Brain function abnormality identification method, device, equipment and medium

Abstract

The disclosed embodiments provide a method, apparatus, device, and medium for identifying abnormal brain function. The method comprises: inputting acquired target brain image data into a brain function identification model, the brain function identification model comprising an input layer, a graph enhancement layer, a graph feature processing layer, and a feature classification layer; in the input layer, constructing an original graph structure based on the target brain image data and extracting a dynamic time feature matrix; in the graph enhancement layer, performing graph enhancement processing on the original graph structure according to the dynamic time feature matrix to obtain an enhanced graph structure; in the graph feature processing layer, obtaining a graph feature representation of the enhanced graph structure; and in the feature classification layer, performing classification prediction of brain function abnormalities based on the graph feature representation to obtain a brain function abnormality identification result. Thus, based on the static spatial features and the dynamic time feature matrix, the graph structure more accurately depicts changes in brain function, thereby improving the accuracy of brain function identification using rs-fMRI data.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a method, device, equipment, and medium for identifying abnormal brain function. Background Art

[0002] Functional magnetic resonance imaging (r-fMRI) is a non-invasive and friendly technology for collecting brain information. The resting-state functional magnetic resonance imaging (rs-fMRI) data collected by this technology can be used to identify abnormal brain function, such as for auxiliary diagnosis of brain diseases such as Alzheimer's disease and schizophrenia.

[0003] In related technologies, a functional connectivity network (FCN) based on a static functional connectivity matrix is constructed based on rs-fMRI data, which simplifies the complex brain network into a static topological structure, and identifies brain dysfunction based on the static topological structure.

[0004] However, the above methods are not accurate enough in depicting changes in brain function, resulting in insufficient accuracy in identifying abnormal brain function based on rs-fMRI data. Summary of the Invention

[0005] The present disclosure provides a method, apparatus, device and medium for identifying abnormal brain function to solve the problem that existing methods are not accurate enough in depicting changes in brain function, resulting in insufficient accuracy in identifying abnormal brain function based on rs-fMRI data.

[0006] In a first aspect, the present disclosure provides a method for identifying abnormal brain function, comprising: acquiring target brain image data, wherein the target brain image data is resting-state functional magnetic resonance imaging (rs-fMRI) data to be identified; inputting the target brain image data into a brain function recognition model, wherein the brain function recognition model comprises an input layer, a graph enhancement layer, a graph feature processing layer, and a feature classification layer; in the input layer, based on the target brain image data, constructing an original graph structure of a first brain function connection network and extracting a dynamic time feature matrix of the first brain function connection network; in the graph enhancement layer, performing graph enhancement processing on the original graph structure according to the dynamic time feature matrix to obtain an enhanced graph structure of the first brain function connection network; in the graph feature processing layer, performing feature processing on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure; in the feature classification layer, performing classification prediction of brain function abnormalities based on the graph feature representation to obtain a brain function abnormality recognition result of the target brain image data.

[0007] In one possible implementation, the original graph structure includes an original node feature matrix, an original edge weight matrix, and an original adjacency matrix; in the graph enhancement layer, the original graph structure is subjected to graph enhancement processing according to the dynamic time feature matrix to obtain an enhanced graph structure of the first brain functional connection network, including: in the graph enhancement layer, based on the original node feature matrix, the original edge weight matrix, and the dynamic time feature matrix, the original adjacency matrix is optimized to obtain an optimized adjacency matrix; the original adjacency matrix is replaced by the optimized adjacency matrix to obtain the enhanced graph structure.

[0008] In one possible implementation, during the training process of the brain function recognition model, the loss function used includes a first loss function constructed based on a first information bottleneck optimization objective, in which the first mutual information is maximized and the second mutual information is minimized; wherein, the first mutual information refers to the mutual information between the optimized adjacency matrix in the original graph structure of the second brain function connection network and the node feature similarity matrix of the second brain function connection network, and the node feature similarity matrix is obtained based on the edge weight matrix in the original graph structure of the second brain function connection network and the node feature matrix in the original graph structure of the second brain function connection network; wherein, the second mutual information refers to the mutual information between the optimized adjacency matrix and the dynamic time feature matrix in the second brain function connection network; wherein, the second brain function connection network is a brain function connection network constructed based on training samples.

[0009] In one possible implementation, the first loss function includes a cross-entropy loss term and a relative entropy loss term; the cross-entropy loss term is a variational lower bound formula constructed based on the first mutual information, and the relative entropy loss term is a variational upper bound formula constructed based on the second mutual information; in the process of solving the first loss function, the optimized adjacency matrix is adjusted using a reparameterization technique.

[0010] In one possible implementation, in the input layer, based on the target brain image data, the original graph structure of the first brain functional connection network is constructed and the dynamic time feature matrix of the first brain functional connection network is extracted, including: in the input layer, extracting the BOLD signal time series corresponding to multiple brain regions from the target brain image data; performing feature extraction on the BOLD signal time series corresponding to the multiple brain regions to obtain the original node feature matrix; performing global correlation analysis on the BOLD signal time series corresponding to the multiple brain regions to obtain the original edge weight matrix; performing local correlation analysis based on a sliding time window on the BOLD signal time series corresponding to the multiple brain regions to obtain the dynamic time feature matrix; and initializing the original adjacency matrix to a fully connected matrix.

[0011] In one possible implementation, the graph feature processing layer includes a graph convolution module and a feature projection module; in the graph feature processing layer, feature processing is performed on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure, including: inputting the enhanced graph structure into the graph convolution module, and performing feature extraction and feature aggregation on the enhanced graph structure in the graph convolution module to obtain a graph convolution feature representation; inputting the graph convolution feature representation into the feature projection module, and performing feature dimensionality reduction on the graph convolution feature representation in the feature projection module to obtain the graph feature representation.

[0012] In one possible implementation, during the training process of the brain function recognition model, the loss function used includes a second loss function constructed based on a second information bottleneck optimization objective, in which the third mutual information is maximized and the fourth mutual information is maximized; wherein, the third mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the first original graph structure; wherein, the fourth mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the second original graph structure; wherein, the first original graph structure and the second original graph structure are original graph structures constructed based on different training samples.

[0013] In a second aspect, an embodiment of the present disclosure provides a device for identifying abnormal brain function, comprising: an acquisition unit for acquiring target brain image data, wherein the target brain image data is rs-fMRI data to be identified; an input unit for inputting the target brain image data into a brain function recognition model, wherein the brain function recognition model comprises an input layer, a graph enhancement layer, a graph feature processing layer, and a feature classification layer; a construction and extraction unit for constructing, in the input layer, an original graph structure of a first brain function connection network based on the target brain image data and extracting a dynamic time feature matrix of the first brain function connection network; a graph enhancement unit for performing graph enhancement processing on the original graph structure according to the dynamic time feature matrix in the graph enhancement layer to obtain an enhanced graph structure of the first brain function connection network; a graph feature extraction unit for performing feature processing on the enhanced graph structure in the graph feature processing layer to obtain a graph feature representation of the enhanced graph structure; and a feature classification unit for performing classification prediction of brain function abnormalities based on the graph feature representation in the feature classification layer to obtain a brain function abnormality recognition result of the target brain image data.

[0014] In a third aspect, an embodiment of the present disclosure provides an electronic device comprising: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method for identifying abnormal brain function as described in the first aspect above.

[0015] In a fourth aspect, an embodiment of the present disclosure provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the method for identifying abnormal brain function as described in the first aspect above is implemented.

[0016] In a fifth aspect, an embodiment of the present disclosure provides a computer program product, which includes computer-executable instructions. When a processor executes the computer-executable instructions, the method for identifying abnormal brain function as described in the first aspect above is implemented.

[0017] In the brain function abnormality identification method, device, equipment and medium provided by the embodiments of the present disclosure, in the brain function identification model, based on the rs-fMRI data to be identified, the original graph structure of the brain function connection network is constructed, and the dynamic time feature matrix of the brain function connection network is extracted. The original graph structure reflects the static spatial characteristics of the brain, and the dynamic time feature matrix reflects the dynamic time characteristics of the brain. This method takes into account the static structure and dynamic changes of brain function, so that the brain function connection network can more accurately depict the brain function; under the guidance of the dynamic time feature matrix, the original graph structure is enhanced to obtain an enhanced graph structure. Compared with the original graph structure, the enhanced graph structure more accurately depicts the dynamic changes of brain function; the enhanced graph structure is feature processed, and based on the extracted graph feature representation, the brain function abnormality identification result is predicted, thereby improving the accuracy of brain function abnormality identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0019] Figure 1 A flowchart of a method for identifying abnormal brain function provided by an embodiment of the present disclosure;

[0020] Figure 2 This is a diagram illustrating an example of the structure of a brain function recognition model provided in an embodiment of the present disclosure;

[0021] Figure 3 A structural block diagram of a device for identifying abnormal brain function provided by an embodiment of the present disclosure;

[0022] Figure 4 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present disclosure without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0024] It should be understood that the terms "first", "second", "third", etc. in the embodiments of the present invention are used to distinguish similar or similar objects or entities, and do not necessarily mean to limit a specific order or sequence. Unless otherwise noted, it should be understood that the terms used in this way can be interchangeable under appropriate circumstances, for example, they can be implemented in an order other than those given in the diagrams or descriptions of the embodiments of the present disclosure.

[0025] First, the technical problems and technical concepts solved by the embodiments of the present disclosure are explained.

[0026] In the method of identifying brain dysfunction based on brain imaging data, a functional connectivity network (FCN) based on a static fully connected functional connectivity matrix is constructed based on resting-state functional Magnetic Resonance Imaging (rs_fMRI) data. The FCN is then processed through graph neural networks (GNN) to obtain brain dysfunction identification results.

[0027] Although rs_fMRI data is resting-state functional magnetic resonance imaging (FMRI), the brain also experiences ongoing neuronal activity during this resting state. This neuronal activity can cause changes in brain function, giving rs_fMRI data temporal dynamics. The aforementioned approach simplifies the complex brain into a static topological connectivity diagram using a static fully connected functional connectivity matrix. This approach fails to account for the temporal dynamics of rs_fMRI data, resulting in inaccurate depictions of brain function changes and, consequently, inaccurate identification of abnormal brain function.

[0028] In order to solve the above problems, the embodiments of the present disclosure construct an original graph structure representing static spatial features and a dynamic time feature matrix in the process of constructing a brain function recognition network. The original graph structure is enhanced based on the dynamic time matrix to obtain an enhanced graph structure, which effectively improves the accuracy of the enhanced graph structure in depicting the static characteristics of brain function and the dynamic changes of brain function. Based on the enhanced graph structure, the brain function recognition model can better capture the complex dynamic interactions between brain intervals, effectively improving the accuracy of the brain function recognition model in identifying brain function abnormalities in rs_fMRI data.

[0029] Figure 1 This is a flow chart of a method for identifying abnormal brain function provided by an embodiment of the present disclosure. Figure 1 As shown, the method for identifying abnormal brain function includes:

[0030] S101, acquiring target brain image data, where the target brain image data is rs-fMRI data to be identified.

[0031] To distinguish the rs-fMRI data used as training samples, the target rs-fMRI data is referred to as target brain imaging data. Target brain imaging data is time-series data, consisting of functional magnetic resonance imaging (fMRI) images acquired continuously over a period of time. Because fMRI images display three-dimensional spatial information about the brain, target brain imaging data also contains rich temporal information.

[0032] In this embodiment, the target brain image data may be obtained from a brain imaging data acquisition device or from a database; or the target brain image data may be received from input by relevant researchers.

[0033] S102, inputting the target brain image data into a brain function recognition model, which includes an input layer, an image enhancement layer, an image feature processing layer, and a feature classification layer.

[0034] Among them, the brain function recognition model is a neural network model for graph structure data processing. Based on training samples, the brain function recognition model can be fully pre-trained, so that the brain function recognition model has the ability to accurately identify brain function abnormalities from rs-fMRI data.

[0035] In this embodiment, the target brain image data is input into the brain function recognition model, and firstly input into the input layer of the brain function recognition model for preliminary data processing.

[0036] S103, in the input layer, based on the target brain image data, constructing the original graph structure of the first brain functional connection network and extracting the dynamic time feature matrix of the first brain functional connection network.

[0037] Among them, in order to distinguish it from the brain functional connection network constructed during the training process, the brain functional connection network constructed based on the target brain imaging data is called the first brain functional connection network.

[0038] In this embodiment, in the input layer, brain region signal extraction can be performed on the target brain image data to obtain blood oxygen level dependent (BOLD) signal time series corresponding to multiple brain regions. The BOLD signal time series of a brain region reflects the neuronal activity of the brain region at a specific moment and the changes in neuronal activity of the brain region over time. It can also reflect the static relationship between brain functions and the dynamic interaction of brain functions between multiple brain regions. Therefore, by performing dynamic correlation analysis and static correlation analysis on the BOLD signal time series corresponding to multiple brain regions, the original graph structure of the first brain functional connectivity network can be constructed and the dynamic time feature matrix of the first brain functional connectivity network can be extracted.

[0039] S104 , in the graph enhancement layer, performing graph enhancement processing on the original graph structure according to the dynamic time feature matrix to obtain an enhanced graph structure of the first brain functional connection network.

[0040] In this embodiment, considering that the dynamic time feature matrix has dynamic time features that the original graph structure does not have, the original graph structure and the dynamic time feature matrix are input into the graph enhancement layer. In the graph enhancement layer, under the guidance of the dynamic time feature matrix, the original graph structure is enhanced to obtain an enhanced graph structure. The enhanced graph structure takes into account both static and dynamic features when depicting the brain function interactions of multiple brain regions, thereby improving the accuracy of depicting the static characteristics and dynamic changes of brain functions through the graph structure.

[0041] S105 , in the graph feature processing layer, feature processing is performed on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure.

[0042] In this embodiment, the enhanced graph structure is input into the graph feature processing layer. The graph feature processing layer may include a convolution layer capable of performing feature processing on the graph structure. In the graph feature processing layer, the enhanced graph structure can be feature processed through the convolution layer to obtain a graph feature representation of the enhanced graph structure.

[0043] S106 , in the feature classification layer, performing classification prediction of brain function abnormality based on the graph feature representation to obtain a brain function abnormality recognition result of the target brain image data.

[0044] The brain function abnormality recognition result of the target brain image data may include whether the target brain image data has brain function abnormality and / or the type of brain function abnormality of the target brain image data.

[0045] In this embodiment, the graph feature representation is input into the feature classification layer. In the feature classification layer, probability information related to brain function abnormality can be determined based on the graph feature representation. The probability information related to brain function abnormality may include: a probability value corresponding to brain function abnormality and a probability value corresponding to normal brain function, and / or probability values corresponding to multiple types of brain function abnormality. Based on the probability information related to brain function abnormality, the brain function abnormality recognition result is determined: when the probability value corresponding to the brain function abnormality is greater than the probability value corresponding to normal brain function, the brain function abnormality recognition result is determined to be the presence of brain function abnormality in the target brain image data; otherwise, the brain function abnormality recognition result is determined to be the absence of brain function abnormality in the target brain image data; and / or, the brain function abnormality recognition result is determined to be the target abnormality type, the target abnormality type being the brain function abnormality type corresponding to the maximum probability value among the probability values corresponding to the multiple types of brain function abnormality, or, the brain function abnormality recognition result is determined to include probability values corresponding to the multiple types of brain function abnormality.

[0046] As an example, abnormal brain function may include brain diseases, and the types of abnormal brain function may include brain disease types, such as Alzheimer's disease, bipolar disorder, depression, etc., so as to improve the accuracy and reliability of brain disease identification through the brain function abnormality identification method provided in the embodiments of the present application.

[0047] In an embodiment of the present application, in the process of constructing a brain function recognition network, an original graph structure representing static spatial features and a dynamic time feature matrix are constructed. By performing graph enhancement on the original graph structure based on the dynamic time matrix, an enhanced graph structure is obtained, thereby improving the accuracy of the enhanced graph structure in depicting the static characteristics of brain function and the dynamic changes of brain function. Based on the enhanced graph structure, the graph feature processing layer in the brain function recognition model can better capture the complex dynamic interactions between brain intervals and obtain accurate and rich graph feature representations. Classification and prediction of brain function abnormalities are performed based on the graph feature representation, thereby effectively improving the accuracy of the brain function recognition model in identifying brain function abnormalities in rs_fMRI data.

[0048] In some embodiments, the original graph structure includes an original node feature matrix, an original edge weight matrix, and an original adjacency matrix. The multiple nodes in the original graph structure represent multiple brain regions, and the multiple nodes correspond one-to-one to the multiple brain regions; the original node feature matrix includes node features corresponding to the multiple nodes, that is, feature information of the BOLD signal time series corresponding to the multiple brain regions; the original edge weight matrix includes the global static weights of the edges between the multiple nodes, and the global static weights of the edges between the multiple nodes represent the global static correlation between the multiple brain regions; the multiple element values in the original adjacency matrix represent the edge connection relationship between the multiple nodes. For example, if the element value is zero, it means that there is no edge between the two nodes corresponding to the element value, and if the element value is 1, it means that there is an edge between the two nodes corresponding to the element value. Thus, the signal characteristics of multiple brain regions are described through the original node feature matrix, and the degree of correlation between multiple brain regions is described by combining the original edge weight matrix and the original adjacency matrix, thereby improving the accuracy of the original graph structure in depicting brain function.

[0049] In the case where the original graph structure includes the original node feature matrix, the original edge weight matrix and the original adjacency matrix, in a possible implementation method, the original graph structure of the first brain functional connection network is constructed in the input layer based on the target brain image data and the dynamic time feature matrix of the first brain functional connection network is extracted, that is, S103 includes: in the input layer, extracting the BOLD signal time series corresponding to multiple brain regions from the target brain image data; performing feature extraction on the BOLD signal time series corresponding to the multiple brain regions to obtain the original node feature matrix; performing global correlation analysis on the BOLD signal time series corresponding to the multiple brain regions to obtain the original edge weight matrix; performing local correlation analysis based on a sliding time window on the BOLD signal time series corresponding to the multiple brain regions to obtain the dynamic time feature matrix; and initializing the original adjacency matrix to a fully connected matrix. Therefore, when constructing the brain functional connectivity network, the global static correlation of the BOLD signal time series corresponding to multiple brain regions (i.e., the global spatial connection of multiple brain regions) and the local dynamic correlation of the BOLD signal time series corresponding to multiple brain regions (i.e., the local temporal dynamics of multiple brain regions) are considered simultaneously, which improves the accuracy of the brain functional connectivity network in depicting changes in brain function.

[0050] In this implementation, in the process of performing global correlation analysis on the BOLD signal time series corresponding to multiple brain regions to obtain the original edge weight matrix, for each brain region, the BOLD signal time series corresponding to the brain region is regarded as a whole, and the global static correlation coefficient between the multiple brain regions is obtained by calculating the correlation coefficient between the BOLD signal time series corresponding to each two brain regions. The global static correlation coefficient between the multiple brain regions is determined as the global static weight of the edges between the multiple nodes, and the original edge weight matrix is formed by the global static weights of the edges between the multiple nodes. In the process of performing a sliding time window-based local correlation analysis on the BOLD signal time series corresponding to multiple brain regions to obtain a dynamic time feature matrix, BOLD signal time series segments of multiple brain regions located within the same time window can be extracted from the BOLD signal time series corresponding to the multiple brain regions; the local dynamic correlation coefficients between the BOLD signal time series segments of the multiple brain regions located within the same time window are calculated, and the local dynamic correlation coefficients are determined as the local dynamic weights of the edges between the multiple nodes; the time window is slid, and the process of calculating the local dynamic coefficients is repeated to ultimately obtain multiple local dynamic weights of the edges between each two nodes; the dynamic time feature matrix is formed from the multiple local dynamic weights of the edges between each two nodes. The original adjacency matrix is initialized as a fully connected matrix, and the element values of the original adjacency matrix can be initialized to 1.

[0051] In the above-mentioned global correlation analysis and local correlation analysis, the correlation coefficient used may be a Pearson correlation coefficient.

[0052] Optionally, the target brain imaging data can be preprocessed in the input layer (e.g., format conversion, time series segmentation, denoising, etc.). After preprocessing, the brain is divided into multiple regions, and the BOLD signal time series are extracted from each region to obtain BOLD signal time series corresponding to the multiple regions. Thus, preprocessing improves the data quality of the target brain imaging data and the accuracy of the BOLD signal time series corresponding to the multiple regions.

[0053] Optionally, multi-band low-frequency amplitude (ALFF) extraction is performed on the BOLD signal time series corresponding to multiple brain regions to obtain multi-band ALFF features corresponding to each of the multiple brain regions. The original node feature matrix is composed of these multi-band ALFF features corresponding to each of the multiple brain regions. This allows for a more comprehensive capture of brain activity characteristics across different frequency bands from the BOLD signal time series corresponding to multiple brain regions.

[0054] In the case where the original graph structure includes the original node feature matrix, the original edge weight matrix and the original adjacency matrix, in a possible implementation method, the original graph structure is subjected to graph enhancement processing in the graph enhancement layer according to the dynamic time feature matrix to obtain an enhanced graph structure of the first brain functional connection network, that is, S104 includes: in the graph enhancement layer, based on the original node feature matrix, the original edge weight matrix and the dynamic time feature matrix, the original adjacency matrix is optimized to obtain an optimized adjacency matrix; the original adjacency matrix is replaced by the optimized adjacency matrix to obtain an enhanced graph structure. Therefore, under the guidance of the original node feature matrix, the original edge weight matrix and the dynamic time feature matrix, the original adjacency matrix is optimized so that the original adjacency matrix can incorporate dynamic time features without causing significant changes in the graph structure, allowing the brain function recognition model to stably learn key features from the graph structure; moreover, based on the original node feature matrix, the original edge weight matrix and the dynamic time feature matrix, the original adjacency matrix is optimized to achieve graph structure enhancement based on the soft edge-dropping strategy, which can avoid destroying key brain function connections compared to graph structure enhancement based on random edge deletion.

[0055] In this implementation, the neural network structure in the graph enhancement layer processes the original node feature matrix, the original edge weight matrix, and the dynamic time feature matrix to generate a mask matrix corresponding to the original adjacency matrix. The mask matrix is associated with the information to be filtered in the original adjacency matrix. The original adjacency matrix is then optimized based on the mask matrix to obtain an optimized adjacency matrix. The original adjacency matrix is then replaced with the optimized adjacency matrix to obtain an enhanced graph structure, which includes the original node feature matrix, the original edge weight matrix, and the optimized adjacency matrix.

[0056] Optionally, the neural network structure in the graph enhancement layer can be called an information bottleneck optimization module. The information bottleneck optimization module is trained based on a loss function constructed based on the information bottleneck optimization objective. The information bottleneck optimization module trained with this loss function can generate a mask matrix that retains information in the original graph structure that is useful for identifying brain dysfunction and removes redundant information, thereby improving the graph structure's ability to express brain function and thereby improving the accuracy and stability of identifying brain dysfunction. The specific training process can be found in the description of the subsequent embodiments and will not be repeated here.

[0057] Optionally, the graph enhancement layer may further include a graph enhancement strategy module. In the graph enhancement strategy model, an optimized adjacency matrix is obtained by performing operations on the mask matrix and the original adjacency matrix.

[0058] In one possible implementation, the graph feature processing layer includes a graph convolution module and a feature projection module. Based on this, in the graph feature processing layer, feature processing is performed on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure, that is, S105 includes: inputting the enhanced graph structure into the graph convolution module, and in the graph convolution module, performing feature extraction and feature aggregation on the enhanced graph structure to obtain a graph convolution feature representation; inputting the graph convolution feature representation into the feature projection module, and in the feature projection module, performing feature dimensionality reduction on the graph convolution feature representation to obtain a graph feature representation. Thus, by utilizing the feature extraction capability of the graph convolution module, a rich and accurate graph convolution feature representation is extracted from the enhanced graph structure, and then through the feature projection module, the dimension of the graph convolution feature representation is made to meet the requirements of the subsequent feature classification layer.

[0059] In this implementation, the enhanced graph structure is input into the graph convolution module, that is, the original node feature matrix, the original edge weight matrix, and the optimized adjacency matrix are input into the graph convolution module. The original node feature matrix, the original edge weight matrix, and the optimized adjacency matrix are subjected to graph convolution calculation to obtain high-dimensional features of multiple nodes in the enhanced graph structure, namely, the graph convolution feature representation. During the graph convolution calculation process, the original edge weight matrix can be first optimized based on the optimized adjacency matrix to obtain the optimized edge weight matrix; then, feature aggregation is performed based on the optimized edge weight matrix and the original node feature matrix. During the feature aggregation process, for any node, the node features of the neighboring nodes of the node are aggregated into the node features of the node through graph convolution. This process can be understood as the aggregation of local neighborhood information. After obtaining the graph convolution feature representation, the graph convolution feature representation is input into the feature projection module. In the feature projection module, feature dimensionality reduction is achieved by performing feature projection on the graph convolution feature representation to obtain the graph feature representation.

[0060] During the training process of the brain function recognition model, the model is trained based on training samples and a loss function of the brain function recognition model. The training samples include rs-fMRI data, which is unlabeled training data. The loss function of the brain function recognition model is constructed based on information bottleneck optimization theory. The processing of the training samples by the brain function recognition model is the same as the processing of the target brain image data by the brain function recognition model, except that the corresponding loss value is calculated based on the loss function. Therefore, the following provides relevant embodiments of the loss function of the brain function recognition model.

[0061] In some embodiments, during the training process of the brain function recognition model, the loss function used includes a first loss function constructed based on a first information bottleneck optimization objective, in which the first mutual information is maximized and the second mutual information is minimized. The first mutual information refers to the mutual information between the optimized adjacency matrix in the original graph structure of the second brain functional connectivity network and the node feature similarity matrix of the second brain functional connectivity network. The node feature similarity matrix is obtained based on the edge weight matrix in the original graph structure of the second brain functional connectivity network and the node feature matrix in the original graph structure of the second brain functional connectivity network. The second mutual information refers to the mutual information between the optimized adjacency matrix and the dynamic time feature matrix in the second brain functional connectivity network. The second brain functional connectivity network is a brain functional connectivity network constructed based on training samples. Thus, through the first loss function constructed based on the information bottleneck theory, an adjacency matrix optimization based on a soft edge discarding strategy is achieved, resulting in an adjacency matrix that is more suitable for brain function abnormality identification. While reducing redundant information in the graph structure, the necessary brain functional connections in the graph structure are retained, thereby improving the stability and accuracy of brain function abnormality identification based on the graph structure.

[0062] The basic idea of information bottleneck optimization theory is to retain the information most relevant to the target task during information transmission, while removing irrelevant or redundant information. For graph structure optimization, information bottleneck optimization theory can be expressed as the following optimization goal:

[0063]

[0064] in, Representation diagram With classification targets Maximizing the mutual information between can ensure that the graph representation contains sufficient task-related information; Representation diagram With the original data Minimizing this term can reduce redundant information and improve the generalization ability of the task processing model (in this embodiment, the brain function recognition model). It is a hyperparameter that controls information constraints and is used to adjust the balance between information preservation and information compression.

[0065] In this embodiment, based on the optimization objective of the information bottleneck optimization theory, a first loss function is constructed to maximize the relevant information between the optimized adjacency matrix and the node features, so that the optimized adjacency matrix is more consistent with the actual distribution of the node features; and minimize the relevant information between the optimized adjacency matrix and the dynamic time feature matrix to reduce the redundant information between the optimized adjacency matrix and the dynamic time feature matrix.

[0066] In this embodiment, in order to distinguish it from the brain functional connection network in the aforementioned embodiment, the brain functional connection network constructed based on the training samples is referred to as the second brain functional connection network. For the second brain functional connection network, the first information bottleneck optimization objective can be expressed as:

[0067]

[0068] in, Represents the optimized adjacency matrix in the second brain functional connection network, The node feature similarity matrix representing the second brain functional connection network is obtained based on the edge weight matrix and the node feature matrix. The element values in the node feature similarity matrix represent the feature similarity between two brain regions. express and The mutual information between them is the first mutual information; represents the dynamic time feature matrix, express A and W The mutual information between them is the second mutual information.

[0069] Optionally, during training, a convolution operation is performed on the edge weight matrix in the original graph structure of the second brain functional connectivity network and the node feature matrix in the original graph structure of the second brain functional connectivity network to obtain a convolved node feature matrix; and a dot product is performed on the convolved node feature matrix to obtain a node feature similarity matrix. Thus, the accuracy of the node feature similarity matrix is improved through the convolution and dot product operations.

[0070] In this optional method, The element value in row i and column j can be expressed as:

[0071]

[0072] Represents brain area and The feature similarity between and Respectively represent the i-th element and the j-th element in the convolved node feature matrix, It is a Sigmoid function, and the dot product operation is implemented by the Sigmoid function.

[0073] Optionally, the first loss function includes a cross-entropy loss term and a relative entropy loss term; the cross-entropy loss term is a variational lower bound formula constructed based on the first mutual information, and the relative entropy loss term is a variational upper bound formula constructed based on the second mutual information. Thus, the first information bottleneck optimization objective is approximated as a solvable variational lower bound formula, resulting in a first loss function suitable for the brain function recognition model.

[0074] In this optional approach, the variational lower bound formula of the first mutual information can be expressed as:

[0075]

[0076] in, represents the optimized adjacency matrix The element in row i and column j in Yes The variational approximation of Represents the adjacency matrix given the optimization Conditional distribution of node feature similarity matrix under the condition of . express The prior distribution of .

[0077] The above variational lower bound formula can be further simplified as:

[0078]

[0079] In this optional method, the variational upper bound formula of the second mutual information can be expressed as:

[0080]

[0081] Among them, here Yes Prior distribution of Variational approximation of .

[0082] Therefore, the first loss function can be expressed as:

[0083]

[0084] in, represents the first loss function, is the cross entropy loss term in the first loss function, is the relative entropy loss term in the first loss function, that is, KL divergence.

[0085] In summary, the first information bottleneck optimization objective can be expressed as:

[0086]

[0087] in, represents the learnable parameters of the graph enhancement layer, Represents the learnable parameters of the graph feature processing layer, in particular, Represents the learnable parameters of the graph convolution module in the graph feature processing layer.

[0088] Optionally, in the process of solving the first loss function, the reparameterization trick can be used to adjust the optimized adjacency matrix so that the loss value of the first loss function during training can be efficiently calculated by gradient descent, thereby improving the training efficiency and enabling the brain function recognition model to be trained end-to-end.

[0089] In this alternative, the reparameterization technique is used to adjust the optimized adjacency matrix, which can be expressed as:

[0090]

[0091] in, , represents a uniformly distributed random variable; is the Sigmoid function, which is used to make the update process of the adjacency matrix differentiable; Represents the temperature parameter, which is used to control the degree of soft edge discarding. For example, a small value Make the distribution of the adjacency matrix closer to binary, and the larger the value Make the edge weights in the graph structure smoother.

[0092] In some embodiments, during the training of the brain function recognition model, the loss function used includes a second loss function constructed based on a second information bottleneck optimization objective, wherein the third mutual information is maximized and the fourth mutual information is maximized in the second information bottleneck optimization objective; the third mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the first original graph structure; the fourth mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the second original graph structure. The first original graph structure and the second original graph structure are original graph structures constructed based on different training samples.

[0093] In this embodiment, considering that enhancing the graph structure can easily cause the model to learn graph feature representations with large deviations, especially when the enhanced graph structure undergoes significant changes, the graph feature representations may shift, adversely affecting the model performance of the brain function recognition model. To address this problem, a second information bottleneck optimization objective is proposed. In this second information bottleneck optimization objective, the graph feature representation of the original graph structure and the graph feature representation of the enhanced graph structure corresponding to the original graph structure are regarded as a pair of positive samples, and the graph feature representation of the original graph structure and the graph feature representation of the enhanced graph structure corresponding to other original graph structures are regarded as a pair of negative samples; the second information bottleneck optimization objective is optimized, that is, the mutual information between positive samples is maximized and the mutual information between negative samples is minimized. Among them, the third mutual information is maximized, that is, the mutual information between positive samples is maximized; and the fourth mutual information is minimized, that is, the mutual information between negative samples is minimized.

[0094] In this embodiment, a second loss function is constructed based on the second information bottleneck optimization objective, and adversarial optimization of the second information bottleneck optimization objective is achieved through the second loss function. Specifically, a batch training method is used to train the brain function recognition model. In each training session of the brain function recognition model, based on batch training samples, multiple original graph structures and enhanced graph structures corresponding to the multiple original graph structures are constructed. Through the graph feature processing layer in the brain function recognition model, graph feature representations corresponding to the multiple original graph structures and graph feature representations of the enhanced graph structures corresponding to the multiple original graph structures are generated. Based on the graph feature representations corresponding to the multiple original graph structures and the graph feature representations of the enhanced graph structures corresponding to the multiple original graph structures, a loss value of the second loss function is calculated, and the loss value of the second loss function is used to adjust the parameters of the brain function recognition model to achieve training of the brain function recognition model.

[0095] The graph feature processing layer of the brain function recognition model may include an adversarial comparison module. The calculation process of the loss value of the second loss function in the above-mentioned training process and the parameter adjustment of the brain function recognition model may be implemented through the adversarial comparison module.

[0096] Through a second loss function, adversarial contrastive learning is introduced during the training of the brain function recognition model. While enhancing the original graph structure to obtain an enhanced graph structure, adversarial contrastive learning maximizes the similarity between the graph feature representations before and after enhancement, ensuring the stability of the graph features before and after enhancement. Furthermore, through this adversarial contrastive learning strategy, the graph enhancement strategy of the brain function recognition model is dynamically adjusted during training to prevent the enhancement operation from significantly affecting the final graph feature learning. This ensures that the graph feature representation of the enhanced graph structure maintains a certain degree of similarity with that of the corresponding original graph structure. This effectively improves the generalization ability of the brain function recognition model, reduces its dependence on specific datasets, avoids biases in the graph feature learning caused by graph structure enhancement, and enhances the accuracy and robustness of the model in identifying brain function abnormalities. In particular, the model maintains stable recognition performance across different types of brain function abnormalities, such as Alzheimer's disease, bipolar disorder, and depression.

[0097] Optionally, the second loss function can be expressed as:

[0098]

[0099] in, is the number of graphs in a batch, i.e. the number of samples; represents the original graph structure, Indicates that the original graph structure is enhanced to obtain an enhanced graph structure; Represents the feature processing operation on the graph structure to obtain the corresponding graph feature representation, Indicates that the feature processing of the original graph structure is performed to obtain the graph feature representation of the original graph structure. Indicates that feature processing is performed on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure; The graph feature representation representing the original graph structure of the i-th training sample, The graph feature representation representing the enhanced graph structure of the i-th training sample, Indicates the Graph feature representation of the enhanced graph structure of training samples; Represents the cosine similarity function.

[0100] Optionally, the second information bottleneck optimization objective is expressed as:

[0101]

[0102] in, Represents the learnable parameters of the graph feature processing layer, in particular, Represents the learnable parameters of the feature projection module in the graph feature processing layer, Represents the learnable parameters of the graph enhancement layer.

[0103] Combining the first information bottleneck optimization goal and the second information bottleneck optimization goal of the aforementioned embodiment, the final optimization goal of the brain function model can be obtained as follows:

[0104]

[0105] In some embodiments, the training process of the brain function recognition model may include: obtaining training samples, which are rs-fMRI data; inputting the training samples into the brain function recognition model, which includes an input layer, a graph enhancement layer, a graph feature processing layer, and a feature classification layer; in the input layer, based on the training samples, constructing the original graph structure of the training brain function connection network and extracting the dynamic time feature matrix of the training brain function connection network; in the graph enhancement layer, performing graph enhancement processing on the original graph structure of the training brain function connection network according to the dynamic time feature matrix of the training brain function connection network to obtain an enhanced graph structure of the training brain function connection network; in the graph feature processing layer, performing feature processing on the enhanced graph structure of the training brain function connection network to obtain an enhanced graph structure of the training brain function connection network. a graph feature representation of the structure; in the feature classification layer, based on the graph feature representation of the enhanced graph structure of the trained brain functional connection network, classification prediction of brain function abnormalities is performed to obtain brain function abnormality recognition results of the target brain imaging data; based on the original graph structure of the brain functional connection network, the dynamic time feature matrix of the trained brain functional connection network and the first loss function, a first loss value is determined, and / or, based on the graph feature representation of the original graph structure of the trained brain functional connection network, the graph feature representation of the enhanced graph structure of the trained brain functional connection network and the second loss function, a second loss value is determined, the graph feature representation of the original graph structure of the trained brain functional connection network is obtained by inputting the original graph structure into the graph feature processing layer for processing; according to the first loss value and / or the second loss value, the brain function recognition model is adjusted.

[0106] As an example, Figure 2 This is an example diagram of the structure of the brain function recognition model provided by the embodiment of the present disclosure. Figure 2 As shown, the brain function recognition model includes an input layer, a graph enhancement layer, a graph feature processing layer and a classification layer. The input layer includes a preprocessing module and a graph construction module. The graph enhancement layer includes an information bottleneck optimization module and a graph enhancement strategy module. The graph feature processing layer includes a graph convolution module, a feature projection module and an adversarial contrast module, wherein the feature projection module and the adversarial contrast module form an adversarial contrast layer.

[0107] based on Figure 2The brain function recognition model shown in the figure may include the following steps: inputting the rs-fMRI data into an input layer; preprocessing the rs-fMRI data in an output layer using a preprocessing module; constructing the original graph structure of the brain functional connectivity network in a graph construction module based on the preprocessed rs-fMRI data and extracting a dynamic temporal feature matrix of the brain functional connectivity network, wherein the original graph structure includes an original node feature matrix, an original adjacency matrix, and an original edge weight matrix; optimizing the original adjacency matrix based on the original node feature matrix in the original graph structure, the original node feature matrix in the original graph structure, and the dynamic temporal feature matrix of the brain functional connectivity network using an information bottleneck optimization module and a graph enhancement strategy module to obtain an optimized adjacency matrix; and calculating the loss value of the first loss function in the aforementioned embodiment based on the optimized adjacency matrix, the original node feature matrix, and the original edge weight matrix during training; and obtaining an enhanced graph structure based on the optimized adjacency matrix. The enhanced graph structure is input into a graph feature processing layer, and feature processing is performed by a graph convolution module and a feature projection module in the graph feature processing layer to obtain a graph feature representation of the enhanced graph structure. The graph feature representation is input into the feature classification layer to ultimately determine the abnormal brain function recognition result. During the training process, the original graph structure can also be input into the graph feature processing layer, where the graph convolution module and feature projection module perform feature processing to obtain the graph feature representation of the original graph structure. In the adversarial comparison module, the loss value of the second loss function in the aforementioned embodiment can be calculated based on the graph feature representation of the original graph structure and the graph feature representation of the enhanced graph structure. The loss value of the first loss function and the loss value of the second loss function are combined to adjust the parameters of the brain function recognition model to obtain a trained brain function recognition module.

[0108] Corresponding to the method for identifying abnormal brain function in the above embodiment, Figure 3 This is a structural block diagram of a device for identifying abnormal brain function provided by an embodiment of the present disclosure. For ease of explanation, only the parts related to the embodiment of the present disclosure are shown. Figure 3The brain function abnormality identification device 300 includes: an acquisition unit 301, which is used to acquire target brain image data, where the target brain image data is rs-fMRI data to be identified; an input unit 302, which is used to input the target brain image data into a brain function identification model, where the brain function identification model includes an input layer, a graph enhancement layer, a graph feature processing layer and a feature classification layer; a construction and extraction unit 303, which is used to construct the original graph structure of the first brain function connection network based on the target brain image data in the input layer and extract the dynamic time feature matrix of the first brain function connection network; a graph enhancement unit 304, which is used to perform graph enhancement processing on the original graph structure according to the dynamic time feature matrix in the graph enhancement layer to obtain an enhanced graph structure of the first brain function connection network; a graph feature extraction unit 305, which is used to perform feature processing on the enhanced graph structure in the graph feature processing layer to obtain a graph feature representation of the enhanced graph structure; and a feature classification unit, which is used to perform classification processing on the graph feature representation in the feature classification layer to determine the brain function abnormality identification result of the target brain image data.

[0109] In some embodiments, the original graph structure includes an original node feature matrix, an original edge weight matrix, and an original adjacency matrix; the graph enhancement unit 304 is specifically used to: in the graph enhancement layer, optimize the original adjacency matrix based on the original node feature matrix, the original edge weight matrix, and the dynamic time feature matrix to obtain an optimized adjacency matrix; replace the original adjacency matrix with the optimized adjacency matrix to obtain an enhanced graph structure.

[0110] In some embodiments, during the training process of the brain function recognition model, the loss function used includes a first loss function constructed based on a first information bottleneck optimization objective, in which the first mutual information is maximized and the second mutual information is minimized; wherein the first mutual information refers to the mutual information between the optimized adjacency matrix in the original graph structure of the second brain function connection network and the node feature similarity matrix of the second brain function connection network, and the node feature similarity matrix is obtained based on the edge weight matrix in the original graph structure of the second brain function connection network and the node feature matrix in the original graph structure of the second brain function connection network; wherein the second mutual information refers to the mutual information between the optimized adjacency matrix and the dynamic time feature matrix in the second brain function connection network; wherein the second brain function connection network is a brain function connection network constructed based on training samples.

[0111] In some embodiments, the first loss function includes a cross-entropy loss term and a relative entropy loss term; the cross-entropy loss term is a variational lower bound formula constructed based on the first mutual information, and the relative entropy loss term is a variational upper bound formula constructed based on the second mutual information; in the process of solving the first loss function, the optimized adjacency matrix is adjusted using a reparameterization technique.

[0112] In some embodiments, the input unit 302 is specifically used to: extract, in the input layer, the BOLD signal time series corresponding to multiple brain regions from the target brain image data; perform feature extraction on the BOLD signal time series corresponding to the multiple brain regions to obtain an original node feature matrix; perform global correlation analysis on the BOLD signal time series corresponding to the multiple brain regions to obtain an original edge weight matrix; perform local correlation analysis based on a sliding time window on the BOLD signal time series corresponding to the multiple brain regions to obtain a dynamic time feature matrix; and initialize the original adjacency matrix to a fully connected matrix.

[0113] In some embodiments, the graph feature processing layer includes a graph convolution module and a feature projection module; the graph enhancement unit 304 is specifically used to: input the enhanced graph structure into the graph convolution module, and perform feature extraction and feature aggregation on the enhanced graph structure in the graph convolution module to obtain a graph convolution feature representation; input the graph convolution feature representation into the feature projection module, and perform feature dimensionality reduction on the graph convolution feature representation in the feature projection module to obtain a graph feature representation.

[0114] In some embodiments, during the training process of the brain function recognition model, the loss function used includes a second loss function constructed based on a second information bottleneck optimization objective, in which the third mutual information is maximized and the fourth mutual information is maximized; the third mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the first original graph structure; the fourth mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the second original graph structure; the first original graph structure and the second original graph structure are original graph structures constructed based on different training samples.

[0115] The brain function abnormality identification device provided in this embodiment can be used to implement the technical solution of the embodiment of the above-mentioned brain function abnormality identification method. Its implementation principle and technical effects are similar and will not be repeated here.

[0116] refer to Figure 4 , which shows a schematic structural diagram of an electronic device 400 suitable for implementing an embodiment of the present disclosure. The electronic device 400 may be a terminal device or a server. The terminal device may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, personal digital assistants (PDAs), tablet computers (Portable Android Devices, PADs), portable multimedia players (PMPs), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The electronic device shown is only an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0117] like Figure 4 As shown, electronic device 400 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 401, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 402 or programs loaded from storage device 408 into random access memory (RAM) 403. RAM 403 also stores various programs and data required for the operation of electronic device 400. Processing device 401, ROM 402, and RAM 403 are interconnected via bus 404. An input / output (I / O) interface 405 is also connected to bus 404.

[0118] Typically, the following devices may be connected to the I / O interface 405: an input device 406 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 407 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 408 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 409. The communication device 409 may allow the electronic device 400 to communicate with other devices wirelessly or by wire to exchange data. Figure 4 The electronic device 400 is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead.

[0119] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network via the communication device 409, or installed from the storage device 408, or installed from the ROM 402. When the computer program is executed by the processing device 401, the above-mentioned functions defined in the method of the embodiment of the present disclosure are performed.

[0120] It should be noted that the computer-readable medium described above in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable storage media may include, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or components, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such a propagated data signal may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wire, optical cable, RF (radio frequency), or any suitable combination thereof.

[0121] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0122] The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device executes the method shown in the above embodiment.

[0123] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0124] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the module, program segment, or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of the boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0125] The units described in the embodiments of the present disclosure may be implemented in software or hardware. In some cases, the name of a unit does not limit the unit itself. For example, the acquisition unit may be described as "a unit for acquiring a page image and page description text of a web page to be detected."

[0126] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0127] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0128] The above description is merely a preferred embodiment of the present disclosure and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the scope of the above disclosure. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in this disclosure.

[0129] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details have been included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0130] Although the subject matter has been described in language specific to structural features and / or methodological logical acts, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms of implementing the claims.

Claims

1. A method for identifying abnormal brain function, characterized in that: include: Acquiring target brain image data, wherein the target brain image data is resting-state functional magnetic resonance imaging (rs-fMRI) data to be identified; Inputting the target brain image data into a brain function recognition model, wherein the brain function recognition model includes an input layer, an image enhancement layer, an image feature processing layer, and a feature classification layer; In the input layer, based on the target brain image data, an original graph structure of a first brain functional connectivity network is constructed and a dynamic time feature matrix of the first brain functional connectivity network is extracted, wherein the original graph structure includes an original node feature matrix, an original edge weight matrix, and an original adjacency matrix; In the graph enhancement layer, the original graph structure is subjected to graph enhancement processing according to the dynamic time feature matrix to obtain an enhanced graph structure of the first brain functional connection network; In the graph feature processing layer, feature processing is performed on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure; In the feature classification layer, classification prediction of brain function abnormality is performed based on the graph feature representation to obtain a brain function abnormality recognition result of the target brain image data; In the graph enhancement layer, the original graph structure is enhanced according to the dynamic time feature matrix to obtain an enhanced graph structure of the first brain functional connection network, including: In the graph enhancement layer, the original adjacency matrix is optimized based on the original node feature matrix, the original edge weight matrix, and the dynamic time feature matrix to obtain an optimized adjacency matrix; Replacing the original adjacency matrix with the optimized adjacency matrix to obtain the enhanced graph structure; In the input layer, based on the target brain image data, constructing the original graph structure of the first brain functional connection network and extracting the dynamic time feature matrix of the first brain functional connection network includes: In the input layer, blood oxygen level-dependent BOLD signal time series corresponding to multiple brain regions are extracted from the target brain image data; Performing feature extraction on the BOLD signal time series corresponding to the multiple brain regions to obtain the original node feature matrix; Performing global correlation analysis on the BOLD signal time series corresponding to the multiple brain regions to obtain the original edge weight matrix; performing a local correlation analysis based on a sliding time window on the BOLD signal time series corresponding to the multiple brain regions to obtain the dynamic time feature matrix; Initialize the original adjacency matrix to a fully connected matrix.

2. The method for identifying abnormal brain function according to claim 1, wherein: During the training of the brain function recognition model, the loss function used includes a first loss function constructed based on a first information bottleneck optimization objective, in which the first mutual information is maximized and the second mutual information is minimized; The first mutual information refers to the mutual information between the optimized adjacency matrix in the original graph structure of the second brain functional connection network and the node feature similarity matrix of the second brain functional connection network, and the node feature similarity matrix is obtained according to the edge weight matrix in the original graph structure of the second brain functional connection network and the node feature matrix in the original graph structure of the second brain functional connection network; The second mutual information refers to the mutual information between the optimized adjacency matrix and the dynamic time feature matrix in the second brain functional connection network; The second brain functional connection network is a brain functional connection network constructed based on training samples.

3. The method for identifying abnormal brain function according to claim 2, wherein: The first loss function includes a cross entropy loss term and a relative entropy loss term; The cross entropy loss term is a variational lower bound formula constructed based on the first mutual information, and the relative entropy loss term is a variational upper bound formula constructed based on the second mutual information; In the process of solving the first loss function, the optimized adjacency matrix is adjusted using a reparameterization technique.

4. The method for identifying abnormal brain function according to any one of claims 1 to 3, characterized in that: The graph feature processing layer includes a graph convolution module and a feature projection module; In the graph feature processing layer, feature processing is performed on the enhanced graph structure to obtain a graph feature representation of the enhanced graph structure, including: Inputting the enhanced graph structure into the graph convolution module, and performing feature extraction and feature aggregation on the enhanced graph structure in the graph convolution module to obtain a graph convolution feature representation; The graph convolution feature representation is input into the feature projection module, and in the feature projection module, feature dimension reduction is performed on the graph convolution feature representation to obtain the graph feature representation.

5. The method for identifying abnormal brain function according to any one of claims 1 to 3, characterized in that: During the training of the brain function recognition model, the loss function used includes a second loss function constructed based on a second information bottleneck optimization objective, in which the third mutual information is maximized and the fourth mutual information is maximized; The third mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the first original graph structure; The fourth mutual information refers to the mutual information between the graph feature representation of the first original graph structure and the graph feature representation of the enhanced graph structure corresponding to the second original graph structure; The first original graph structure and the second original graph structure are original graph structures constructed based on different training samples.

6. A device for identifying abnormal brain function, characterized in that: include: an acquisition unit, configured to acquire target brain image data, wherein the target brain image data is rs-fMRI data to be identified; An input unit, configured to input the target brain image data into a brain function recognition model, wherein the brain function recognition model comprises an input layer, an image enhancement layer, an image feature processing layer, and a feature classification layer; Constructing an extraction unit, configured to construct, in the input layer, an original graph structure of a first brain functional connectivity network based on the target brain image data and extract a dynamic time feature matrix of the first brain functional connectivity network, wherein the original graph structure includes an original node feature matrix, an original edge weight matrix, and an original adjacency matrix; a graph enhancement unit, configured to perform graph enhancement processing on the original graph structure according to the dynamic time feature matrix in the graph enhancement layer to obtain an enhanced graph structure of the first brain functional connection network; A graph feature extraction unit, configured to perform feature processing on the enhanced graph structure in the graph feature processing layer to obtain a graph feature representation of the enhanced graph structure; a feature classification unit, configured to perform classification prediction of brain function abnormality based on the graph feature representation in the feature classification layer, and obtain a brain function abnormality recognition result of the target brain image data; The graph enhancement unit performs graph enhancement processing on the original graph structure according to the dynamic time feature matrix in the graph enhancement layer to obtain the enhanced graph structure of the first brain functional connection network, including: In the graph enhancement layer, the original adjacency matrix is optimized based on the original node feature matrix, the original edge weight matrix, and the dynamic time feature matrix to obtain an optimized adjacency matrix; Replacing the original adjacency matrix with the optimized adjacency matrix to obtain the enhanced graph structure; The construction and extraction unit, in the input layer, constructs the original graph structure of the first brain functional connection network based on the target brain image data and extracts the dynamic time feature matrix of the first brain functional connection network, including: In the input layer, blood oxygen level-dependent BOLD signal time series corresponding to multiple brain regions are extracted from the target brain image data; Performing feature extraction on the BOLD signal time series corresponding to the multiple brain regions to obtain the original node feature matrix; Performing global correlation analysis on the BOLD signal time series corresponding to the multiple brain regions to obtain the original edge weight matrix; performing a local correlation analysis based on a sliding time window on the BOLD signal time series corresponding to the multiple brain regions to obtain the dynamic time feature matrix; Initialize the original adjacency matrix to a fully connected matrix.

7. The device for identifying abnormal brain function according to claim 6, wherein: During the training of the brain function recognition model, the loss function used includes a first loss function constructed based on a first information bottleneck optimization objective, in which the first mutual information is maximized and the second mutual information is minimized; The first mutual information refers to the mutual information between the optimized adjacency matrix in the original graph structure of the second brain functional connection network and the node feature similarity matrix of the second brain functional connection network, and the node feature similarity matrix is obtained according to the edge weight matrix in the original graph structure of the second brain functional connection network and the node feature matrix in the original graph structure of the second brain functional connection network; The second mutual information refers to the mutual information between the optimized adjacency matrix and the dynamic time feature matrix in the second brain functional connection network; The second brain functional connection network is a brain functional connection network constructed based on training samples.

8. The device for identifying abnormal brain function according to claim 7, wherein: The first loss function includes a cross entropy loss term and a relative entropy loss term; The cross entropy loss term is a variational lower bound formula constructed based on the first mutual information, and the relative entropy loss term is a variational upper bound formula constructed based on the second mutual information; In the process of solving the first loss function, the optimized adjacency matrix is adjusted using a reparameterization technique.

9. An electronic device, characterized in that: include: memory and at least one processor; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method for identifying abnormal brain function according to any one of claims 1 to 5.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying abnormal brain function as claimed in any one of claims 1 to 5 are implemented.

Images