A brain network analysis method based on space-time dual-dimension multi-scale

Abstract

The application relates to the technical field of medical image diagnosis, and specifically discloses a brain network analysis method based on space-time dual-dimension multi-scale, which comprises the following steps: obtaining a functional magnetic resonance image and extracting time series of each brain region; constructing a functional brain map on fine, medium and coarse multi-scales based on Pearson correlation combined with functional homogeneity weighting; extracting multi-scale space features through a hybrid structure of a U-Net encoder and a multi-scale graph isomorphic network; for the BOLD signal of the brain region, adopting an internal-external dual-branch selective state space model architecture to capture short-term instantaneous fluctuations and long-term trend dependence respectively, obtaining time features through multi-scale context extraction and pathological time sequence attention mechanism to strengthen disease-related time sequence fragments; and inputting the space-time features after residual gate fusion into a classifier to obtain a classification result. The application breaks through the limitation of single-scale and single-time sequence modeling, realizes fine and functional fusion of brain space-time features, and improves the accuracy and interpretability of brain disease auxiliary diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and computer-aided diagnosis technology, specifically to a brain network analysis method based on spatiotemporal dual-dimensional multi-scale. Background Technology

[0002] Autism spectrum disorder is a common neurodevelopmental disorder, primarily characterized by social communication impairments and stereotyped behaviors. Currently, its clinical diagnosis mainly relies on behavioral observation, which is highly subjective and prone to misdiagnosis. Resting-state functional magnetic resonance imaging (fMRI), by measuring blood oxygenation level-dependent signals, can non-invasively reflect the connectivity strength between different functional areas of the brain, and has become an important tool for studying brain diseases and providing auxiliary diagnosis.

[0003] Traditional brain network analysis methods typically construct functional connectivity networks based on Pearson correlation coefficients. However, this approach only captures statistical synchronicity between brain regions, is susceptible to noise interference, and lacks consideration for the inherent functional organization of the brain, resulting in semantically ambiguous brain networks. In the feature extraction stage, existing methods suffer from two major limitations: in the spatial dimension, they often employ single-scale graph neural networks, making it difficult to simultaneously model local fine connections and global topological dependencies; in the temporal dimension, recurrent neural networks struggle to handle long-range dependencies, Transformer models have high computational complexity, and single state-space models mix short-term fluctuations and long-term trends in brain signals, failing to distinguish their different physiological meanings and leading to inaccurate capture of disease-specific temporal patterns.

[0004] Therefore, there is an urgent need for an analytical method that can construct more biologically meaningful brain networks and simultaneously refine the modeling of brain spatiotemporal features to improve the accuracy and interpretability of assisted diagnosis. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a brain network analysis method based on spatiotemporal dual-dimensional multi-scale. This method optimizes brain network construction by introducing functional homogeneity constraints and designs a collaborative multi-scale spatial feature extraction module and a dual-branch temporal feature extraction module to achieve deep fusion and efficient analysis of brain spatiotemporal features.

[0006] This invention is achieved through the following technical solution:

[0007] A method for analyzing brain networks based on spatiotemporal dual dimensions and multiple scales is provided, including the following steps:

[0008] S1. Obtain functional magnetic resonance imaging (fMRI) images of the subject's brain;

[0009] S2. Based on the time series of brain regions in functional magnetic resonance imaging, construct a brain functional network with functional homogeneity constraints at multiple scales (fine, medium, and coarse).

[0010] S3. Extract multi-scale spatial features from brain functional networks that incorporate functional homogeneity constraints using multi-scale graph isomorphic networks.

[0011] S4. The time series of each brain region is processed using a state-space model-based bi-branch temporal module to extract multi-scale temporal features. The multi-scale temporal features are then weighted using a pathological temporal attention mechanism to obtain weighted temporal features.

[0012] S5. By integrating multi-scale spatial features and weighted temporal features, joint spatiotemporal features are obtained;

[0013] S6. Input the joint spatiotemporal features into the classifier and output the brain functional state classification results.

[0014] Furthermore, in S2, a brain functional network with integrated functional homogeneity constraints is constructed, including:

[0015] Calculate the Pearson correlation coefficients between time series data of different brain regions to obtain the correlation matrix at each scale. ;

[0016] Based on a pre-defined brain functional network atlas, the weight vectors of each brain region belonging to different functional networks are obtained through the U-Net encoder.

[0017] Calculate the cosine similarity of the weight vectors between each brain region pair to obtain the functional homogeneity matrix at each scale. ;

[0018] For the correlation matrix With functional homogeneity matrix By performing weighted fusion, the adjacency matrix of the brain functional network with functional homogeneity constraints is obtained. :

[0019] ;

[0020] in: These are the weighting coefficients.

[0021] Furthermore, in S3, multi-scale spatial features are extracted using a multi-scale graph isomorphic network, including:

[0022] The adjacency matrix of the fine-scale functionalized brain network is input into the U-Net encoder to obtain feature maps corresponding to the three spatial scales of fine, medium and coarse.

[0023] The feature maps at each scale are input into independent graph isomorphic networks to extract the topological features at the corresponding scales;

[0024] The topological features output from each graph isomorphic network are fused, and the fused features are input into the U-Net decoder for upsampling to output multi-scale spatial features.

[0025] Furthermore, in S4, the dual-branch time series module based on the state space model includes a high-resolution branch and a low-resolution branch; and both the high-resolution branch and the low-resolution branch use a selective state space model for sequence modeling; the high-resolution branch is used to directly input the time series of each brain region into the first state space model to extract high-resolution time series features representing short-term fluctuations; the low-resolution branch is used to downsample the time series of each brain region and input it into the second state space model to extract low-resolution time series features representing long-term trends.

[0026] Furthermore, in S4, a pathological temporal attention mechanism is used to weight multi-scale temporal features, including:

[0027] High-resolution temporal features are fused with low-resolution temporal features to obtain preliminary fused temporal features;

[0028] By using multiple one-dimensional convolutional layers with different kernel sizes, multi-scale contextual features are extracted from the initially fused temporal features and then fused to obtain multi-scale temporal features.

[0029] The pathological temporal attention weighting is applied to the multi-scale temporal features. Specifically, the temporal importance score is obtained by passing a linear layer and a Sigmoid activation function, and the temporal importance score is multiplied element-wise with the multi-scale temporal features to obtain the weighted temporal features.

[0030] Furthermore, it also includes introducing functional homogeneity constraint loss during model training; functional homogeneity constraint loss is used to constrain the brain region feature vectors learned by the multi-scale graph isomorphic network to maximize the cosine similarity between them and the prototype feature vectors of the functional network to which they belong.

[0031] Furthermore, in S5, the step of fusing multi-scale spatial features and weighted temporal features adopts a residual gating fusion mechanism.

[0032] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method.

[0033] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0034] The beneficial effects of this invention are:

[0035] I. Brain Network Construction More Consistent with Biological Priors: By fusing Pearson correlation coefficients with functional homogeneity-based weighted fusion from functional network maps, a "functionalized" brain network with less noise and clearer functional semantics is constructed, providing a higher quality foundation for subsequent analysis. This invention introduces "functional homogeneity" as a construction constraint. The principle is that, based on the brain's inherent functional network maps (such as the default mode network, visual network, etc.), the U-Net encoder learns the "attribution weights" of each brain region to different functional networks and calculates the similarity (cosine similarity) of the attribution patterns between brain region pairs. Finally, the statistical correlation matrix and the functional homogeneity matrix are weighted and fused, essentially using prior knowledge of functional organization to filter and correct data-driven statistical connections. This suppresses statistically random but functionally unreasonable spurious connections while strengthening connections between brain regions belonging to the same functional system. The constructed brain network is "functionalized," with higher basic quality, less noise, and clearer functional semantics.

[0036] II. More Comprehensive Spatial Feature Extraction: This invention designs a hybrid architecture of "U-Net encoder - multi-scale GIN decoder." Its core principle is to leverage the inherent multi-resolution representation capabilities of U-Net and the powerful topological relationship modeling capabilities of GIN. The U-Net encoder downsamples fine-scale functionalized brain maps, automatically generating multi-scale feature maps aligned with the receptive fields of medium- and coarse-scale brain region divisions. Each scale's feature map is fed into an independent GIN, which operates only on the functionalized brain network topology at its corresponding scale, learning the interaction patterns between nodes (brain regions or networks) at that scale. The outputs of the multi-scale GINs are fused and then upsampled by the U-Net decoder to restore details, ensuring a comprehensive capture of brain spatial patterns. Fine-scale GINs focus on the fine functional collaboration of local brain region clusters, while coarse-scale GINs characterize the integration and separation relationships between large-scale functional networks. This overcomes the shortcomings of single-scale models.

[0037] III. More Refined Temporal Feature Modeling: The innovative use of an internal and external bi-branch selective state-space model (SSM) explicitly separates and models the short-term transient fluctuations and long-term trend dependence of the BOLD signal, better reflecting its physiological characteristics. Furthermore, it enhances disease-related segments through a pathological temporal attention mechanism. This avoids mutual interference between information at different time scales, resulting in higher purity and clearer physical meaning of the extracted temporal features. The model can more accurately capture transient abnormalities or long-term pattern changes that are potentially more relevant to the disease. Employing the latest selective state-space model (Mamba), its linear computational complexity overcomes the quadratic complexity bottleneck of the Transformer when processing long fMRI sequences, achieving a balance between efficiency and performance.

[0038] IV. Enhanced Diagnostic Specificity, Feature Interpretability, and Generalization: The pathological temporal attention mechanism enables the model to dynamically focus, significantly enhancing the disease-discriminating power of features; functional homogeneity loss, as a form of structured regularization, guides the model to learn features that conform to known brain functional organization patterns. This not only makes the model's decisions more understandable (features are associated with known functional networks) but also introduces powerful biological priors, helping to improve the model's generalization ability on unseen data and preventing overfitting to data noise.

[0039] In this invention, the functionalized brain map provides a higher quality input benchmark for all subsequent spatial and temporal analyses. Multi-scale spatial features and separately modeled temporal features characterize brain abnormalities from different dimensions and are deeply complementary through residual-gated fusion. Pathological temporal attention and functional homogeneity loss continuously guide and correct the direction of feature learning throughout the training process, optimizing it towards high diagnostic performance and high interpretability. This invention proposes a novel and synergistically enhanced technical solution that can construct a clearer brain connectivity map from noisy fMRI data, extracting more comprehensive, purer, and more interpretable spatiotemporal features. This provides more accurate, reliable, and understandable quantitative evidence for the auxiliary diagnosis of brain diseases, possessing significant theoretical breakthrough value and clinical translational potential. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the implementation of the spatiotemporal dual-dimensional multi-scale brain network analysis method of this invention.

[0041] Figure 2 This is a model framework diagram of the brain network analysis method based on spatiotemporal dual-dimensional multi-scale of the present invention. Detailed Implementation

[0042] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0043] This embodiment uses the publicly available Autism Brain Imaging Data Exchange (ABIDE) dataset as an example. This dataset contains resting-state functional magnetic resonance imaging (rs-fMRI) data of 184 subjects (79 autistic patients and 105 normal controls). The method provided by this invention is applied to assist in the identification of autism spectrum disorder. Figure 1 As shown, the specific steps are as follows:

[0044] S1. Obtain and preprocess the functional magnetic resonance imaging (fMRI) images of the subjects' brains;

[0045] The following preprocessing operations were performed on the acquired raw brain functional image data using the DPARSF toolbox in MATLAB:

[0046] Remove the first few time points: To eliminate the influence of the initial unstable signal of the magnetic field, the first 10 time points of each image being attempted are removed.

[0047] Head movement and temporal correction: Head movement artifact correction and temporal correction are performed on the images to reduce the impact of subject head movement and scanning sequence.

[0048] Covariate regression: Regression removes the effects of ventricular and white matter signals and higher-order head movement effects (Friston 24 parameters) to eliminate non-neuronal physiological noise.

[0049] Spatial normalization: The corrected image is registered to the standard brain template space of the Montreal Neuroscience Institute to achieve spatial alignment.

[0050] Filtering: The image time series to be viewed is subjected to a bandpass filter of 0.01~0.1 Hz to retain low-frequency oscillation signals and reduce the influence of high-frequency physiological noise (such as heartbeat and respiration) and low-frequency drift.

[0051] S2. Based on the time series of brain regions in functional magnetic resonance imaging, construct a brain functional network with functional homogeneity constraints at multiple scales (fine, medium, and coarse).

[0052] The preprocessed brain images were divided into 116 brain regions, including 90 cerebral regions and 26 cerebellar regions, using an automated anatomical labeling (AAL) atlas. The average BOLD signal of all voxels in each brain region at each time point was extracted to form a time series for each brain region.

[0053] The time series data of all brain regions are combined and represented as a time series matrix. ,in Indicates the first Time series of each node Indicates the length of the time series. Indicates the number of brain regions (in this embodiment) =116).

[0054] We constructed a brain functional network that incorporates functional homogeneity constraints at three spatial scales: fine, medium, and coarse.

[0055] Define three spatial scales:

[0056] Fine-scale brain mapping directly uses the original brain region division results; meso-scale brain mapping merges adjacent functional brain regions into complex brain regions; coarse-scale brain mapping aggregates multiple complex brain regions into large-scale functional networks according to the brain's functional network system, resulting in brain region sets of fine, meso, and coarse scales, denoted as... , , ,in , and These represent the number of brain regions at the fine, intermediate, and coarse scales, respectively. In this embodiment, the fine scale directly uses the 116 brain regions defined by the AAL template. =116, and based on anatomical and functional proximity, the 116 brain regions were merged into 58 complex brain regions at the mesoscale. =58, coarse-scale based on the Yeo-7 network functional map, the brain regions are aggregated into 7 large-scale functional networks, resulting in =7.

[0057] Calculate the Pearson correlation matrix:

[0058] For brain regions and Its Pearson correlation coefficient:

[0059] ;

[0060] in It is for brain regions The time series data was obtained by z-score standardization, resulting in a mean of 0 and a standard deviation of 1.

[0061] ;

[0062] in brain region At the point of time The signal value; and brain regions The mean and standard deviation. Traversing all brain regions... , total calculation Secondary correlation coefficient, output symmetric correlation matrix .

[0063] Calculate the functional homogeneity matrix:

[0064] First, spatial features of brain regions are extracted using the U-Net encoder to obtain the weight matrix of each brain region belonging to different functional networks. ,in K The number of functional networks (in this example, ) K= 7) For each brain region i, concatenate its attribution weights to form an attribution feature vector. Calculate the functional homogeneity of any two brain regions i and j. The cosine similarity of its function-attributed feature vectors:

[0065] ;

[0066] Obtain the functional homogeneity matrix at each scale. .

[0067] By weighting and fusing the Pearson correlation matrix and the functional homogeneity matrix proportionally, the adjacency matrices of functional brain networks at various scales are obtained:

[0068] ;

[0069] in The weighting coefficients are α, which is an adjustable hyperparameter. In the experiment, the value was set to 0.5, thus obtaining the adjacency matrices of the functional brain network at fine, medium, and coarse scales. , and .

[0070] S3. Extract multi-scale spatial features from brain functional networks that incorporate functional homogeneity constraints using multi-scale graph isomorphic networks.

[0071] For the U-Net encoder, we will obtain the fine-scale functionalized brain network adjacency matrix. As input, the encoder extracts local functional details step by step through multi-layer convolution and non-linear activation operations, and outputs multi-scale feature maps that can be mapped to predefined fine, medium, and coarse brain regions respectively. , and Semantic alignment prepares the ground for subsequent multi-scale graph neural network processing.

[0072] The scale-aligned feature maps output from the U-Net encoder are input into three independent graph isomorphic networks. Each GIN is a functionalized brain network at its corresponding scale. Message passing and node feature updates are performed on the defined topology.

[0073] The multi-scale graph isomorphic network defined under spatial features is shown below:

[0074] ; ;

[0075] in It is a functional brain network constructed using Pearson correlation and functional homogeneity weighting. For the first Layer nodes The feature vectors are initialized. It is a multilayer perceptron; Here are learnable parameters, or you can simply set a fixed scalar; It is with nodes An adjacent pair of nodes, Meaning: aggregation node The neighbor node information. Fine, medium, and coarse-scale GINs are updated independently and iteratively, outputting the global spatial feature matrix at each scale:

[0076]

[0077] in K The total number of layers in the GIN is given. After stitching, the spatial features at fine, medium, and coarse scales are obtained as follows: , and .

[0078] The spatial features output by fine, medium, and coarse-scale GINs are aligned by dimension and then weighted and fused to obtain fused multi-scale spatial features. .Will As input to the U-Net decoder, fine-scale spatial details are gradually recovered through deconvolution upsampling, ultimately outputting complete spatial features. , D For feature dimensions.

[0079] S4. The time series of each brain region is processed using a state-space model-based bi-branch temporal module to extract multi-scale temporal features. The multi-scale temporal features are then weighted using a pathological temporal attention mechanism to obtain weighted temporal features.

[0080] Disease-discriminating temporal features are extracted from the original BOLD time series, and the acquired BOLD signals are input into... It is divided into two paths: the outer Mamba (high-resolution branch) directly inputs the original resolution signal to capture short-term transient fluctuations; the inner Mamba (low-resolution branch)... Perform downsampling (average pooling) ), to obtain low-resolution BOLD signal To capture long-term trend dependence.

[0081] Obtained through selective mapping , and Based on the state-space model (SSM):

[0082] ;

[0083] ;

[0084] Because solving the above differential equations in a deep learning environment is very difficult, we need to use parameters. Using zero-order hold technique and Discretize, and then obtain and input Multiply, then discretize to get and the original state Multiplying these two terms and adding them together gives a new state. Finally, the new state and Multiplication yields the output :

[0085] ;

[0086] ;

[0087] in , , It is the identity matrix. The discretized SSM is equivalent to the following convolution:

[0088] ;

[0089] ;

[0090] After concatenating and aligning the outputs of the inner and outer Mamba, they are passed in parallel through 1 3.1 5.1 One-dimensional convolution with 7 kernels combined with SSM extracts short, medium, and long-term temporal context features, which are then weighted and fused to obtain multi-scale temporal features. .

[0091] Next, we need to... Pathological time-series attention weighting is performed by mapping the feature dimension from D to 1D through a linear layer, and then obtaining the time-series importance score after Sigmoid activation. Then, it is multiplied element-wise with the original feature to enhance the disease-related time series segments, and finally outputs the time feature. .

[0092] S5. By integrating multi-scale spatial features and weighted temporal features, joint spatiotemporal features are obtained;

[0093] spatial features With time characteristics Fusion is performed. This embodiment employs a residual gating fusion mechanism: firstly, ... and The features are mapped to the same feature space and added together. Then, an adaptive fusion weight is generated through a gating unit consisting of a fully connected layer and a sigmoid function. This weight modulates the fused features. Finally, the fused features are added together with the initial fused features to form the final joint spatiotemporal feature representation used for classification.

[0094] S6. Input the joint spatiotemporal features into the classifier and output the brain functional state classification results;

[0095] The joint spatiotemporal features are input into a fully connected layer classifier, and finally the predicted probability of the subject belonging to "autism patient" or "normal control" is output through the Softmax function. The category with the higher probability value is taken as the model's diagnostic result.

[0096] The entire model described above is trained end-to-end under supervised conditions using a labeled training set.

[0097] The loss function is constructed using cross-entropy loss plus functional homogeneity loss, as shown in the following formula:

[0098] The supervised loss (cross-entropy loss) is calculated based on the model's predictions and the true labels.

[0099] ;

[0100] Where M is the number of disease diagnosis categories, The true label of the sample (one-hot encoding). This represents the predicted probability of the model (softmax output).

[0101] Functional homogeneity loss constrains the functional homogeneity consistency of multi-scale GIN output features, and its formula is as follows:

[0102] ;

[0103] in for scale The feature vector of the i-th brain region at the scale, i.e. ; The functional network to which this brain region belongs The central eigenvector.

[0104] The formula for the total loss function is:

[0105] ;

[0106] in This is the loss balancing coefficient, used to control the weight of losses due to functional homogeneity.

[0107] The model framework diagram of this invention based on a spatiotemporal dual-dimensional multi-scale brain network analysis method is shown below. Figure 2 As shown, the training parameters are set as follows:

[0108] Optimizer: Use the Adam optimizer.

[0109] Learning rate: The initial learning rate is set to 1×10. -4 Because the BOLD signal value is relatively small (approximately 10).-2 (Scale), a smaller learning rate helps stabilize training.

[0110] Batch size: Set to 16. Given that the sample size of medical image data is usually limited, an appropriate batch size combined with a small learning rate can ensure the stability and generalization ability of model training.

[0111] Early stopping mechanism: Set the patience value to 10. When the validation set loss no longer decreases for 10 consecutive training epochs, terminate training early and save the model parameters with the best performance on the validation set to prevent overfitting.

[0112] To verify the effectiveness of the method of the present invention and quantitatively evaluate its performance, a comparative experiment was conducted on the publicly available Autism Brain Imaging Data Exchange (ABIDE I) dataset. The experiment selected samples from the NYU site within this dataset and extracted time series data from 116 brain regions based on the AAL atlas. This subset of the dataset is often referred to as NYU116 in related studies and includes a certain number of autistic patients and normal control subjects.

[0113] Experimental Setup: To ensure the fairness and reproducibility of the evaluation, the data from the NYU116 site were randomly divided into training, validation, and test sets in a 7:1:2 ratio. All comparison methods used the same training set for model training, the validation set for hyperparameter tuning and early stopping, and finally, performance was reported on a separate test set. Evaluation metrics included classification accuracy, AUC, sensitivity, and specificity.

[0114] Comparison Methods: Several representative advanced methods in the field of brain network analysis in recent years were selected as baselines for comparison, including:

[0115] BrainGNN: A brain network analysis method based on graph neural networks.

[0116] BrainNetCNN: A convolutional neural network specifically designed for brain connectome data.

[0117] BrainGB: A brain network analysis model that employs graph attention mechanisms and self-supervised learning.

[0118] BrainNetTF: A brain network classification model based on the Transformer architecture.

[0119] BrainMamba: A method for processing temporal brain signals using a selective state-space model.

[0120] BrainNetMLP: A method for classifying brain network features based on multilayer perceptron.

[0121] Table 1. Comparative experimental results of NYU116 sites on the ABIDE dataset.

[0122]

[0123] Results Analysis: As shown in Table 1, the method of this invention (Ours) achieved a classification accuracy of 71.85% on the NYU116 test set, significantly outperforming all the listed comparison methods. Specifically:

[0124] Accuracy: The method of this invention outperforms the second-best performing BrainNetMLP (68.21%) by about 3.6 percentage points, demonstrating the effectiveness of this invention in integrating spatiotemporal multi-scale features.

[0125] AUC value: The method of this invention achieved 73.20, which is close to the highest BrainNetMLP (73.70), indicating that it has good overall ranking ability.

[0126] Sensitivity and Specificity: The method of this invention achieves a more balanced performance in terms of sensitivity (69.40%) and specificity (74.90%). Compared to some baseline methods (such as BrainMamba, which has high specificity but extremely low sensitivity), the method of this invention is more robust in identifying patients and excluding healthy controls, which is crucial in actual clinical auxiliary diagnosis.

[0127] In summary, the comparative experimental results show that the brain network construction method with functional homogeneity constraints, spatiotemporal dual-dimensional multi-scale feature extraction and fusion method proposed in this invention can more effectively capture abnormal brain network patterns related to autism spectrum disorder, thus surpassing many existing mainstream technologies in key performance indicators and verifying the beneficial effects of this invention.

[0128] To verify the necessity and effectiveness of each core module in this invention, ablation experiments were designed and conducted on the same NYU116 site. Based on the complete model (denoted as Ours-Full), the following ablation variant models were constructed sequentially for comparison:

[0129] 1. Ours w / o FH: Remove functional homogeneity constraints, i.e., when constructing the brain network, set the weighting coefficient α=1 and use only the Pearson correlation matrix.

[0130] 2. Ours w / o MS-GIN: Remove the multi-scale graph isomorphic network and directly input the features output by the U-Net encoder into the decoder, that is, use a single-scale (fine-scale) GIN.

[0131] 3. Ours w / o Dual-Mamba: Replaces the dual-branch selective state-space model of the temporal feature extraction module with a single-branch Mamba model.

[0132] 4. Ours w / o PA: Remove pathological temporal attention mechanism, that is, do not weight multi-scale temporal features.

[0133] All ablation models used the exact same training / validation / test set splits, hyperparameter settings, and random seeds as the full models. Experimental results are shown in Table 2 below:

[0134]

[0135] Results analysis:

[0136] Table 2 shows that each module affected the model performance to varying degrees. The complete model achieved the best results across all metrics, indicating a good synergistic effect among its components. Removing the dual-branch temporal modeling module (w / o Dual-Mamba) resulted in the most significant performance decline, particularly in sensitivity, which dropped from 69.40% to 62.30%, demonstrating the module's crucial role in capturing disease-related temporal dynamics. Removing multi-scale spatial modeling (w / o MS-GIN) also led to a significant decrease in Acc and AUC, highlighting the importance of multi-scale topological information for brain network representation. Removing the pathological temporal attention mechanism (w / o PA) reduced model sensitivity, indicating that this mechanism effectively reinforces key temporal segments. In contrast, removing the functional homogeneity constraint (w / o FH) had a smaller impact on overall performance, but resulted in a significant decrease in specificity, suggesting its primary function was to improve the model's robustness and noise resistance.

[0137] The ablation experiment results above demonstrate that each technical feature of this invention contributes positively to the final performance, and their synergistic effect enables the complete model to achieve optimal comprehensive diagnostic performance. This empirically proves the completeness and innovation of the technical solution of this invention.

[0138] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A brain network analysis method based on spatiotemporal dual-dimensional multi-scale, characterized in that: Includes the following steps: S1. Obtain functional magnetic resonance imaging (fMRI) images of the subject's brain; S2. Based on the time series of brain regions in functional magnetic resonance imaging, construct a brain functional network with functional homogeneity constraints at multiple scales (fine, medium, and coarse). Constructing brain functional networks with integrated functional homogeneity constraints, including: Calculate the Pearson correlation coefficients between time series data of different brain regions to obtain the correlation matrix at each scale. ; Based on a pre-defined brain functional network atlas, the weight vectors of each brain region belonging to different functional networks are obtained through the U-Net encoder. Calculate the cosine similarity of the weight vectors between each brain region pair to obtain the functional homogeneity matrix at each scale. ; For the correlation matrix With functional homogeneity matrix By performing weighted fusion, the adjacency matrix of the brain functional network with functional homogeneity constraints is obtained. : ； in: These are weighting coefficients; S3. Extract multi-scale spatial features from brain functional networks that incorporate functional homogeneity constraints using multi-scale graph isomorphic networks; the multi-scale graph isomorphic network defined under spatial features is shown below: ； ； in It is a functional brain network constructed using Pearson correlation and functional homogeneity weighting. For the first Layer nodes The feature vector, initialized It is a multilayer perceptron; Here are learnable parameters, or you can simply set a fixed scalar; It is with nodes An adjacent pair of nodes, Meaning: aggregation node The neighbor node information; fine-, medium-, and coarse-scale GINs are updated independently and iteratively, outputting the global spatial feature matrix at each scale: ； in K The total number of layers in the GIN is given. After stitching, the spatial features at fine, medium, and coarse scales are obtained as follows: , and ; The spatial features output by fine, medium, and coarse-scale GINs are aligned by dimension and then weighted and fused to obtain fused multi-scale spatial features. ;Will As input to the U-Net decoder, fine-scale spatial details are gradually recovered through deconvolution upsampling, ultimately outputting complete spatial features. , D For feature dimensions; S4. The time series of each brain region is processed using a state-space model-based bi-branch temporal module to extract multi-scale temporal features. The multi-scale temporal features are then weighted using a pathological temporal attention mechanism to obtain weighted temporal features. S5. By integrating multi-scale spatial features and weighted temporal features, joint spatiotemporal features are obtained; S6. Input the joint spatiotemporal features into the classifier and output the brain functional state classification results.

2. The brain network analysis method based on spatiotemporal dual-dimensional multi-scale as described in claim 1, characterized in that: In S3, multi-scale spatial features are extracted using a multi-scale graph isomorphic network, including: The adjacency matrix of the fine-scale functionalized brain network is input into the U-Net encoder to obtain feature maps corresponding to the three spatial scales of fine, medium and coarse. The feature maps at each scale are input into independent graph isomorphic networks to extract the topological features at the corresponding scales; The topological features output from each graph isomorphic network are fused, and the fused features are input into the U-Net decoder for upsampling to output multi-scale spatial features.

3. The brain network analysis method based on spatiotemporal dual-dimensional multi-scale according to claim 1, characterized in that: In S4, the dual-branch time series module based on the state space model includes a high-resolution branch and a low-resolution branch; and both the high-resolution branch and the low-resolution branch use a selective state space model for sequence modeling; the high-resolution branch is used to directly input the time series of each brain region into the first state space model to extract high-resolution time series features that characterize short-term fluctuations. The low-resolution branch is used to downsample the time series of each brain region and input it into the second state space model to extract low-resolution temporal features that represent long-term trends.

4. The brain network analysis method based on spatiotemporal dual-dimensional multi-scale as described in claim 3, characterized in that: In S4, a pathological temporal attention mechanism is used to weight multi-scale temporal features, including: High-resolution temporal features are fused with low-resolution temporal features to obtain preliminary fused temporal features; By using multiple one-dimensional convolutional layers with different kernel sizes, multi-scale contextual features are extracted from the initially fused temporal features and then fused to obtain multi-scale temporal features. The pathological temporal attention weighting is applied to the multi-scale temporal features. Specifically, the temporal importance score is obtained by passing a linear layer and a Sigmoid activation function, and the temporal importance score is multiplied element-wise with the multi-scale temporal features to obtain the weighted temporal features.

5. The brain network analysis method based on spatiotemporal dual-dimensional multi-scale as described in claim 1, characterized in that: It also includes introducing functional homogeneity constraint loss during model training; functional homogeneity constraint loss is used to constrain the brain region feature vectors learned by the multi-scale graph isomorphic network to maximize the cosine similarity between them and the prototype feature vectors of the functional network to which they belong.

6. The brain network analysis method based on spatiotemporal dual-dimensional multi-scale as described in claim 1, characterized in that: In S5, the step of fusing multi-scale spatial features and weighted temporal features adopts a residual gating fusion mechanism.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images