Brain signal analysis method and system based on spatial-temporal feature enhancement
By combining CRF and MRF models with manifold learning attention mechanisms for comparative learning, the feature representation of fMRI data is enhanced, solving the problem of insufficient utilization of temporal characteristics in existing technologies and improving the diagnostic accuracy and robustness of neurological diseases.
Patent Information
- Application Number
- CN202511364827.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-23
- Publication Date
- 2025-12-19
AI Technical Summary
Existing fMRI data analysis methods struggle to effectively utilize temporal characteristics, lack multi-scale temporal information integration, and their models lack robustness and versatility in the diagnosis of neurological diseases.
We employ CRF and MRF models combined with manifold learning attention mechanism to enhance feature representation through contrastive learning, construct a brain network map, and integrate temporal and spatial features.
It significantly improves the feature representation ability of fMRI data, enhancing the accuracy and robustness of auxiliary diagnosis of neurological diseases.
Smart Images
Figure CN121167445A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electroencephalogram (EEG) data processing technology, and in particular to a brain signal analysis method and system based on spatiotemporal feature enhancement. Background Technology
[0002] Due to its non-invasive nature and ability to capture real-time brain activity, fMRI (functional magnetic resonance imaging) has become a valuable tool for the auxiliary diagnosis of neurological diseases. Although significant progress has been made in fMRI-based analysis, effectively utilizing the temporal properties of fMRI data in a structured and discriminative manner remains a challenging task. Several key challenges continue to hinder further progress: (1) models have limited ability to jointly capture spatial and temporal dependencies within fMRI time series, with most models only modeling single features; (2) it is difficult to learn discriminative representations that can sensitively reflect subtle pathological changes among subjects; and (3) there is a lack of structure-aware fusion mechanisms to integrate multi-scale temporal information into robust and interpretable representations of brain activity. Most existing methods treat CRF (conditional random field) and MRF (Markov random field) as separate modules, rarely considering their joint integration within a unified deep learning architecture, especially in the context of spatiotemporal modeling of brain functional data.
[0003] In recent years, numerous diagnostic methods based on fMRI data have been developed, including graph neural network-based models such as BrainGNN and BrainGB, and attention-based models such as Brain network Transformer and BC-MGAT. While these methods effectively capture the static topological features of brain networks, they may not fully utilize the inherent high dimensionality and temporal characteristics of fMRI data. This simplification limits the model's ability to capture dynamic interactions between regions and may hinder the accurate detection of disease-specific abnormal patterns. Furthermore, challenges such as noise, individual variability, and uncertain mappings between pathological features and functional activation reduce the robustness and versatility of these models in clinical practice. Therefore, there is an urgent need for a time-series augmentation strategy that can effectively simulate temporal dynamics and enhance resilience to better support the auxiliary diagnosis of neurological diseases.
[0004] Contrastive learning excels at learning semantic relationships between different augmented views, offering a compelling approach to feature fusion. By distinguishing the similarities and differences between samples under various augmentations, it improves feature discriminability and enables models to learn more robust and generalizable representations. Existing research has explored this idea in the context of graph augmentation of fMRI-derived brain networks—for example, Hu et al. used two-stage spectral augmentation to generate different functional connectivity maps constructed from fMRI data for contrastive learning; while Wang et al. applied contrastive learning between the original map and the diffuse augmented map constructed from fMRI time series. These methods primarily focus on augmenting the encoder by introducing structural changes into the input map. Complementing this direction, some recent work has begun to explore improvements in the decoding stage, incorporating attention mechanisms or feature weighting strategies to optimize task-specific outputs. However, these methods typically focus on node-level or edge-level features without explicitly modeling the multi-scale hierarchical organization or temporal evolution patterns of brain activity. While they can capture local interactions, they may struggle to characterize the global, dynamic, and non-Euclidean properties of the nervous system, potentially limiting their ability in complex decoding tasks. In contrast, manifold learning offers tremendous potential for modeling the intrinsic geometry of neural signals, and its integration with attention mechanisms has shown promising results in decoding tasks such as EEG analysis. These findings suggest that a structure-aware, multi-directionally guided fusion mechanism is crucial for capturing meaningful patterns of brain activity and improving the robustness and interpretability of diagnostic models. Summary of the Invention
[0005] The purpose of this invention is to provide a brain signal analysis method and system based on spatiotemporal feature enhancement, which significantly enhances feature representation by analyzing the time series of brain signals from both temporal and spatial perspectives.
[0006] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, the present invention provides a brain signal analysis method based on spatiotemporal feature enhancement, comprising the following steps: Acquire fMRI data and preprocess it; Based on the preprocessed data, a CRF model is used to learn the dynamic transformation relationship between activation labels, and weighted and functional connectivity features are extracted from the time series data to obtain a weighted enhanced time series. ; The activation probability of each ROI (Region of Interest) is used as input to the MRF model. By iteratively minimizing the energy function, the activation state of each ROI is optimized in each time frame. The activation confidence of the ROI at time step is combined with the corresponding original time series to obtain a weighted enhanced time series. ; A weighted and enhanced time series algorithm is constructed based on manifold learning attention mechanism. and Functional connections result in enhanced features. It is used to construct brain network maps.
[0007] In some embodiments, the dynamic transformation relationship between activation labels is learned through a CRF model, and time-series weighted and functional connectivity features are extracted, including the following steps: Activation labels are generated based on statistical thresholds; Based on the activation state at each time point, the weights of the state features learned by the CRF model are... Feature vectors used in construction By weighting each component of the feature vector accordingly, a score reflecting the relative importance of neural activity at that moment is obtained. ; Using weighted scores Constructing weighted augmented time series .
[0008] In some embodiments, the feature vector is extracted from the average BOLD (Blood Oxygen Level Dependent) time series of each ROI. The following is constructed from: ; In the formula, This represents the current value of the time series. The difference from the previous time point, for Time series values at time points, for Time series values at any given time; For adjacent time series values, This represents the window width and is set to a constant. This is the deviation term.
[0009] In some embodiments, the weighted enhanced time series The calculation expression is: ; in, ; ; In the formula, A one-dimensional standard Gaussian kernel, This indicates the time offset from the current time point. , This represents the window width and is set to a constant. Standard deviation; for The score of the relative importance of neural activity at any given moment; for Activation tag at any moment; The weights obtained from training the CRF model. For the first One ROI, For feature vectors; For the first ROI Weights obtained during CRF model training at time step; For feature vectors; For the first ROI at time point The BOLD signal.
[0010] In some embodiments, the energy function includes single-point energy and paired energy, wherein the single-point energy is used to reflect the cost associated with whether each ROI is activated, and the paired energy is used to ensure the spatial consistency of the activation state of each ROI with its neighboring regions.
[0011] In some embodiments, the expression for the energy function is: ; in, for: ; ; for: ; ; In the formula, For the first ROI The energy function at a single point in time, where 1 indicates an active state; For the first ROI The energy function at a single point in time, where 0 represents the inactive state; For the first ROI in The activation probability at any given time; The energy of the activated region increases with the number of adjacent regions; To penalize inactive ROIs adjacent to the active region; For hyperparameters; For the first The ROI and the An adjacency matrix is constructed from ROIs, defining regions with both positive and both negative Pearson correlation coefficients as neighbors. For the first The number of neighboring nodes activated by each ROI.
[0012] In some embodiments, the weighted enhanced time series The expression is: ; in, ; In the formula, In order to be in Time of the first Confidence level of each ROI activation; For the first ROI at time point The BOLD signal; For the first ROI The energy function at a single point in time, where 0 represents the inactive state; For the first ROI The energy function at a single point in time, where 1 indicates an active state.
[0013] In some embodiments, the enhanced features are obtained through the following methods : Using Pearson correlation coefficient from weighted enhanced time series and Transformation to obtain projection features and ; Applying linear transformations to enhance the projection features of time series and To obtain the Query and Key representations; The query and key are mapped from Euclidean space to a Poincaré spherical manifold through an exponential mapping to simulate complex inter-region interactions in a non-Euclidean latent space. Will and The input is fed into a shared dense layer to perform an initial transformation on the features, and then a feature fusion model with attention weights is used. Combining them yields enhanced features. .
[0014] In some embodiments, the feature fusion model is: ; in, ; In the formula, For enhanced features; For hyperparameters, ; These are the initial features for fusion; The functional connectivity matrix obtained by weighting the CRF; The functional connectivity matrix obtained by weighting the MRF; The initial feature matrix for fusion The value; For shared dense layers; The functional connectivity matrix obtained by weighting the CRF ; Functional connectivity matrix obtained by weighting MRF .
[0015] On the other hand, the present invention provides a brain signal analysis system based on spatiotemporal feature enhancement for performing the above-described method, comprising: Data acquisition module: used to acquire fMRI data and perform preprocessing; The first enhancement module: Based on the preprocessed data, it learns the dynamic transformation relationship between activation labels through a CRF model, and extracts weighted functional connectivity features from the time series to obtain a weighted enhanced time series. ; The second enhancement module uses the activation probability of each ROI as input to the MRF model and iteratively minimizes the energy function to optimize the activation state of each ROI in each time frame. The activation confidence of the ROI at time step is combined with the corresponding original time series to obtain a weighted enhanced time series. ; Fusion Module: Based on the manifold learning attention mechanism, a weighted augmented time series is obtained by fusing the first augmentation module and the second augmentation module. and The functional connections ultimately result in enhanced features. It is used to construct brain network maps.
[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention enhances fMRI time series data by extracting the training weights of CRF and MRF models, thereby leveraging the complementary advantages of CRF models in temporal modeling and MRF models in spatial processing.
[0017] This invention uniquely integrates weighted features from CRF and MRF models after contrastive learning using an improved manifold structure attention framework, significantly enhancing feature representation. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall structure of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of feature fusion in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of Embodiment 2 of the present invention. Detailed Implementation
[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0020] Example 1: Please see Figures 1-2 A brain signal analysis method based on spatiotemporal feature enhancement includes the following steps: S1. Acquire fMRI (functional magnetic resonance imaging) data and preprocess it using the CONN toolbox.
[0021] The preprocessing includes temporal and spatial preprocessing, such as slice timing correction, spatial realignment, normalization, and noise reduction.
[0022] ROIs (Regions of Interest) are defined based on AAL (Automated Anatomical Labeling) atlases, and the average BOLD (Blood Oxygen Level Dependent Effect) time series extracted from each ROI is denoted as... To ensure temporal alignment and comparability among subjects, all time series were normalized to a fixed length T, which facilitates consistent modeling of subjects, enabling reliable analysis of sequence patterns and functional connectivity without temporal mismatch due to variations in scan duration.
[0023] S2. Based on the preprocessed data, the dynamic transformation relationship between activation labels is learned through the CRF model, and the time series is weighted and functional connectivity features are extracted to obtain a weighted enhanced time series. .
[0024] The CRF model not only captures the temporal dependencies in the activated label sequence, but also highlights the key stages of neural activity, thereby enhancing the expressive power of the data at the temporal level.
[0025] Specifically, firstly, activation labels are generated based on statistical thresholds. Then, a CRF model is used to learn the dynamic transformation relationship between activation labels, thereby achieving more refined time-series weighted and functional connectivity feature extraction.
[0026] In one specific embodiment, activation labels are calculated using statistical analysis based on the characteristics of the time series. The definition is as follows: ; In the formula, a value of 1 indicates an active state, while 0 indicates an inactive state.
[0027] in, ; ; In the formula, For the first ROI at time point The BOLD signal; For the first Time series of average BOLD extracted from each ROI The average value; For the first Time series of average BOLD extracted from each ROI The standard deviation; This represents the length of the time series.
[0028] For each time point, extract the average BOLD time series from each ROI. Construct an eigenvector from To enhance the model's sensitivity to local temporal dynamics in time series: ; In the formula, This represents the current value of the time series. The difference from the previous time point, for Time series values at time points, for Time series values at any given time; For adjacent time series values, This represents the window width and is set to a constant with a value of 2. This is the deviation term.
[0029] Linear-chain CRF models are used to model the temporal structure of time series, thereby generating conditional probability distributions on the feature label sequences. By capturing feature label associations and temporal dependencies, CRF models can make structured predictions of the entire sequence. The training objective is to maximize the true label sequence (…). The conditional log-likelihood of ).
[0030] The conditional probability distribution is: ; ; In the formula, The weights, which are related to the features, reflect the contribution of each time point to the label assignment. It represents the indirect weights between labels at adjacent time points, emphasizing the temporal continuity and dependencies within the label sequence; This is a normalization factor used to ensure that the sum of the conditional probabilities of all possible label sequences is 1. , All are weighted balancing terms, with a value of 0.1; Adjacent , Indirect weights between time tags.
[0031] Based on the activation state at each time point, the weights of the state features learned by the CRF model are... Feature vectors used in construction Each component of the feature vector is weighted accordingly to generate a score that reflects the relative importance of neural activity at that moment. : ; In the formula, for Activation tag at any moment; Weights obtained from training the CRF model For the first One ROI, For feature vectors; For the first ROI Weights obtained during CRF model training at time step; These are the eigenvectors.
[0032] Considering the temporal propagation and influence of neural activity, activation at a single time point can spread to adjacent time points. Therefore, a one-dimensional Gaussian filter is used to further smooth the weighted sequence to simulate the temporal spread of activation states. This aims to improve the temporal continuity of activation patterns and reduce the influence of local signal fluctuations. These weighted scores are then used to reconstruct the time series, resulting in a CRF-enhanced time series. : ; in, ; In the formula, A one-dimensional standard Gaussian kernel, This indicates the time offset from the current time point; for The score of the relative importance of neural activity at any given moment; The standard deviation is 1.5. For the first ROI at time point The BOLD signal.
[0033] These reconstructed signals combine CRF-informed weighting and smoothing to improve the sensitivity and robustness of subsequent analyses.
[0034] S3. Using the activation probability of each ROI as input to the MRF model, the activation state of each ROI is optimized in each time frame by iteratively minimizing the energy function. The activation confidence of the ROI at time step is combined with the corresponding original time series to obtain a weighted enhanced time series. .
[0035] Brain region activity is influenced not only by itself but also by spatial dependence and interactions with neighboring regions. To capture spatial dependencies in time series, this invention employs an MRF model based on the Iterative Conditional Pattern (ICM) algorithm for spatial modeling. Unlike existing hard-threshold binarization methods, this embodiment uses a probabilistic soft threshold obtained through sigmoid transformation as a label to preserve activation uncertainty while embedding spatial functional relationships to ensure consistent and stable activation inference.
[0036] The specific steps are as follows: First, based on average value ( ) and standard deviation ( The statistical information calculated in the original text is used to apply the sigmoid function transformation to obtain the activation probability of each ROI: ; In the formula, It is a constant. .
[0037] Secondly, the probability distribution of each time frame The observations used as input to the MRF model. The spatial adjacency graph is... Based on the Pearson correlation between different ROI time series (positive correlation is set to 1, otherwise 0), it is integrated with the ICM method for frame-by-frame spatial inference. This process is mainly used to model the activation state of each ROI, where "inactive" (0) and "active" (1) represent two possible activation conditions.
[0038] Set energy function The energy function measures the rationality and spatial consistency of these activation states. It consists of two parts: single-point energy ( ) and paired energy ( ).
[0039] ; The single-point energy ( This reflects the cost associated with whether each ROI is activated. From the activation probability... The derived latent function quantifies this cost, ensures smooth transitions between states, and effectively manages uncertainty. The soft-thresholding method employed in this embodiment allows for a more gradual and stable inference process.
[0040] Single-point energy ( )for: ; ; In the formula, For the first ROI The energy function at a single point in time, where 1 indicates an active state; For the first ROI The energy function at a single point in time, where 0 represents the inactive state; For the first ROI in The activation probability at any given time.
[0041] The pairwise energy ( This is used to ensure the spatial consistency of the activation state of each ROI with its neighboring regions.
[0042] ; ; In the formula, For hyperparameters; For the first The ROI and the An adjacency matrix is constructed from ROIs, defining regions with both positive and both negative Pearson correlation coefficients as neighbors. For the first The number of neighboring nodes activated by each ROI; The energy of the activated region increases with the number of adjacent regions; To penalize inactive ROIs that are adjacent to the active region.
[0043] Iteratively minimize the energy function using the ICM algorithm The activation state of each ROI is optimized in each time frame. This process effectively combines local activation information with spatial consistency constraints, improving the accuracy of ROI activation inference.
[0044] Then, using the updated activation state, calculate... The confidence level of ROI activation at a given time quantifies the uncertainty of the activation state of each ROI at that specific time point.
[0045] ; After normalizing the confidence score of the ROI activation at that moment, it is combined with the corresponding original time series to generate a weighted enhanced time series. The sequence has been denoised and is more robust.
[0046] ; In the formula, For the first ROI in Activation confidence at any given moment.
[0047] S4. Constructing a weighted and enhanced time series based on manifold learning attention mechanism. and Functional connections result in enhanced features. It is used to construct brain network maps.
[0048] To better learn high-quality feature representations from enhanced time series data, this embodiment employs functional connections constructed from time series data enhanced by CRF and MRF models to represent the data and capture the complex spatiotemporal dependencies between time series data and brain structures. Then, the SimCLR contrastive learning framework is used to further enhance the discriminative power of the learned representations.
[0049] Specifically, First, Pearson correlation coefficient was used to analyze the enhanced time series. and Calculate the function link matrix , It is then projected into a new latent space through a nonlinear transformation.
[0050] ; In the formula, These are the features obtained after linear transformation; It is a linear transformation function; For activation functions; for , The abbreviation for indicates one of them (the entire two-layer transformation is collectively referred to as nonlinear transformation).
[0051] The projection features are represented as follows: ; ; In the formula, for Features obtained after linear transformation; for Features obtained after linear transformation; The number of subjects; The number of features; For feature dimensions.
[0052] Positive and negative sample pairs are defined based on the ROI correspondence: features from the same ROI are considered positive pairs, while all other features are considered negative pairs.
[0053] A similarity matrix is constructed and contrastive loss is applied to guide the learning of the joint discriminative representation of the data; ; ; In the formula, This is the similarity matrix calculated using cosine similarity. For the number of subjects (i.e., the number obtained) , (Number of matrices) Each type of matrix One, total indivual; This section discusses the method for calculating cosine similarity using the Euclidean norm.
[0054] Then, a novel attention mechanism based on manifold learning is employed to fuse time series features weighted by CRF and MRF models. First, a linear transformation is applied to the enhanced projected features of the time series. and To obtain Query(Q) and Key(K) representations, they are then mapped from Euclidean space to a Poincaré spherical manifold via an exponential mapping to simulate complex inter-region interactions in a non-Euclidean latent space.
[0055] ; In the formula, For popular spaces express; For popular spaces express; For exponential mapping, it represents the way a mapping is made from Euclidean space to manifold space.
[0056] To simulate non-Euclidean interactions, the hyperbolic distance between embedded Q and K on the Poincaré spherical manifold is calculated using the Arcosh-based geodesic distance formula. And use it as a measure of feature similarity (i and j represent two different features).
[0057] ; ; ; In the formula, Let Q be the hyperbolic distance between Q and K; It is a hyperbolic function; It is the Euclidean norm.
[0058] The expression for calculating manifold attention weights is as follows: ; Input and Firstly, through a shared dense layer An initial transformation is performed on the features, where features are fused in pairs at corresponding locations. The transformed features obtained through this step are then used with a feature fusion model and pre-computed attention weights. The combination resulted in enhanced features. It effectively integrates the enhanced features of CRF and MRF models from different perspectives.
[0059] The feature fusion model is as follows: ; in, ; In the formula, For enhanced features; For hyperparameters, ; These are the initial features for fusion; The functional connectivity matrix obtained by weighting the CRF; Functional connectivity matrix obtained by weighting MRF ; represents the initial feature matrix for fusion. The value; For shared dense layers; The functional connectivity matrix obtained by weighting the CRF ; Functional connectivity matrix obtained by weighting MRF .
[0060] like Figure 2 As shown, a Support Vector Machine (SVM) is used as the classifier. The decoded features are used as input to the SVM, and the SVM outputs the predicted probability of the target class. These probabilities are then compared with the true label y to calculate the cross-entropy loss. Combined with the contrastive loss, the final loss function is constructed. The model is trained multiple times. To better balance the contributions of classification and contrastive objectives, a weighting factor is used. Initialized to 0.1, the final result is used to construct a brain network map.
[0061] ; In the formula, Number of participants; For the first One real label; For the first Predicted labels.
[0062] Verification Example To verify the effectiveness of this invention, this validation example used fMRI data from three representative neuroimaging datasets at different scales: ABIDE (Autism Spectrum Disorder), ADHD-200 (ADHD, Attention Deficit Hyperactivity Disorder), and ADNI (AD, Alzheimer's Disease). For ASD, 884 subjects (408 with ASD and 476 CNs (control group)) from 20 ABIDE sites were used for binary classification. For ADHD, data from three locations—Beijing (194), KKI (83), and New York University (216)—totaling 493 subjects (218 with ADHD and 275 CNs) were used for binary classification. For AD, 161 participants from ADNI1 and ADNI2 (29 with AD, 32 with LMCI (late-stage mild cognitive impairment), 51 with EMCI (early-stage mild cognitive impairment), and 49 with CN) were used for four-class classification. All experiments were conducted on an NVIDIA GeForce RTX 3090 Ti GPU with 7-fold cross-validation. This validation example uses four metrics to measure the effectiveness of the model: accuracy (ACC), recall, F1 score, and precision. The results are shown in Tables 1-3.
[0063] Table 1. Results of ASD Assisted Diagnosis
[0064] Table 2. Results of ADHD Auxiliary Diagnosis
[0065] Table 3 AD Auxiliary Diagnosis Table
[0066] Experiments show that removing any module leads to varying degrees of performance degradation. By optimizing the CRF and MRF probabilistic graphical models, weights that better represent feature dependencies from both spatial and temporal perspectives are obtained. These weights determine how downstream classification tasks utilize temporal context to make decisions. Unlike traditional methods that use discrete features or labels, this invention represents the activation probability at each time step as a continuous value in the MRF feature function. Furthermore, Gaussian smoothing is applied to the discrete activation state weights learned by the CRF model, thereby improving the smoothness and continuity within the time series.
[0067] Manifold attention mechanisms significantly enhance the fusion of CRF and MRF-enhanced features, outperforming the use of manifold structures for individual features only. A multifaceted attention mechanism—combining spatial and temporal features of the brain—first refines the outputs of two probabilistic graphical models (CRF or MRF) using contrastive learning. The refined features are used as inputs for queries (Q) and keywords (K) and processed through the attention mechanism in the manifold space. Balanced attention weights computed from different perspectives are integrated into the features. In the middle, features This is achieved through a linear transformation of the refined features. Compared to models without a contrastive learning module, this invention not only reduces dimensionality and eliminates redundant information but also significantly improves performance. Specifically, the addition of the contrastive learning module improved accuracy by 3.05% on the ABIDE dataset, 0.39% on the ADHD dataset, and 6.21% on the ADNI dataset, thus enabling the correct identification of more individuals across all cases. To further enhance cross-branch functional integration, parameters... Adaptively controlling the contribution of multiple attentional information sources from each part reflects not only numerical weighting but also the rationality of branch selection based on different perspectives. Leveraging the geometry of the brain, it calculates the similarity between weighted time series from different ROIs and enhancement methods, promoting global fusion by more effectively modeling inter-regional relationships.
[0068] The CRF and MRF models in this application assign higher importance weights to features closely related to neurological diseases. While identifying overlapping key features, they also capture distinct, complementary information. This synergy and complementarity makes downstream tasks more effective. Ablation studies confirm that models augmented with CRF and MRF outperform models directly used for decoding tasks. High-quality data augmentation enhances the discriminative power of contrastive learning and allows manifold-based attention mechanisms to better focus on key features.
[0069] Example 2 This application provides a brain signal analysis system based on spatiotemporal feature enhancement for performing the above-mentioned method, including: Data acquisition module: used to acquire fMRI data and preprocess the data; The first enhancement module: Based on the preprocessed data, it learns the dynamic transformation relationship between activation labels through a CRF model, and extracts weighted functional connectivity features from the time series to obtain a weighted enhanced time series. ; The second enhancement module uses the activation probability of each ROI as input to the MRF model. By iteratively minimizing the energy function and considering spatial characteristics, it optimizes the activation state of each ROI in each time frame. The activation confidence of the ROI at time step is combined with the corresponding original time series to obtain a weighted enhanced time series. ; Fusion Module: Based on the manifold learning attention mechanism, a weighted augmented time series is obtained by fusing the first augmentation module and the second augmentation module. and The functional connections ultimately result in enhanced features. It is used to construct brain network maps.
[0070] This invention discloses a brain signal analysis system based on spatiotemporal feature enhancement, which can be installed in a computer device. The computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a brain signal analysis program based on spatiotemporal feature enhancement. The memory includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic storage, disk, optical disk, etc. The processor is the control core of the electronic device, connecting various components of the computer device via various interfaces and lines. It executes programs or modules stored in the memory and calls data stored in the memory to perform various functions of the computer device and process data.
[0071] The module described in this invention refers to a series of computer program segments that can be executed by the processor of a computer device and can perform a fixed function, and which are stored in the memory of the computer device.
[0072] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.
[0073] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall fall within the scope of the present invention.
Claims
1. A brain signal analysis method based on spatiotemporal feature enhancement, characterized in that, Includes the following steps: Acquire fMRI data and preprocess it; Based on the preprocessed data, a CRF model is used to learn the dynamic transformation relationship between activation labels, and weighted and functional connectivity features are extracted from the time series data to obtain a weighted enhanced time series. ; The activation probability of each ROI is used as input to the MRF model. By iteratively minimizing the energy function, the activation state of each ROI is optimized in each time frame. The activation confidence of the ROI at time step is combined with the corresponding original time series to obtain a weighted enhanced time series. ; A weighted and enhanced time series based on manifold learning attention mechanism was constructed. and Functional connections result in enhanced features. It is used to construct brain network maps.
2. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 1, characterized in that, By learning the dynamic transformation relationship between activation labels using a CRF model, and extracting time-series weighted and functional connectivity features, the following steps are included: Activation labels are generated based on statistical thresholds; Based on the activation state at each time point, the weights of the state features learned by the CRF model are... Feature vectors used in construction By weighting each component of the feature vector accordingly, a score reflecting the relative importance of neural activity at that moment is obtained. ; Using weighted scores Constructing weighted augmented time series .
3. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 2, characterized in that, The feature vector is extracted from the average BOLD time series of each ROI. The following is constructed from: ； In the formula, This represents the current value of the time series. The difference from the previous time point, for Time series values at time points, for Time series values at any given time; For adjacent time series values, This represents the window width and is set to a constant. This is the deviation term.
4. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 2, characterized in that, The weighted enhanced time series The calculation expression is: ； in, ； ； In the formula, A one-dimensional standard Gaussian kernel, This indicates the time offset from the current time point. , This represents the window width and is set to a constant. Standard deviation; for The score of the relative importance of neural activity at any given moment; for Activation tag at any moment; The weights obtained from training the CRF model. For the first One ROI, For feature vectors; For the first ROI Weights obtained during CRF model training at time step; For feature vectors; For the first ROI at time point The BOLD signal.
5. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 1, characterized in that, The energy function includes single-point energy and paired energy. The single-point energy is used to reflect the cost associated with whether each ROI is activated, and the paired energy is used to ensure the spatial consistency of the activation state of each ROI with its neighboring regions.
6. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 5, characterized in that, The expression for the energy function is: ； in, for: ； ； for: ； ； In the formula, For the first ROI The energy function at a single point in time, where 1 indicates an active state; For the first ROI The energy function at a single point in time, where 0 represents the inactive state; For the first ROI in The activation probability at any given time; The energy of the activated region increases with the number of adjacent regions; To penalize inactive ROIs adjacent to the active region; For hyperparameters; For the first The ROI and the An adjacency matrix is constructed from ROIs, defining regions with both positive and both negative Pearson correlation coefficients as neighbors. For the first The number of neighboring nodes activated by each ROI.
7. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 1, characterized in that, The weighted enhanced time series The expression is: ； in, ； In the formula, In order to be in Time of the first Confidence level of each ROI activation; For the first ROI at time point The BOLD signal; For the first ROI The energy function at a single point in time, where 0 represents the inactive state; For the first ROI The energy function at a single point in time, where 1 indicates an active state.
8. The brain signal analysis method based on spatiotemporal feature enhancement according to claim 1, characterized in that, Enhanced features are obtained through the following methods. : Using Pearson correlation coefficient from weighted enhanced time series and Transformation to obtain projection features and ; Applying linear transformations to enhance the projection features of time series and To obtain the Query and Key representations; The query and key are mapped from Euclidean space to a Poincaré spherical manifold through an exponential mapping to simulate complex inter-region interactions in a non-Euclidean latent space. Will and The input is fed into a shared dense layer to perform an initial transformation on the features, and then a feature fusion model with attention weights is used. Combining them yields enhanced features. .
9. A brain signal analysis method based on spatiotemporal feature enhancement according to claim 8, characterized in that, The feature fusion model is as follows: ； in, ； In the formula, For enhanced features; For hyperparameters, ; These are the initial features for fusion; The functional connectivity matrix obtained by weighting the CRF; The functional connectivity matrix obtained by weighting the MRF; The initial feature matrix for fusion The value; For shared dense layers; The functional connectivity matrix obtained by weighting the CRF ; Functional connectivity matrix obtained by weighting MRF .
10. A brain signal analysis system based on spatiotemporal feature enhancement, executing the brain signal analysis method based on spatiotemporal feature enhancement as described in any one of claims 1-9, characterized in that, include: Data acquisition module: used to acquire fMRI data and perform preprocessing; The first enhancement module: Based on the preprocessed data, it learns the dynamic transformation relationship between activation labels through a CRF model, and extracts weighted functional connectivity features from the time series to obtain a weighted enhanced time series. ; The second enhancement module uses the activation probability of each ROI as input to the MRF model and iteratively minimizes the energy function to optimize the activation state of each ROI in each time frame. The activation confidence of the ROI at time step is combined with the corresponding original time series to obtain a weighted enhanced time series. ; Fusion Module: Based on the manifold learning attention mechanism, a weighted augmented time series is obtained by fusing the first augmentation module and the second augmentation module. and The functional connections ultimately result in enhanced features. It is used to construct brain network maps.