier, outputs the accurate predicted probability of whether the user has major depressive disorder through the Softmax activation function, and calculates the cross-entropy disease classification loss L. cls L cls It reflects the gap between the predicted probability and the patient's actual disease status; S5.2 Identifying Data Source Hospitals and Sites: The final graph-level representation vector z is simultaneously passed through a gradient inversion layer (GRL) and input to the site discriminator. The site discriminator outputs the probability distribution of each sample belonging to different hospitals and sites, thereby completing site classification and calculating the domain classification loss L. domain The core purpose of this classification is to improve the site classification accuracy as much as possible by making the site discriminator compete with the feature extractor whose gradient is reversed by the gradient inversion layer, thus forcing the backbone network to learn disease features that are irrelevant to the site. S5.3, Game Theory Mechanism: During forward propagation of the network, the gradient reversal layer behaves as an identity transformation; while during backpropagation to update the weights, the gradient reversal layer will change L... domain The returned gradient is forcibly negative and multiplied by the adversarial penalty coefficient -λ; to ensure the stability of the network in the early stages of training, the preferred update strategy for parameter λ is to adaptively increase with each training epoch: , where p is the current training progress percentage. This mathematical mechanism forces the backbone feature extraction network at the front end of the system to form a zero-sum game with the site discriminator, ultimately completely erasing the "physical watermark" of the site in the latent space, achieving zero-sample cross-site generalization diagnosis.

[0010] At the level of multimodal feature fusion, when processing multimodal data such as fMRI (functional MRI) and sMRI (static anatomical MRI), existing technologies mostly adopt strategies of "early splicing" (flattening features and directly concatenating them at the input layer) or "late fusion" (training two classifiers separately and then weighting the output probabilities to perform a simple combination of extracted latent features. The fused features have extremely high redundancy and cannot effectively capture the deep nonlinear interactions and complementary relationships between different modalities, easily missing key joint pathological biomarkers.

[0011] In addressing batch effect elimination in multicenter data, existing research primarily relies on traditional statistical harmonic reconciliation methods (with ComBat and its variants being the most typical) to address data distribution bias caused by cross-hospital equipment. These methods are typically performed independently after imaging feature extraction, attempting to calibrate and eliminate batch effects from data from different centers using linear models. However, traditional methods cannot completely remove the "physical watermark" of MRI instruments in deep, high-dimensional nonlinear spaces. When the model is deployed to entirely new target sites not involved in training, severe domain bias persists, often resulting in a precipitous drop in cross-center generalization accuracy.

[0012] To address the fundamental shortcomings of existing deep learning-based depression diagnostic techniques in areas such as shallow multimodal fusion and multi-site batch effects, this invention aims to overcome the heterogeneous semantic gap and achieve deep algebraic coupling and redundancy removal of cross-modal features by establishing a bidirectional cross-attention mechanism between fMRI and sMRI in the deep feature space. Simultaneously, a domain adversarial decoupling mechanism based on negative gradient game theory is introduced at the bottom layer of the end-to-end graph network training, forcing the model to actively remove physical batch watermarks from different MRI scanners. By collaboratively solving the two major technical challenges of "shallow splicing of multimodal features" and "post-hoc linear calibration of multi-site batch effects," this invention successfully extracts pure and domain-invariant "structure-function" joint pathological biomarkers, ultimately achieving high-precision and robust objective auxiliary diagnosis of multi-center MDD in real and complex clinical environments.

[0013] Compared with the prior art, the present invention has the following advantages: 1. The diagnostic system of this invention connects a site discriminator in parallel at the end of the backbone network and innovatively introduces a gradient inversion layer. This forces the diagnostic system to actively forget the field strength and sequence physical characteristics of different brands of MRI in a deep nonlinear manifold space, achieving absolute nonlinear decoupling between pure "MDD disease characteristics" and "multi-center equipment noise." This diagnostic system can be deployed directly with high precision without collecting calibration data for new hospitals. Existing post-hoc statistical calibration techniques such as ComBat can only handle linear batch effects and are completely disconnected from feature extraction in deep networks, resulting in extremely unstable generalization performance at target sites. This invention effectively avoids these problems.

[0014] 2. The diagnostic method of this invention adopts a bidirectional cross-attention mechanism, allowing fMRI and sMRI to query and guide each other. Subsequently, dimensionality reduction and redundancy removal are forced through a single-layer nonlinear bottleneck layer, which causes nonlinear algebraic coupling between the dynamic rhythm of the time series and the gray matter atrophy of the anatomical structure in the deep manifold space. This completely surpasses the traditional simple linear splicing and accurately extracts the core joint biomarkers of "structure-function disharmony" in depression. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structure of a multicenter depression diagnostic system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network according to the present invention.

[0016] In the diagram: 1- Spatiotemporal heterogeneous feature extraction module; 2- Cross-modal deep fusion module; 3- Prior static graph construction module; 4- Higher-order graph information aggregation and pooling module; 5- Domain adversarial and disease classification module. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] The full English names and Chinese definitions of abbreviations: MDD: Major Depressive Disorder.

[0019] fMRI: Functional Magnetic Resonance Imaging (used to extract rhythmic sequences of blood oxygenation activity in the brain over time).

[0020] sMRI: Structural Magnetic Resonance Imaging (used to extract high-precision three-dimensional anatomical and structural information of the brain).

[0021] BOLD: Blood Oxygenation Level Dependent (the underlying physical signal sequence in fMRI that reflects neuronal activity).

[0022] AAL: Automated Anatomical Labeling (a medical brain atlas used to divide the brain into 116 regions of interest).

[0023] TCN: Temporal Convolutional Network (a lightweight base model that uses dilated causal convolution to extract long-range temporal features).

[0024] GCN: Graph Convolutional Network (a core model for high-order cross-brain region information transfer based on node features and connected topology).

[0025] GRL: Gradient Reversal Layer (the core network layer that forcibly negates the gradient during backpropagation to achieve adversarial training).

[0026] MLP: Multi-Layer Perceptron, a fully connected network (the infrastructure used for final output disease probability prediction and site prediction).

[0027] like Figure 1 As shown, a multicenter depression diagnostic system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network includes a spatiotemporal heterogeneous feature extraction module 1, a cross-modal deep fusion module 2, a prior static graph construction module 3, a high-order graph information aggregation and pooling module 4, and a domain adversarial and disease classification module 5. The spatiotemporal heterogeneous feature extraction module 1 contains two parallel lightweight extraction branches. One branch consists of a temporal convolutional network (TCN) and a linear projection layer without activation functions, which receives the preprocessed dynamic temporal signals from fMRI. The other branch consists of only a linear projection layer without activation functions, which receives the static gray matter volume scalar from sMRI. The cross-modal deep fusion module 2 is connected after the spatiotemporal heterogeneous feature extraction module 1. It contains a bidirectional cross-attention unit, a channel splicing unit, a nonlinear bottleneck layer, and a single-layer cross-modal nonlinear fusion module. The prior static graph construction module 3 directly receives the raw fMRI signal and contains a Pearson correlation calculator and a sparsity threshold trimmer. The high-order graph information aggregation and pooling module 4 receives the node features output by the cross-modal deep fusion module 2 and the adjacency matrix (skeleton) output by the prior static graph construction module 3, which contains a two-layer graph convolutional network (GCN) and a global attention pooling layer (Global Attention Readout). The domain adversarial and disease classification module 5 (zero-sum game core) consists of two parallel branches, one bright and one dark, connected to the output of the high-order graph information aggregation and pooling module 4. One branch (bright line) is the MDD master classifier (MLP, implemented by a multilayer perceptron consisting of two fully connected networks), which is used to output the disease probability. The other branch (dark line) contains a gradient inversion layer (GRL) and a site discriminator, which is used to identify the hospital and site from which the data originates.

[0028] A multicenter depression diagnosis method based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network, employing the aforementioned multicenter depression diagnosis system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network, includes the following steps: S1. Lightweight extraction and latent space alignment of spatiotemporal heterogeneous features: S1.1, fMRI dynamic stream processing: The spatiotemporal heterogeneous feature extraction module 1 inputs the BOLD temporal signal matrix (input dimension set to 116×T, where T is the time step) of 116 brain regions divided by preprocessing and AAL atlas into a temporal convolutional network (TCN). In this embodiment, the kernel size of the TCN is preferably set to K=3, and a dilated causal convolution mechanism is adopted (the dilation factor is set to d={1,2,4} in sequence) to exponentially expand the receptive field to extract long-range temporal rhythm features. After feature extraction, without going through a nonlinear activation function, the temporal features are directly mapped to the feature space through a linear projection layer with shared weights and no activation function, reducing the dimension and outputting a dynamic spatial feature matrix with a fixed dimension of 116×128. S1.2, sMRI static flow processing: The spatiotemporal heterogeneous feature extraction module 1 takes the highly condensed gray matter volume scalars of 116 brain regions (input dimension is highly condensed 116×1) after voxel-based morphological analysis (VBM) preprocessing and directly inputs them into a linear projection layer without activation function (configured as Linear(in_features=1,out_features=128)). This linear projection layer without activation function directly upscales the 1-dimensional physical volume scalars and aligns them to a 116×128 static spatial feature matrix through a learnable parameter matrix.

[0029] S2, Fusion of cross-modal bidirectional cross-attention and nonlinear bottleneck: S2.1, Bidirectional Interactive Inspection: The cross-modal deep fusion module 2 inputs the obtained fMRI features and sMRI features (both 116×128) into the bidirectional cross-attention unit (attention heads h=8). The fMRI features are used as queries to inspect the key and value of the sMRI, generating a dynamic augmented feature Z containing structural constraints. f←s Similarly, sMRI features are used as keys and values ​​to examine fMRI, generating structural enhancement features Z that incorporate temporal evolution. s←f ; S2.2, Nonlinear bottleneck dimensionality reduction: The channel splicing unit will generate Z... f←s With Z s←fConcatenation is performed along the channel dimension to form a 116×256 feature redundancy concatenation matrix. This matrix is ​​then fed into a single-layer cross-modal nonlinear fusion module with a fixed parameter structure. This module sequentially includes: a fully connected linear layer that compresses features from 256 dimensions to 128 dimensions; a batch normalization layer (BatchNorm) to prevent internal covariate shift; and a leaky ReLU linear unit (with a negative half-axis slope parameter α set to 0.2) that introduces nonlinear manifold distortion, thus creating a nonlinear bottleneck layer. Through the physical compression of this bottleneck layer, deep algebraic coupling of heterogeneous features is achieved, outputting the final unified node feature matrix X. fusion (The final dimensions are 116×128).

[0030] S3. Static topological graph based on prior medical knowledge: The prior static graph construction module 3 directly reuses the initial input fMRI raw time-series signal matrix (116×T) to calculate the Pearson correlation coefficient between pairwise time series of 116 brain regions. In order to filter out high-frequency noise and generate a sparse skeleton with anatomical interpretability, sparsity pruning is performed using a preset fixed empirical threshold. In this embodiment, the preferred topology pruning parameter is to retain the edges with an absolute value of Pearson correlation coefficient |r|>0.4. After pruning, a global static prior functional connectivity adjacency matrix A (dimension 116×116) is generated. This adjacency matrix A constitutes the static brain network topology skeleton, providing an interpretable topology structure with medical prior constraints for subsequent high-order graph convolution information aggregation.

[0031] S4. Graph Convolutional Information Aggregation and Global Attention Pooling: The high-order graph information aggregation and pooling module 4 will refine the unified node feature matrix X. fusion As a node attribute, the static prior function connects the adjacency matrix A as the topology and is synchronously input into a two-layer graph convolutional network (GCN). In the information transmission configuration, the output dimension of the first layer GCN remains at 128, and the output dimension of the second layer GCN is compressed to 64. The pathological features of each brain region are strictly aggregated into second-order neighborhoods along the static brain network topology framework. The brain region features updated by the graph convolutional network are connected to the global attention pooling layer for global pooling (Readout). The global attention pooling layer learns a 116×1 attention weight vector and dynamically weights and sums the 116 brain regions, compressing the local node features into a global representation with extremely high information density, and outputting the final graph-level representation vector z (with a fixed dimension of 1×64) representing the patient's overall brain pathological state.

[0032] S5, Domain Adversarial Decoupling and MDD Classification: S5.1 Probability Prediction: Domain Adversarial and Disease Classification Module 5 inputs the ultimate graph-level representation vector z into the MDD main classifier. The main classifier is implemented by a multilayer perceptron consisting of two fully connected networks, with 32 hidden layer neurons and 2 output layers. It outputs the accurate predicted probability of whether the user has major depressive disorder through the Softmax activation function and calculates the cross-entropy disease classification loss L. cls L cls It reflects the gap between the predicted probability and the patient's actual disease status; S5.2 Identifying Data Source Hospitals and Sites: The ultimate graph-level representation vector z is simultaneously passed through a gradient inversion layer (GRL) and input to the site discriminator (whose output dimension is equal to the total number N of hospitals and sites in the multi-center dataset). sites The site discriminator outputs the probability distribution of each sample belonging to different hospitals and sites, thereby completing site classification and calculating the domain classification loss L. domain ; S5.3, Game Theory Mechanism: During forward propagation of the network, the gradient reversal layer behaves as an identity transformation; while during backpropagation to update the weights, the gradient reversal layer will change L... domain The returned gradient is forcibly negative and multiplied by the adversarial penalty coefficient -λ; to ensure the stability of the network in the early stages of training, the preferred update strategy for parameter λ is to adaptively increase with each training epoch: , where p is the current training progress percentage.

Claims

1. A multicenter depression diagnostic system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph networks, characterized in that: It includes a spatiotemporal heterogeneous feature extraction module (1), a cross-modal deep fusion module (2), a priori static graph construction module (3), a high-order graph information aggregation and pooling module (4), and a domain adversarial and disease classification module (5). The spatiotemporal heterogeneous feature extraction module (1) includes two parallel lightweight extraction branches. One branch is a temporal convolutional network and a linear projection layer without activation function, which receives the preprocessed fMRI dynamic temporal signal; the other branch is only a linear projection layer without activation function, which receives the sMRI static gray matter volume scalar. The cross-modal deep fusion module (2) is connected after the spatiotemporal heterogeneous feature extraction module (1), and contains a bidirectional cross-attention unit, a channel splicing unit, a nonlinear bottleneck layer and a single-layer cross-modal nonlinear fusion module. The prior static graph construction module (3) directly receives the raw fMRI signal and contains a Pearson correlation calculator and a sparsity threshold trimmer. The higher-order graph information aggregation and pooling module (4) receives the node features output by the cross-modal deep fusion module (2) and the adjacency matrix output by the prior static graph construction module (3), and contains a two-layer graph convolutional network and a global attention pooling layer. The domain adversarial and disease classification module (5) consists of two branches connected in parallel to the output of the high-order graph information aggregation and pooling module (4). One branch is the MDD master classifier, which is used to output the disease probability; the other branch contains a gradient inversion layer and a site discriminator, which is used to identify the hospital and site from which the data originates.

2. A multicenter diagnostic method for depression based on spatiotemporal heterogeneous feature fusion and domain adversarial graph networks, characterized in that: The multicenter depression diagnostic system based on spatiotemporal heterogeneous feature fusion and domain adversarial graph network as described in claim 1 includes the following steps: S1. Lightweight extraction and latent space alignment of spatiotemporal heterogeneous features: S1.1, fMRI dynamic flow processing: spatiotemporal heterogeneous feature extraction module (1) inputs the BOLD time-series signal matrix of 116 brain regions divided by preprocessing and AAL map into the temporal convolutional network. After feature extraction, without going through a nonlinear activation function, the temporal features are directly mapped to the feature space through a linear projection layer without activation function and the dimensions are reduced and the dynamic spatial feature matrix is ​​output. S1.2, sMRI static flow processing: the spatiotemporal heterogeneous feature extraction module (1) directly inputs the highly condensed gray matter volume scalars of 116 brain regions after voxel-based morphological analysis preprocessing into a linear projection layer without activation function. This linear projection layer without activation function directly upscales the physical volume scalars and aligns them to the static spatial feature matrix through a learnable parameter matrix. S2, Fusion of cross-modal bidirectional cross-attention and nonlinear bottleneck: S2.1, Bidirectional Interactive Inspection: The cross-modal deep fusion module (2) inputs the obtained fMRI features and sMRI features into the bidirectional cross-attention unit. The fMRI features are used as the keys and values ​​for querying the sMRI, generating dynamic enhancement features Z containing structural constraints. f←s ; sMRI features are used as keys and values ​​for querying fMRI to generate structural enhancement features Z that include temporal evolution. s←f ; S2.2, Nonlinear bottleneck dimensionality reduction: The channel splicing unit will generate Z... f←s With Z s←f The features are concatenated along the channel dimension to form a feature redundancy concatenated matrix. This matrix is ​​then fed into a single-layer cross-modal nonlinear fusion module with a fixed parameter structure. This module sequentially includes: a fully connected linear layer, a batch normalization layer to prevent internal covariate shift, and a leakage-corrected linear unit that introduces nonlinear manifold distortion, thus creating a nonlinear bottleneck layer. Through the physical compression of this bottleneck layer, deep algebraic coupling of heterogeneous features is achieved, outputting the final unified node feature matrix X. fusion ; S3. Static topological graph based on prior medical knowledge: The prior static graph construction module (3) directly reuses the initial input fMRI original time sequence signal matrix to calculate the Pearson correlation coefficient between the pairwise time sequences of 116 brain regions; in order to filter out high-frequency noise and generate a sparse skeleton with anatomical interpretability, it performs sparsification pruning using a preset fixed empirical threshold. After pruning, a global static prior functional connectivity adjacency matrix A is generated. This adjacency matrix A constitutes the static brain network topology skeleton, providing an interpretable topology with medical prior constraints for subsequent high-order graph convolution information aggregation. S4. Graph Convolutional Information Aggregation and Global Attention Pooling: The high-order graph information aggregation and pooling module (4) will refine the unified node feature matrix X. fusion As a node attribute, the static prior function connects the adjacency matrix A as the topology and is synchronously input into a two-layer graph convolutional network. The pathological features of each brain region are strictly aggregated into second-order neighborhoods along the topological skeleton of the static brain network. The brain region features updated by the graph convolutional network are connected to the global attention pooling layer for global pooling. The global attention pooling layer learns an attention weight vector and dynamically weights and sums the 116 brain regions, compressing the local node features into a global representation with extremely high information density, and outputting the final graph-level representation vector z representing the patient's overall brain pathological state. S5, Domain Adversarial Decoupling and MDD Classification: S5.1 Probability Prediction: Domain Adversarial and Disease Classification Module (5) inputs the ultimate graph-level representation vector z into the MDD main classifier, outputs the accurate prediction probability of whether or not the patient has major depressive disorder through the Softmax activation function, and calculates the cross-entropy disease classification loss L. cls L cls It reflects the gap between the predicted probability and the patient's actual disease status; S5.2 Identifying Data Sources (Hospitals and Sites): The final graph-level representation vector z is simultaneously passed through the gradient inversion layer and input to the site discriminator. The site discriminator outputs the probability distribution of each sample belonging to different hospitals and sites, thereby completing site classification and calculating the domain classification loss L. domain ; S5.3, Game Theory Mechanism: During forward propagation of the network, the gradient reversal layer behaves as an identity transformation; while during backpropagation to update the weights, the gradient reversal layer will change L... domain The returned gradient is forcibly negative and multiplied by the adversarial penalty coefficient -λ; to ensure the stability of the network in the early stages of training, the preferred update strategy for parameter λ is to adaptively increase it with each training round. , where p is the current training progress percentage.

Images