Function connection network analysis method based on group sparse attention converter

Abstract

The invention discloses a functional connection network analysis method based on a group sparse attention converter. The method mainly comprises the following steps: preprocessing brain functional magnetic resonance imaging data output by a functional magnetic resonance instrument; dividing the preprocessed brain function image into a plurality of brain regions according to a standard; based on Pearson correlation and group sparse characteristics, calculating a dependency relationship between brain regions, and respectively constructing an initialized brain network and a guidance network; taking the initialized brain network as the input of a group sparse attention converter, and carrying out learning under the assistance of a guidance network so as to learn a brain network with a group similar topological structure; and subsequent classification is carried out. The method integrates priori knowledge of a population level, allows more comprehensive modeling across multiple brain networks, is designed for end-to-end adaptive learning, dynamically updates the brain networks through back propagation, and significantly improves the construction and analysis performance of the brain networks.

Description

Technical Field

[0001] The present invention relates to medical care information processing, in particular to a method for functional connection network analysis based on a group sparse attention converter, which diagnoses brain diseases of subjects through functional connection networks. Background Art

[0002] Resting-state functional magnetic resonance imaging is a non-invasive neuroimaging technology that measures blood oxygen level-dependent signals without the subject having to perform any specific tasks. It has been widely used in the diagnosis of brain diseases such as major depression, autism spectrum disorder, and Alzheimer's disease. Brain diseases may be caused by genetic, environmental, or psychological factors, resulting in disordered neural circuits or interrupted communication between brain regions.

[0003] In recent years, scientists have gained more knowledge about the brain's functional connectivity network and have conducted a series of studies on it. The brain's functional connectivity network represents the dependency between different brain regions. Through the study of the functional connectivity network, we can further explore the potential relationship between brain regions and provide strong support for the analysis of brain diseases.

[0004] The paper "Constructing hierarchical attentive functional brain networks for early AD diagnosis" (J. Zhang, Y. Guo, L. Zhou, L. Wang, W. Wu, and D. Shen,'Constructing hierarchical attentive functional brain networks for early AD diagnosis', Medical Image Analysis., vol. 94, p. 103137, May 2024, doi:10.1016 / j.media.2024.103137.) mentioned that the traditional research on brain functional connection networks is mainly divided into the following three steps: (1) constructing brain functional connection networks, (2) feature selection, and (3) classification. First, the dependency relationship between predefined brain regions is estimated to construct the brain functional connection network, and then the features of the brain functional connection network are selected based on prior knowledge. Finally, the selected features are placed in the classifier for classification. However, these three steps are independent, and the entire process cannot adjust the construction of the early brain functional connection network through feedback from the classification stage, which will lead to suboptimal results. Moreover, due to the complexity and high dimensionality of the brain, it is difficult to manually determine the importance of features. In addition, once the brain functional connection network is constructed, it remains unchanged and cannot be automatically adjusted according to the learning process. This static nature cannot simulate the real brain activity well.

[0005] In addition, deep learning has introduced a unified framework for the study of brain functional connection networks to integrate steps such as feature selection and classification. For example, graph neural networks are mentioned in "Graph neural networks: A review of methods and applications" (J. Zhou et al., 'Graph neural networks: A review of methods and applications', AI Open, vol. 1, pp. 57–81, Jan. 2020, doi: 10.1016 / j.aiopen.2021.01.001.). Its disadvantages are that it operates in a fixed "information passing" framework, so the topological structure of the graph remains static during the learning process, and graph neural networks tend to capture local information and can only aggregate data from first-order neighbors.

[0006] The paper "Brain network transformer" (X. Kan, W. Dai, H. Cui, Z. Zhang, Y. Guo, and C. Yang, 'Brain network transformer', Oct. 15, 2022, doi: 10.48550 / arXiv.2210.06681.) introduces a deep learning model with an attention mechanism in the analysis of brain functional connectivity networks, which can better capture global information. Using the self-attention mechanism to obtain the global relationship between nodes can effectively model the complex relationship between brain regions. The disadvantage is that this method only focuses on the brain network of individuals and cannot analyze shared patterns or topological structures at the population level. Summary of the invention

[0007] The purpose of the present invention is to provide a method for functional connection network analysis based on group sparse attention converter. This method adds group sparse constraints and introduces the maximum mean difference method to ensure that the learned brain network is aligned with the guidance network, and applies group sparse constraints to capture the similarity relationship between subjects. Finally, brain disease diagnosis is performed in the fully connected layer, thereby significantly improving the construction and analysis performance of the brain network.

[0008] In order to solve the above problems, the present invention adopts the following technical solutions:

[0009] A method for functional connectivity network analysis based on group sparse attention transformers comprises the following steps: (1) preprocessing brain functional magnetic resonance imaging data acquired using an functional magnetic resonance device using a standardized data processing pipeline based on the rs-fMRI data processing assistant (DPARSF). Then, a 6-degree-of-freedom linear transformation without resampling is used to co-register each T1-weighted image to an average functional image, and spatial normalization is performed from a single native space to MNI space. Finally, brain regions are divided based on a standardized brain partition template to obtain an average time series.

[0010] MRI is a multi-parameter imaging. For the convenience of image analysis, it is hoped that the grayscale of a frame of MRI image is mainly determined by a specific imaging parameter. This is the so-called weighted imaging (WI). For example, when the image grayscale is mainly determined by T1, it is a T1-weighted image; when it is mainly determined by T2, it is a T2-weighted image; and when it is mainly determined by proton density, it is a proton density-weighted image.

[0011] (2) Extract the average time series of the nth subject , in represents the time series of the i-th brain region of the n-th subject, T represents the length of time in the time series, and M represents the number of nodes. The dependency relationship between the two brain regions is calculated based on the Pearson correlation, that is, the initialization of the brain network .

[0012] (3) Since the subjects have the same topological structure, based on the following formula: (1) Where N represents the number of subjects. yes norm, represents the time series matrix of the nth subject excluding the i-th brain region, represents the influence of other brain regions on the i-th brain region in the n-th subject, represents the influence of other brain regions of N subjects except the i-th brain region on the i-th brain region. According to formula (1), the guidance network with consistent topological structure among subjects is obtained: .

[0013] (4) Initialize the brain network As the input of the multi-head self-attention layer, through the linear projection matrix , ,and Get the query separately , key , and value : (2) Calculate along dimension D and The dot product of ,pass Scale the result to get the scaled result; apply the softmax function to the scaled result and then add it to the value Perform inner product to get the final attention output, as shown in formula (3): (3) The resulting matrix Defined as the original brain network.

[0014] use Norm regularization term, used for the original brain network To ensure group sparsity constraints at the subject level, the original brain network Reshape to size A matrix where each column corresponds to a subject: (4)

[0015] (6) Using the maximum mean difference method, the original brain network learned Network with mentors Perform an alignment operation to encourage the brain network to capture similar relationships across subjects during training. The specific loss function is defined as: (5) The maximum mean difference method is used to reduce the differences between the subjects’ brain networks, ensuring that the learned brain network does not deviate from the original brain network. , solve the degradation problem in deep neural networks and maintain the stability of performance, thereby obtaining the output of the multi-head attention layer.

[0016] (7) Based on the multi-head attention layer, a model of a group sparse attention-based converter is constructed. The model contains n encoding layers, each of which contains two sublayers: a multi-head attention layer and a feedforward network layer. Each sublayer uses a residual connection to maintain the original function. The features output by the encoding layer are embedded from the node level to the graph level through the readout function, and finally the graph level embedding is input into the fully connected layer for brain disease classification. Cross entropy loss is used for classification. Update and optimize the classification. Thus, the optimal model is obtained, and its objective function is as follows: (6) The hyperparameters and Used to balance the weights of the two losses. Update the model through back propagation, set the number of training updates epoch, and output the final classification result with the indicator of the optimal number of updates.

[0017] As a further improvement of the present invention:

[0018] In step (2), the dependency relationship between the two brain regions is calculated based on the Pearson correlation, that is, the brain network is initialized. By selecting a sparse representation mathematical model to describe the effective connectivity between brain regions, and by minimizing non-important connections, the dependency between two brain regions is estimated. It is more suitable for extracting key information from high-dimensional, noisy data and generating a simpler brain network.

[0019] The multi-head attention layer contains H attention heads, and the input maps out H results respectively, which are input into the corresponding attention heads respectively. After calculation, the H results are concatenated or added to generate the final output result. Using the multi-head attention layer can obtain learning from multiple perspectives, making the learned brain network more interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1It is an implementation flow chart of the functional connection network analysis model based on the group sparse attention converter of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation examples.

[0022] like Figure 1 As shown, the specific implementation example of the functional connection network analysis model based on the group sparse attention converter includes the following steps:

[0023] (1) In the present invention, a public Autism Brain Imaging Data Exchange (ABIDE) dataset was used, which contains 184 subjects, including 79 autistic patients and 105 healthy controls. The resting-state functional magnetic resonance data were collected using a 3T Allegra scanner and preprocessed using a standardized data processing flow based on the rs-fMRI data processing assistant (DPARSF). The data of the first 10 time points were discarded to obtain a stable magnetization state, and then slice timing correction, head motion correction, and bandpass filtering (0.01-0.1 ), nuisance covariates (i.e., head motion parameters, white matter, and cerebrospinal fluid) were regressed out. The individual T1-weighted images were then co-registered to the mean functional image using a 6-degree-of-freedom linear transformation without resampling and spatially normalized from a single native space to MNI space.

[0024] A standardized brain partition template is selected for the processed data for brain region division. In the present invention, the brain is divided into 116 brain regions using the automated anatomy labeling (AAL) atlas, including 90 cerebral regions and 16 cerebellar regions. The 116 brain regions represent 116 nodes of the brain functional connection network.

[0025] (2) The average time series of the nth subject is recorded as , in represents the time series of the i-th brain region of the n-th subject, T represents the length of time in the time series, and M represents the number of nodes. The dependency relationship between the two brain regions is calculated based on the Pearson correlation, that is, the initialization of the brain network .

[0026] (3) Since the subjects have the same topological structure, based on the following formula: (1) Where N represents the number of subjects. yes norm, represents the time series matrix of the nth subject excluding the i-th brain region, represents the influence of other brain regions on the i-th brain region in the n-th subject, It represents the influence of other brain regions of N subjects on brain region i except for the i-th brain region. This norm encourages subjects to group the weight matrix and obtain a guidance network with consistent topological structure among subjects according to formula (1). .

[0027] (4) Initialize the brain network As the input of the multi-head self-attention layer, through the linear projection matrix , ,and Get the query separately , key , and value : (2) in , and It is a learnable parameter. In this invention, in order to simplify the calculation, we set D=M=116 to keep the feature dimension and the number of brain regions consistent.

[0028] Calculate along dimension D and The dot product of In order to maintain the stability of data and training effect, Scale the result to get the scaled result; apply the softmax function to the scaled result and then add it to the value Perform inner product to get the final attention output, as shown in formula (3): (3) The resulting matrix It can also be rewritten as ,in represents the attention coefficient between the i-th and j-th brain regions in the attention matrix of the n-th subject. To measure the dependency between the i-th and j-th brain regions, it directly reflects their functional connectivity, so we transform the matrix Defined as the original brain network.

[0029] (5) Considering the similar topological structures between subjects, a group sparsity constraint is introduced to guide the refinement of the attention matrix, ensuring that these structured matrices capture consistent structures across subjects. Norm regularization term, used for the original brain network To ensure group sparsity constraints at the subject level, we put Reshape to size A matrix where each column corresponds to a subject and the following constraints are imposed on it: (4)

[0030] (6) In order to deal with the variability between subjects and enhance the robustness of the method, the maximum mean difference method was used to transform the learned original brain network Network with mentors Alignment is performed to encourage the brain network to capture shared topological structures across subjects during training. This alignment method acts as a regularization mechanism, and the specific loss function is defined as: (5) The maximum mean difference method is used to reduce the differences between the subjects’ brain networks, ensuring that the learned brain network does not deviate from the original brain network. , which also solves the degradation problem in deep networks and maintains the stability of performance, thereby obtaining the output of the multi-head attention layer.

[0031] (7) Based on the multi-head attention layer in step (6), a model of a group sparse attention-based converter is constructed. The model contains 2 encoding layers, each of which contains two sublayers: a multi-head attention layer and a feedforward network layer. Each sublayer uses a residual connection to maintain the original function. The features output by the encoding layer are embedded from the node level to the graph level through the readout function, and finally the graph level embedding is input into the fully connected layer for brain disease classification. Cross entropy loss is used for classification. Update and optimize the classification to obtain the objective function of the model: (6) Among them, the hyperparameters and Used to balance the weights of the two losses. Set the number of training updates to epoch=200, use back propagation to update the parameters, and the optimization algorithm used here is ADAM Optimizer.

[0032] In step (2), the dependency relationship between the two brain regions is calculated based on the Pearson correlation, which is to initialize the brain network. By selecting a sparse representation mathematical model to describe the effective connectivity between brain regions, and by minimizing non-important connections, the dependency between two brain regions is estimated. It is more suitable for extracting key information from high-dimensional, noisy data and generating a simpler brain network.

[0033] The multi-head attention layer contains H attention heads, and the input maps out H results respectively, which are input into the corresponding attention heads respectively. After calculation, the H results are concatenated or added to generate the final output result. Using the multi-head attention layer can obtain learning from multiple perspectives, making the learned brain network more interpretable.

[0034] In order to demonstrate the performance of the proposed method, the present invention uses eight common indicators, namely accuracy (ACC), sensitivity (SEN), specificity (SPE), balance accuracy (BAC), positive predictive value (PPV), negative predictive value (NPV), F1 score (F1_score), and area under the receiver operating characteristic (ROC) curve (AUC). The definitions of the above indicators are: , , , , , , , where TP, TN, FP and FN represent true positive, true negative, false positive and false positive, respectively.

[0035] Table 1: Comparison of the results of nine methods on the identification of patients with autism spectrum disorders under eight indicators Method ACC SEN SPE BAC PPV NPV F1_score AUC PC+SVM 0.6141 0.5823 0.6381 0.6102 0.5476 0.6700 0.5644 0.6890 SR+SVM 0.6467 0.5190 0.7429 0.6309 0.6029 0.6724 0.5578 0.7015 GSR+SVM 0.6630 0.5570 0.7429 0.6499 0.6197 0.6903 0.5867 0.7350 GCN 0.6251 0.5696 0.6691 0.6194 0.5835 0.6680 0.5703 0.6566 GAT 0.6591 0.6178 0.6918 0.6548 0.6176 0.7117 0.6045 0.7024 Transformer 0.6942 0.5554 0.7964 0.6759 0.6903 0.7087 0.6027 0.7346 BNT 0.6474 0.4497 0.7706 0.6101 - - 0.5665 0.7181 MDRL 0.5857 0.4846 0.6583 0.5715 0.5224 0.6332 0.4898 0.5849 Ours 0.7318 0.6054 0.8264 0.7159 0.7635 0.7476 0.6468 0.7878 .

[0036] As shown in Table 1, the classification index performances are based on three traditional methods (PC, SR, GSR), two popular deep methods (GCN, GAT), three improved new methods (Transformer, Brain Network Transformer, MDRL) and the present invention.

[0037] Among them, the present invention has achieved the best performance in all indicators: ACC=0.7318, SEN=0.6054, SPE=0.8264, BAC=0.7159, PPV=0.7635, NPV=0.7476, F1_score=0.6468, AUC=0.7878. On the classification accuracy ACC for autistic patients, the present invention has improved by 0.118 over the traditional PC method, improved by 0.073 over the GAT method, and improved by 0.038 over the Transformer method. In terms of AUC, the closer the numerical value is to 1, the more authentic the experiment is. As shown in the table, the AUC value of the present invention is 0.7878, which is 0.099 higher than the traditional PC method, improved by 0.085 over the GAT method, and improved by 0.053 over the Transformer method. By comparison, it is found that the results obtained by the present invention in the above eight indicators are all higher than other eight comparison methods.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement and partial supplement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for functional connection network analysis based on group sparse attention transformer, characterized in that The following steps are involved: (1) The brain functional magnetic resonance imaging data collected by the functional magnetic resonance imaging device were preprocessed using a standardized data processing process based on the rs-fMRI data processing assistant. Then, each T1-weighted image was co-registered to the average functional image using a 6-degree-of-freedom linear transformation without resampling, and spatial normalization was performed from a single native space to MNI space. Finally, brain regions were divided based on a standardized brain partition template to obtain an average time series. (2) The average time series of the nth subject is , in represents the time series of the i-th brain region of the n-th subject, T represents the length of time in the time series, M represents the number of nodes, and the dependency relationship between the two brain regions is calculated based on the Pearson correlation, that is, the initialization of the brain network ; (3) Since the subjects have the same topological structure, based on the following formula: （1） Where N represents the number of subjects. yes norm, represents the time series matrix of the nth subject excluding the i-th brain region, represents the influence of other brain regions on the i-th brain region in the n-th subject, represents the influence of other brain regions of N subjects except the i-th brain region on the i-th brain region; according to formula (1), the guidance network with consistent topological structure among subjects is obtained ; (4) Initialize the brain network As the input of the multi-head self-attention layer, through the linear projection matrix , ,and Get the query separately , key , and value : （2） Calculate along dimension D and The dot product of ,pass Scale the result to get the scaled result; apply the softmax function to the scaled result and then add it to the value Perform inner product to get the final attention output, as shown in formula (3): （3） The resulting matrix Defined as primitive brain networks; (5) Use Norm regularization term, used for the original brain network To ensure group sparsity constraints at the subject level, the original brain network Reshape to size A matrix where each column corresponds to a subject, with the following group sparse constraints: （4） (6) Using the maximum mean difference method, the original brain network learned Network with mentors Perform alignment operations to encourage the brain network to capture similar relationships between subjects during training. The specific loss function is defined as: （5） The maximum mean difference method is used to reduce the differences between the subjects’ brain networks, ensuring that the learned brain network does not deviate from the original brain network. , solve the degradation problem in deep neural networks and maintain the stability of performance, thereby obtaining the output of the multi-head attention layer; (7) Based on the multi-head attention layer in step (6), a model of a group sparse attention-based converter is constructed. The model contains n encoding layers, each of which contains two sublayers: a multi-head attention layer and a feedforward network layer. Each sublayer uses a residual connection to maintain the original function. The features output by the encoding layer are embedded from the node level to the graph level through a readout function. Finally, the graph level embedding is input into the fully connected layer for brain disease classification. The cross entropy loss is used for classification. Update and optimize the classification to obtain the optimal model. Its objective function is as follows: （6） The hyperparameters and It is used to balance the weights of the two losses, update the model through back propagation, set the number of training updates epoch, and output the final classification result with the indicator of the optimal number of updates.

2. The method for functional connection network analysis based on group sparse attention converter according to claim 1, characterized in that The following steps are involved: In step (2), the dependency relationship between the two brain regions is calculated based on the Pearson correlation, that is, the brain network is initialized. By selecting a sparse representation mathematical model to describe the effective connectivity between brain regions, and by minimizing non-important connections, the dependency between two brain regions is estimated. It is more suitable for extracting key information from high-dimensional, noisy data and generating a simpler brain network.

3. The method for functional connection network analysis based on group sparse attention converter according to claim 1, characterized in that The following steps are involved: The multi-head attention layer contains H attention heads, and the input maps out H results respectively, which are input into the corresponding attention heads respectively. After calculation, the H results are spliced ​​or added to generate the final output result. The use of the multi-head attention layer can obtain learning from multiple perspectives, making the learned brain network more interpretable.

Images