Dual-atlas brain network feature extraction and data processing method and application thereof

Abstract

This invention proposes a dual-scale brain network feature extraction and data processing method and its application. Addressing the issues of semantic inconsistency in multi-scale fusion and the lack of neurobiological constraints in feature extraction, this method acquires resting-state fMRI data and constructs a dual-scale brain functional network. It introduces the Yeo7 network as a spatial prior, mapping ROIs to corresponding functional communities to obtain community feature sequences. Convolutional networks are used in parallel to extract global topological features, and Transformers are used to extract enhanced features of cross-community interactions. A two-level fusion mechanism is employed to obtain fusion features within each scale and the final fusion feature, which are then optimized using a multi-objective composite loss incorporating HSIC and scale-related losses. This method enhances the interpretability and generalization ability of the model and can be used for the auxiliary diagnosis of neuropsychiatric disorders such as autism spectrum disorder.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and auxiliary diagnosis of neuropsychiatric diseases, and in particular to a method for extracting and processing features of dual-map brain networks and its application. Background Technology

[0002] Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder, and resting-state functional magnetic resonance imaging has been widely used to uncover patterns of abnormal brain function.

[0003] Currently, fMRI-based ASD diagnostic methods typically rely on constructing brain functional networks from single brain region maps for feature extraction. Among these, the AAL atlas has advantages in anatomical consistency, while the CC200 atlas excels in the fineness of functional partitioning. To accommodate information at different spatial resolutions, some existing techniques attempt to simply stitch together or jointly model features from multiple atlases. However, different atlases differ fundamentally in their spatial definitions and functional partitioning; the AAL atlas focuses on macroscopic partitioning of anatomical structures, while the CC200 atlas focuses on function-driven microscopic partitioning, and the semantic alignment of their nodes is not aligned. Summary of the Invention

[0004] This invention provides a method for feature extraction and data processing of dual-map brain networks and its application. It addresses the problems of existing methods lacking a unified functional semantic alignment mechanism and failing to effectively utilize neurobiological structural priors during multi-scale brain map fusion, which leads to semantic conflicts and information redundancy in multi-scale features during fusion, thereby limiting the generalization ability and interpretability of the model.

[0005] The core technology of this invention is to introduce a prior functional network as a community prior for dual-map hierarchical mapping and cross-network interaction modeling, and to use a composite loss function containing independence and structural consistency constraints to optimize the multi-scale feature adaptive fusion of dynamic weights, thereby achieving low redundancy and high interpretability extraction of abnormal brain network patterns.

[0006] In a first aspect, the present invention provides a method for feature extraction and data processing of dual-map brain networks, the method comprising the following steps:

[0007] Acquire resting-state functional magnetic resonance imaging data, and construct first-scale brain functional networks and second-scale brain functional networks based on first-scale brain atlases and second-scale brain atlases, respectively. By introducing a prior functional network as a spatial prior, regions of interest in the first-scale and second-scale brain functional networks are mapped to corresponding functional communities to obtain community feature sequences at each scale. For each scale of brain functional network, global topological features are extracted, and community enhancement features are extracted based on community feature sequences. The global topological features and community enhancement features are then fused to obtain the internal fusion features for each scale. Independence constraints and structural consistency constraints are applied to the internal fusion features at each scale, and the internal fusion features at each scale are adaptively weighted and fused based on a dynamic weight strategy to obtain the final fusion features. Classification and recognition are then performed based on the final fusion features.

[0008] Furthermore, the region of interest is mapped to the corresponding functional community to obtain community feature sequences at various scales, including: For the first-scale brain functional network, based on the prior mapping rules of anatomy and function, the region of interest is assigned to the corresponding prior functional network; For second-scale brain functional networks, membership is quantified by calculating the spatial overlap rate between each region of interest and the prior functional network template in standard space, and functional labels are determined based on the principle of maximum membership. By using mapping relationships or functional labels, node features are aggregated hierarchically based on functional domains, and network-level features are extracted as community feature sequences.

[0009] Furthermore, global topological features are extracted, including: The functional connectivity matrix of the brain functional network is processed by a hierarchical convolutional network. The connection pattern features are extracted by an edge-to-edge convolutional layer, the edge features are mapped to node-level features by an edge-to-point convolutional layer, and the node information of the whole graph is aggregated into a global vector by a point-to-graph convolutional layer to obtain global topological features.

[0010] Furthermore, community enhancement features are extracted based on community feature sequences, including: The community feature sequence is input into an encoder containing a self-attention mechanism. The self-attention weight matrix is ​​used to dynamically capture the semantic associations between different functional communities, capture the high-order interaction relationships across functional communities, and output community-enhanced features.

[0011] Furthermore, global topological features and community enhancement features are fused to obtain internal fused features at various scales, including: The community enhancement features are pooled to obtain the enhancement feature vector. The enhancement feature vector is then concatenated with the global topological features and subjected to nonlinear transformation before being weighted and fused according to the calculated weights.

[0012] Furthermore, independence constraints and structural consistency constraints are imposed, including: A composite loss function, including independence constraint loss and scale correlation loss, is used for collaborative optimization. Among them, the independence constraint loss uses the Hilbert-Schmidt independence criterion to constrain the internal fusion features at different scales to remain orthogonal and independent; the scale correlation loss is used to constrain the statistical structure consistency of the internal fusion features at different scales at the sample level.

[0013] Furthermore, an adaptive weighted fusion of the internal fusion features at each scale is performed based on a dynamic weighting strategy, including: Calculate the performance evaluation index of each scale branch on the validation set, calculate the initial dynamic weight components using exponential normalization, and combine the learnable weight parameters with the historical learning experience scaling factor to calculate the final graph weights through a hybrid update strategy.

[0014] Secondly, the present invention provides a dual-map brain network feature extraction and data processing device, comprising: The network construction module is used to acquire resting-state functional magnetic resonance imaging data and construct a first-scale brain functional network and a second-scale brain functional network based on the first-scale brain atlas and the second-scale brain atlas, respectively. The community mapping module is used to introduce the prior functional network as a spatial prior, map the regions of interest in the first-scale brain functional network and the second-scale brain functional network to the corresponding functional communities, and obtain the community feature sequences at each scale. The feature extraction and fusion module is used to extract global topological features for each scale of brain functional network, extract community enhancement features based on community feature sequences, and fuse global topological features and community enhancement features to obtain internal fusion features for each scale. The classification and recognition module is used to apply independence constraints and structural consistency constraints to the internal fusion features at each scale, and to perform adaptive weighted fusion of the internal fusion features at each scale based on a dynamic weight strategy to obtain the final fusion features, and to perform classification and recognition based on the final fusion features.

[0015] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the above-described dual-map brain network feature extraction and data processing method.

[0016] Fourthly, the present invention provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the above-described dual-map brain network feature extraction and data processing method.

[0017] The main contributions and innovations of this invention are as follows: 1. Achieving complementary fusion of multi-scale brain functional information: By constructing a multi-scale brain network joint modeling method based on dual-brain atlases, the complementary advantages of AAL atlas in anatomical consistency and CC200 atlas in functional partitioning precision are fully utilized, overcoming the information loss problem caused by single atlas modeling; 2. Enhance the interpretability and neuroscience consistency of the model: Introduce a prior functional network as a community prior constraint, and map the original functional connections at the region of interest level to the functional community level with clear neurobiological significance, effectively overcoming the "black box" problem of traditional deep learning models; 3. Enhance the ability to model large-scale functional abnormalities: By modeling high-order interaction relationships across functional networks through community-level Transformer coding structures, it is possible to accurately capture abnormal communication patterns between large-scale brain networks such as default mode networks and executive control networks. 4. Optimize the discriminative power and robustness of feature representation: A hierarchical feature fusion mechanism is adopted to collaboratively model global connectivity features and community enhancement features within a single graph, while retaining the overall network topology and local functional module information, and adaptively allocating graph contributions through a dynamic dual-graph weighted fusion strategy; 5. Suppress feature redundancy and improve generalization performance: A composite constraint mechanism is designed, which includes Hilbert-Schmidt independence criterion loss and scale correlation loss. While forcing different map features to be orthogonal to remove redundancy, it maintains the consistency of statistical structure at the sample level, effectively alleviating the overfitting and feature conflict problems under small sample medical data.

[0018] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a dual-map brain network feature extraction and data processing method according to an embodiment of the present invention; Figure 2 This is a DACT module diagram of the dual-map brain network feature extraction and data processing method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0021] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0022] Example 1 This invention aims to propose a method for feature extraction and data processing of dual-map brain networks, specifically, referring to... Figure 1 The method includes the following steps: Step 1: Construction and preprocessing of multi-scale brain functional networks.

[0023] In this embodiment, raw resting-state functional magnetic resonance imaging (rs-fMRI) data of the subjects were first acquired. The data was obtained from the publicly available large-scale multicenter dataset ABIDE I. Data processing employed a standard preprocessing pipeline (Configurable Pipeline for the Analysis of Connectomes, CPAC), including head movement correction, skull dissection, MNI152 spatial normalization based on the ANTs algorithm, and bandpass filtering from 0.01 to 0.1 Hz. Subsequently, time-series signals of n anatomical brain regions were extracted based on a specified brain atlas (AAL / CC200). .

[0024] After preprocessing, corresponding brain topology maps G1 and G2 are constructed for the first-scale brain atlas (anatomical scale, such as the AAL atlas) and the second-scale brain atlas (functional scale, such as the CC200 atlas), respectively. The connectivity strength between brain regions n is calculated using the Pearson correlation coefficient, and a brain functional connectivity matrix C is constructed.

[0025] in: This represents the functional connectivity strength between brain region i and brain region j, and its value range is [ 1,1]; T represents the total number of time points scanned; t is the sampling index of the time series, t=1,2,...,T; x i (t) and x j (t) represent the observed BOLD signal values ​​of brain regions i and j at time point t, respectively; The mean values ​​of brain regions i and j over the entire time series are respectively defined as:

[0026] Step 2: Community feature mapping based on spatial priors.

[0027] like Figure 2 As shown, this method introduces the Yeo7 resting-state network distribution as prior knowledge of the community, mapping Regions of Interest (ROI) nodes at different scales to a unified functional community space. A mapping probability matrix M is constructed to quantify the membership relationship between the i-th brain region and the k-th Yeo7 functional network, with dimensions of 116×7 (for the AAL map) and 200×7 (for the CC200 map). Different mapping rules are formulated for the first scale (AAL, 116 brain regions) and the second scale (CC200, 200 brain regions): (1) Prior-based mapping was used for the first-scale brain functional network (AAL atlas). By establishing an anatomy-function lookup table, 116 anatomical brain regions were directly assigned to their corresponding Yeo7 functional networks, and the matrix elements... It is a binary representation (1 represents membership, 0 represents non-membership).

[0028] (2) For the second-scale brain functional network (CC200 atlas), the membership degree is quantified by calculating the spatial overlap ratio between each ROI and the Yeo7 template in the standard space:

[0029] in: Represents the number of voxels; This represents the set of voxels for the i-th region of interest in standard space. Let represent the set of voxels of the k-th prior functional network in standard space; ∩ represents the intersection of the sets. This method allows an ROI to span multiple functional networks and precisely quantifies its spatial weight distribution among the networks through numerical values. In the final label assignment, this method follows the "maximum membership principle," that is, taking the index with the largest value in the row vector of the M matrix as the functional label for that brain region. For CC200, this maximum overlap ratio is used as the confidence score of the mapping; for AAL, the uniqueness and authority of the mapping are ensured through prior logic.

[0030] Subsequently, using the mapping operator For the first Node features of each graph branch Functional domain-based hierarchical aggregation was performed to extract seven biologically significant network-level features. :

[0031] in: Pool uses mean pooling to ensure that the average activation level of the network region is extracted; The initial feature vector of the i-th brain region node is obtained by linear mapping or direct extraction from the time series signal X.

[0032] Step 3: Layered feature extraction and fusion.

[0033] For each scale of brain functional network, global and community features are extracted in parallel. Global topological features and community enhancement features are then fused to obtain the internal fused features for each scale. Specifically, this includes: (1) Community sequence Input a community-level Transformer encoder. Self-attention weight matrix. The calculation is as follows:

[0034]

[0035] in: Represents a learnable linear projection weight matrix used to map input features to a query (Q), key (K), and value (V) vector space; To calculate the attention weight matrix, the Softmax function is applied to represent the dimension of the attention head. Finally, the self-attention weight matrix is ​​used. AND-value matrix Multiplication yields a weighted representation of community features. .

[0036] (2) To enhance the model's ability to capture complex patterns in different subspaces, this model employs a dual-head attention mechanism (num_heads=2). The outputs of each head are concatenated and then subjected to a nonlinear transformation via a Position-wise Feed-Forward Network (FFN):

[0037] in: x is the input feature vector; This is the first weight matrix, used for linear transformation, mapping the input feature x to an intermediate representation space; Add to the first bias term The results are used to adjust the calculation results; ReLU (Rectified Linear Unit) is the most commonly used nonlinear activation function. Its function is to truncate the part of the input that is less than zero to zero, while keeping the part that is greater than zero unchanged. This is the second weight matrix, used to further map the result after the activation function to the output space; Add to the second bias term The results are used to adjust the output.

[0038] In addition, to prevent the vanishing gradient problem during deep network training, residual connections and layer normalization are introduced in each sub-layer:

[0039] in: This represents the input features of the current sublayer. This represents the output features after residual connection and layer normalization.

[0040] (3) Generate community enhancement features: After the above-mentioned multi-layer processing of the Transformer encoder, the original community feature sequence is generated. Evolved into community-enhanced feature sequences Each of these enhanced community features All representations are recalculated under a global perspective. It dynamically absorbs information from other functional communities through the self-attention matrix A, thereby achieving modeling of cross-network interactions.

[0041] (4) The ROI-level global encoding module utilizes the BrainNetCNN architecture to obtain the original brain network functional connectivity matrix. Global topological features are extracted. The input matrix C is a normalized Pearson correlation coefficient matrix, representing the edge weights between N ROIs in the whole brain. This process captures spatial features at different levels through hierarchical convolution: Edge-to-edge (E2E) convolutional layers utilize row convolutional kernels. Convolution kernel Extract connection mode:

[0042] in: and These are row convolution kernels and column convolution kernels, respectively; The activation function is represented by , and in this embodiment, the LeakyReLU activation function with a negative slope of 0.33 is used.

[0043] Edge-to-Node (E2N) convolutional layers use edge-to-node convolutional kernels. Mapping edge features to node (ROI) level features:

[0044] This step aggregates the connection distribution of each node, transforming the feature map dimension from 64×N×N to 128×N×1, and identifies the ROIs that play a key pivotal role in the whole-brain network.

[0045] Node-to-Graph convolutional layers use node-to-graph convolutional kernels. (dmodel) aggregates all node information in the graph into a global vector:

[0046] The final generated This represents the macroscopic topology of whole-brain networks at the current scale. To accelerate model convergence and suppress overfitting, this invention integrates a batch normalization (BN) layer after each convolutional layer. Specifically, the convolutional output first undergoes mean centering and variance scaling through the BN layer, followed by a nonlinear transformation using the LeakyReLU activation function (negative slope α=0.33). Furthermore, at key nodes of feature aggregation (such as after the E2N layer), a Dropout regularization layer (dropout rate p=0.5) is added. By randomly blocking the weight transfer of some neurons during processing, this forces the model to learn more robust feature representations, further suppressing the risk of overfitting in small sample data environments.

[0047] (5) Intra-scale vertical interaction and dynamic fusion of maps: The extracted global features are integrated using a hierarchical fusion module. and enhance community-level features Deep integration is performed: 'a' represents the graph branch. The enhanced feature sequence is obtained by average pooling, resulting in an enhanced feature vector, which is then combined with the global features through a nonlinear transformation. The features are then weighted and fused according to the calculated weights ω to obtain the fused feature vector. :

[0048] Step 4: Multi-objective constraint optimization and adaptive fusion.

[0049] To address the overfitting and feature redundancy issues in medical image data with small sample sizes, a composite loss function is used for collaborative optimization. The HSIC loss is primarily used to constrain the final feature vector after the fusion of different map branches. and , characteristic matrix The fusion features corresponding to the AAL map branches after internal hierarchical encoding ,and Fusion features corresponding to the CC200 scale branch This enhances feature independence; while scale correlation loss constrains the consistency of statistical structure at the sample level across different scales, thus achieving a balance between "redundancy removal" and "maintaining structural consistency." The total loss function is defined as follows:

[0050] The components and parameters are defined as follows: The cross-entropy loss is used to ensure recognition accuracy, and its formula is:

[0051] in: For real labels, Predict the probability that a sample belongs to the ASD class for the model; The Hilbert-Schmidt Independence Criterion (HSIC) loss is used to force the first-scale feature. With second-scale features Maintaining orthogonality and independence ensures that the model can extract complementary biological information for feature matrices at different scales. ( (For batch size), construct a kernel matrix K using a linear kernel and combine it with a centered matrix H for processing:

[0052] in: This represents the trace operation of the matrix; B is the batch size, and this loss term forces internal fusion features at different scales. and Maintain orthogonal independence; and It is a linear kernel matrix, and its calculation method is as follows: ; For a centered matrix, its definition is: ,in It is the identity matrix, and 1 is a column vector of all 1s.

[0053] It is scale-related loss, achieved through covariance. and standard deviation Constraining features at different scales Consistency of covariance structure among samples:

[0054] In addition, hyperparameters These are the independence constraint weights and scale correlation weights, used to balance classification accuracy and feature independence. for The regularization term is used to penalize excessively large model parameters and prevent overfitting. The regularization coefficient is . This is the set of learnable parameters for the model.

[0055] Preferably, the model adjusts the dynamic weights of the graph in real time based on the performance on the validation set, achieving adaptive integration of multi-scale information. Finally, the fused features are input into a multilayer perceptron (MLP) classifier to identify abnormal patterns.

[0056] Step 5: Performance verification.

[0057] The fused global feature vector F is input into a multi-layer perceptron (MLP) classifier to identify abnormal patterns in the subject's brain network state using nonlinear mapping.

[0058] The following hyperparameter design was used during the experiment: 1. Optimizer: The Adam optimizer is used, with a learning rate set to 10. -4 ; 2. Batch size: The batch size for each training session is 16; 3. Dropout ratio: Set to 0.5 to prevent overfitting; 4. Loss function hyperparameters: Set λ1 to 0.001 and λ2 to 0.01 to balance the influence of different loss terms.

[0059] Furthermore, this invention introduces a dynamic weight update mechanism based on validation set performance. Specifically, after each epoch, the contribution of each graph branch is redistributed based on its performance on the validation set. The specific implementation is as follows: First, calculate the performance evaluation metric of the a-th graph branch on the validation set. ( To amplify performance differences and ensure non-negative weights, exponential normalization (Softmax mapping) is used to calculate the initial dynamic weight components. :

[0060] in: The temperature parameter is set to 1.0 to smooth the weight distribution. It is a minimal constant ( This is used to ensure the stability of numerical calculations. Then, it is combined with learnable weight parameters from the model. The ratio factor between dynamic performance and historical learning experience (Set to 0.5), which ultimately affects the graph weights during forward propagation. A hybrid update strategy is adopted:

[0061] To verify the robustness of this data processing method in feature extraction and pattern recognition, an automated computational processing experiment was conducted using the publicly available offline brain image feature dataset (ABIDE I). In the experiment, the data was divided into two categories based on their feature labels for pattern recognition testing. The specific feature processing and analysis results are shown in Table 1 below: Table 1

[0062] As can be seen, the indicators in the table are the core evaluation metrics for the binary classification model: ACC (Accuracy) reflects the overall percentage of correctly classified samples, AUC (Area Under the ROC Curve) measures the model's ability to distinguish between positive and negative samples, SEN (Sensitivity) represents the proportion of true positive samples correctly identified, and SPE (Specificity) represents the proportion of true negative samples correctly identified. In this comparison, the DACT method of this invention achieved the best performance in ACC, AUC, and SEN, only slightly lagging behind the comparative method in SPE, and its overall performance is superior to the comparative method.

[0063] Example 2 Based on the same concept, this invention also proposes a dual-graph brain network feature extraction and data processing device, including: a network construction module, a community mapping module, a feature extraction and fusion module, and a classification and recognition module.

[0064] The network construction module acquires rs-fMRI data and constructs a multi-scale brain functional network. The community mapping module performs functional semantic space alignment based on the Yeo7 prior. The feature extraction and fusion module processes the BrainNetCNN global branch and the Transformer community enhancement branch in parallel and completes the deep integration of features at each scale. The classification and recognition module, guided by a multi-objective loss function (HSIC constraint and correlation constraint), utilizes a dynamic weight update mechanism to achieve adaptive weighted classification. The specific execution flow and parameter settings of each module of this device are consistent with the steps described in Example 1.

[0065] Example 3 This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0066] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0067] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0068] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0069] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the dual-map brain network feature extraction and data processing methods in the above embodiments.

[0070] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0071] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0072] Input / output device 408 is used to input or output information.

[0073] Example 4 This embodiment also provides a readable storage medium storing a computer program, the computer program including program code for controlling the process to execute the process, the process including the dual-map brain network feature extraction and data processing method according to Embodiment 1.

[0074] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0075] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0076] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 1 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0077] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for feature extraction and data processing of dual-map brain networks, characterized in that, Includes the following steps: Acquire resting-state functional magnetic resonance imaging data, and construct first-scale brain functional networks and second-scale brain functional networks based on first-scale brain atlases and second-scale brain atlases, respectively. A prior functional network is introduced as a spatial prior, and the regions of interest in the first-scale brain functional network and the second-scale brain functional network are mapped to the corresponding functional communities to obtain the community feature sequences at each scale. For each scale of brain functional network, global topological features are extracted, and community enhancement features are extracted based on the community feature sequence. The global topological features and the community enhancement features are then fused to obtain the internal fusion features for each scale. Independence constraints and structural consistency constraints are applied to the internal fusion features of each scale, and adaptive weighted fusion is performed on the internal fusion features of each scale based on a dynamic weight strategy to obtain the final fusion features. Classification and recognition are then performed based on the final fusion features.

2. The method for feature extraction and data processing of dual-map brain networks as described in claim 1, characterized in that, Mapping the region of interest to the corresponding functional community, and obtaining community feature sequences at various scales, including: For the first-scale brain functional network, based on the prior mapping rules of anatomy and function, the region of interest is assigned to the corresponding prior functional network; For the second-scale brain functional network, the membership degree is quantified by calculating the spatial overlap rate between each region of interest and the prior functional network template in the standard space, and the functional label is determined based on the principle of maximum membership degree. The node features are aggregated hierarchically based on functional domains using mapping relationships or functional labels, and network-level features are extracted as the community feature sequence.

3. The method for feature extraction and data processing of dual-map brain networks as described in claim 1, characterized in that, Extract global topological features, including: The functional connectivity matrix of the brain functional network is processed using a hierarchical convolutional network. The global topological features are obtained by sequentially extracting connectivity pattern features through edge-to-edge convolutional layers, mapping edge features to node-level features through edge-to-point convolutional layers, and aggregating the node information of the entire graph into a global vector through point-to-graph convolutional layers.

4. The method for feature extraction and data processing of dual-map brain networks as described in claim 1, characterized in that, Based on the community feature sequence, community enhancement features are extracted, including: The community feature sequence is input into an encoder containing a self-attention mechanism. The self-attention weight matrix is ​​used to dynamically capture the semantic associations between different functional communities, capture the high-order interaction relationships across functional communities, and output the community enhancement features.

5. The method for feature extraction and data processing of dual-map brain networks as described in claim 1, characterized in that, The global topological features and the community enhancement features are fused to obtain internal fused features at each scale, including: The community enhancement features are pooled to obtain an enhancement feature vector. The enhancement feature vector is then concatenated with the global topological features and subjected to a nonlinear transformation before being weighted and fused according to the calculated weights.

6. The method for feature extraction and data processing of dual-map brain networks as described in claim 1, characterized in that, Imposing independence constraints and structural consistency constraints, including: A composite loss function, including independence constraint loss and scale correlation loss, is used for collaborative optimization. The independence constraint loss uses the Hilbert-Schmidt independence criterion to constrain the internal fusion features at different scales to remain orthogonal and independent; the scale correlation loss is used to constrain the statistical structural consistency of the internal fusion features at different scales at the sample level.

7. The method for feature extraction and data processing of dual-map brain networks as described in claim 1, characterized in that, Adaptive weighted fusion of the internal fusion features at each scale based on a dynamic weighting strategy includes: Calculate the performance evaluation index of each scale branch on the validation set, calculate the initial dynamic weight components using exponential normalization, and combine the learnable weight parameters with the historical learning experience scaling factor to calculate the final graph weights through a hybrid update strategy.

8. A device for feature extraction and data processing of dual-map brain networks, characterized in that, include: The network construction module is used to acquire resting-state functional magnetic resonance imaging data and construct a first-scale brain functional network and a second-scale brain functional network based on the first-scale brain atlas and the second-scale brain atlas, respectively. The community mapping module is used to introduce a prior functional network as a spatial prior, map the regions of interest in the first-scale brain functional network and the second-scale brain functional network to the corresponding functional communities, and obtain the community feature sequences at each scale. The feature extraction and fusion module is used to extract global topological features for each scale of brain functional network, extract community enhancement features based on the community feature sequence, and fuse the global topological features and the community enhancement features to obtain the internal fusion features of each scale. The classification and recognition module is used to apply independence constraints and structural consistency constraints to the internal fusion features of each scale, and to perform adaptive weighted fusion of the internal fusion features of each scale based on a dynamic weight strategy to obtain the final fusion features, and to perform classification and recognition based on the final fusion features.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the dual-map brain network feature extraction and data processing method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, the computer program including program code for controlling the process to execute the process, the process including the dual-map brain network feature extraction and data processing method according to any one of claims 1 to 7.

Images