A two-stream spatio-temporal brain network analysis method embedded with group prior

Abstract

The present invention relates to the technical field of brain network construction, and in particular to a dual-stream spatiotemporal brain network analysis method embedded with group priors, comprising: S1, preprocessing brain functional images; S2, dividing the brain into several brain regions and extracting average time series; S3, calculating connection weights to construct edges of the brain map and outputting a symmetrical correlation matrix; S4, defining a graph isomorphism network under spatial features, using the correlation matrix R as the graph isomorphism network input, and outputting a spatial feature label Z1; S5, collecting BOLD signals of brain functional images and inputting them into a model to obtain a temporal feature label Z2; S6, respectively converting the spatial feature label Z1 and the temporal feature label Z2 into the spatial feature label Z1 and the temporal feature label Z2. 2 Dimensionality reduction and aggregation are performed to obtain the same-dimensional label Z; S7, using the same-dimensional label Z to establish a group-based attraction graph to achieve classification and recognition. The present invention uses spatiotemporal feature labels to construct a group graph G p The new labels embedded with group priors are obtained in conjunction with node feature updates, which can effectively improve the classification and recognition accuracy of brain functional images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain network construction, and particularly relates to a double-flow spatio-temporal brain network analysis method embedded with group priori. BACKGROUND

[0002] Resting-state functional magnetic resonance imaging (rs-fMRI) is a non-invasive neuroimaging technique used to measure the functional connectivity of the brain when no specific task is performed. It obtains the functional image of brain activity by detecting the fluctuation of blood oxygen level dependent (BOLD) signal, which is used to infer the functional interaction between different regions of the brain, and has been widely used in early studies of autism. Since autism is found to have means of damaging the neural connections between brain regions, constructing high-quality brain networks has become a key means to assist doctors in clinical research of autism.

[0003] The region of interest (ROI) of the brain obtained after preprocessing of resting-state magnetic resonance imaging is represented by the blood oxygen level dependent (BOLD) signal, and the BOLD signal is composed of a time series. Mamba is a new sequence modeling architecture based on state space model (SSM), which performs excellently in long sequence modeling, can efficiently process large-scale data, and at the same time solves the problems of traditional recurrent neural networks (RNNs) that are difficult to capture long-term dependencies and the quadratic computational cost caused by the attention mechanism of Transformer, providing a more efficient sequence modeling method with linear complexity. Graph isomorphism network (GIN) is a kind of graph neural network, mainly used to improve the graph representation ability, and the core idea is to imitate the Weisfeiler-Lehman graph isomorphism test, so as to improve the ability of GNN to distinguish different graph structures. GIN can capture spatial features more effectively than traditional GNN through means such as MLP transformation and neighborhood feature aggregation.

[0004] Most existing brain function disease diagnosis models only focus on a single spatial feature in the BOLD signal, making it difficult to capture deep feature information that changes dynamically over time, and the processing efficiency is low when facing large-scale samples, which limits its application and promotion in actual clinical practice. In addition, most of the existing brain function network construction methods are independent of individual subjects, ignoring the potential common structure and functional connection mode between subjects, which is easy to cause the diagnosis model to overfit to noise or individual differences, and has poor generalization ability. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a double-flow space-time brain network analysis method embedded with group priori, which obtains labels containing space-time characteristics by establishing a space-time double-flow model, constructs a group graph G p The node feature is updated by using the group attraction graph to obtain new labels embedded with group priori, so that the classification and recognition accuracy of the brain function image can be effectively improved.

[0006] The present application is realized by the following technical solutions:

[0007] The present application provides a double-flow space-time brain network analysis method embedded with group priori, comprising the following steps:

[0008] S1, pre-processing the collected brain function image;

[0009] S2, dividing the brain into a plurality of brain regions based on an automatic anatomical marker atlas, and extracting the average time sequence of each brain region and representing it as , wherein represents the time sequence of the i-th brain region, represents the length of the time sequence, represents the number of brain regions;

[0010] S3, calculating the connection weight by the Pearson correlation coefficient to construct the edge of the brain graph as follows:

[0011] For brain regions and , the Pearson correlation coefficient is:

[0012] ;

[0013] wherein: is obtained by z-score standardization of the time sequence of the brain region , so that the mean value is 0 and the standard deviation is 1:

[0014] ;

[0015] wherein: is the signal value of the brain region at the time point ; and are the mean value and the standard deviation of the brain region , respectively; all brain region pairs are traversed, and correlation coefficients are calculated, and a symmetric correlation matrix is outputted;

[0016] ​S4, define the graph isomorphism network under the space feature, take the correlation matrix R and the original data X output in step S3 as the input of the graph isomorphism network, and output the label Z1 learned with the space feature;

[0017] S5, collect the BOLD signal of the preprocessed brain function image as the input Give the model, and obtain the label Z2 learned with the time feature;

[0018] S6, respectively, the label Z1 learned with the space feature obtained in step S4 and the label Z2 learned with the time feature obtained in step S5 are reduced and aggregated to obtain a label Z fused with the space feature and the time feature in the original data;

[0019] S7, establish a group graph , the fused label Z is input to the group graph , and the new label Z_cat embedded with the group prior is obtained by updating the feature of the node Z by using the group attraction graph, so as to realize the classification and recognition of the brain function image.

[0020] Further, in step S5, the , and are obtained by selecting a mapping mechanism, cooperating with the state space model SSM formula, using the parameters , and are discretized by zero-order holding, and after discretization, the and the input are multiplied, and after discretization, the and the original state are multiplied, and the two items are added to obtain a new state , finally, the new state and are multiplied to obtain the output :

[0021] ;

[0022] ;

[0023] wherein: , , is a unit matrix, and the discretized SSM is equivalent to the following convolution:

[0024] ;

[0025] .

[0026] Further, in step S5 and step S6, a regularizer is defined when calculating the loss:

[0027] ;

[0028] In the training process, the regularization term is added to the total loss function:

[0029] ;

[0030] ;

[0031] Wherein: is the standard cross-entropy loss, is the number of categories; is the real label, represented by one-hot encoding, only the target category corresponds to 1, and the other positions are 0; is the label probability predicted by the model, that is, the output value of ; is a learnable parameter to control the influence of the regularization term; is the regularization loss defined by the regularizer, which should be applied in vector form as , wherein is the heat kernel weight.

[0032] The standard cross-entropy loss is introduced with a semi-supervised regularizer with a heat kernel weight, and the optimization loss function is optimized.

[0033] Further, in step S7, the label Z obtained by the space-time double-flow model is embedded into the patient node, and each patient in the healthy group is regarded as a positive sample, so as to gather and close in the feature space, while each patient in the patient group is regarded as a negative sample, so as to disperse and push away in the feature space; each patient in the patient group is regarded as a positive sample, so as to gather and close in the feature space, while each patient in the healthy group is regarded as a negative sample, so as to disperse and push away in the feature space, and finally a group graph G composed of two groups of clustered clusters is obtained p .

[0034] Further, in step S7, the edges of the group graph are constructed as Top-k neighborhood method, and the Top-k edges are connected in both directions by selecting the k neighbor nodes with the highest similarity metric of the input node. The similarity metric can be calculated in the following three ways:

[0035] (1) Dot product similarity: ;

[0036] The meaning is to measure the projection similarity of two vectors, that is, how many components they have in the same direction;

[0037] (2) Cosine similarity: ;

[0038] The meaning is to measure the similarity of the included angle between two vectors, regardless of the length of the module;

[0039] (3) Gauss kernel similarity: ;

[0040] The meaning is to construct a thermal kernel based on the Euclidean distance, and the closer the distance, the closer the similarity to 1.

[0041] Further, in step S4, the graph isomorphism network is:

[0042] ;

[0043] ;

[0044] Wherein: The brain network is constructed using Pearson correlation, is the feature vector of the node of the layer , and the initialization is a full connection layer; is a learnable parameter or a fixed scalar setting; is a group of nodes adjacent to the node , and is the neighbor node information of the aggregation node .

[0045] Further, in step S1, the DPARSF toolbox of matlab is used for preprocessing of brain function image, and the first time point of the magnetic resonance image is removed, the head motion artifact and time layer correction of the image are carried out, the influence generated by the ventricle, white matter signal and head motion high-order effect is removed, and the corrected image is registered to the standard space; the subject image is filtered by using time band-pass filter to reduce the influence of heartbeat and respiration.

[0046] The beneficial effects of the present application are:

[0047] The present application introduces two encoders Mamba and GIN, which fuse the ability of the former to efficiently capture time features in long sequence modeling and the ability of the latter to extract spatial correlation features on graph structure, greatly improve the calculation efficiency while ensuring the modeling performance. Through the spatiotemporal double-flow model cooperating with different feature encoders, the time and space features are captured in parallel, a label fused with space-time features is obtained, and a group graph composed of two groups of aggregation clusters is constructed by using the obtained label, so that the classification and recognition of disease images and healthy images in brain function images can be realized, and the accuracy of classification and recognition can be effectively improved by continuous training and replacement.

[0048] The application also innovatively introduces group prior information, establishes a connection between different subject samples based on similarity measurement by constructing a population graph, realizes cross-subject information fusion and structure induction, and thus embeds a group-level structure prior in the model training process. This design can effectively improve the stability and robustness of the model, especially when the sample number is limited or the data quality is uneven, significantly enhancing the generalization ability and practical application potential of the model under multi-site and multi-modal data.

[0049] By introducing the Mamba encoder, the two advantages of GIN, which can effectively capture spatial features, and Mamba, which is suitable for long sequence modeling and can efficiently process large-scale data, are combined. The group prior is embedded, which is different from the existing brain function networks that are mostly constructed independently for a single subject. The correlation between subjects is established according to the similarity, and the recognition and classification are more effective and accurate.

[0050] The double-flow spatio-temporal brain network analysis method proposed in the application can use the dot product similarity between subject labels as a similarity measure suitable for group-based classification. Compared with existing methods that do not recognize the information of adjacent subjects in the subject group and predict the label of an individual subject based on brain function images, the proposed method can achieve more accurate classification and recognition, effectively improving the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flowchart of the application.

[0052] Figure 2 The population graph in the application.

[0053] Figure 3 The model evaluation index table in the application on the ABIDE dataset.

[0054] Figure 4 The model evaluation index table in the application on the MDD dataset. DETAILED DESCRIPTION

[0055] To clearly illustrate the technical features of the present application, the following specific embodiments will be described.

[0056] In this embodiment, a publicly available Autism Brain Imaging Data Exchange (ABIDE) dataset is used, which includes 184 subjects, including 79 autism patients and 105 normal people.

[0057] As shown in Figure 1 A double-flow spatio-temporal brain network analysis method embedding group prior information includes the following steps:

[0058] S1. The brain functional image acquired by the functional magnetic resonance imaging device was preprocessed using the DPARSF toolbox of Matlab, including format reading and conversion. In order to keep the time series stable, the first 10 time points of each subject's magnetic resonance image were removed; the image was then corrected for head motion artifacts and time layers; the effects of ventricular, white matter signals and head motion high-order effects were removed; the corrected image was aligned to the standard space; the subject image was used Temporal bandpass filtering to reduce the effects of heartbeat and respiration.

[0059] S2. Select a standardized brain partitioning template. In this example, the brain is divided into 116 brain regions based on the AAL template, including 90 cerebral regions and 16 cerebellar regions. The 116 brain regions represent 116 nodes of the brain functional connectivity network.

[0060] The time series average of all voxels in each brain region in the magnetic resonance image is calculated, and the time series corresponding to different standard partitions are extracted, and the extracted time series are expressed as ,in Indicates the The time series of nodes, Indicates the number of times in the time series, Indicates the number of nodes.

[0061] S3. The connection weights are calculated by Pearson correlation coefficient to construct the edges of the brain map as follows:

[0062] For brain areas and , and its Pearson correlation coefficient:

[0063] ;

[0064] in It is the brain area The time series is z-score standardized to make the mean 0 and the standard deviation 1:

[0065] ;

[0066] in For brain area At the time point The signal value of and brain regions The mean and standard deviation of . , a total of Secondary correlation coefficient, output symmetrical correlation matrix .

[0067] S4, define the graph isomorphism network under the space feature, take the correlation matrix R output in step S3 as the input of the graph isomorphism network, and output the label Z1 about the space feature:

[0068] The space feature learning module is as follows:

[0069] ;

[0070] ;

[0071] Wherein: The input BOLD signal is captured by GIN to capture the spatial features between brain regions, which facilitates the subsequent aggregation of brain regions with similar functions.

[0072] S5, collect the BOLD signal input of the preprocessed brain function image Give The model is obtained, and the label Z2 about the time feature is obtained;

[0073] The time feature learning module is as follows:

[0074] ;

[0075] ;

[0076] Wherein: , , Is the unit matrix.

[0077] S6, respectively, the space feature label Z1 obtained in step S4 and the time feature label Z2 obtained in step S5 are reduced and aggregated to obtain the same dimension label Z;

[0078] S7, a group-based attraction graph is established by using the same dimension label Z , and the group attraction graph is used to realize the classification and recognition of brain function images.

[0079] The edges of the group graph G_p are constructed by Top-k neighborhood method, and the Top-k edges are connected bidirectionally by selecting the k neighbor nodes with the highest similarity measure of the input node.

[0080] Wherein, the calculation method of similarity measure can be selected from one of the following three methods:

[0081] (1) dot product similarity: ;

[0082] Its meaning is to measure the projection similarity of two vectors, that is, how many components they have in the same direction;

[0083] (2) cosine similarity: ;

[0084] The meaning is to measure the similarity of the included angle between two vectors, regardless of the length of the module;

[0085] (3) Gaussian kernel similarity: ;

[0086] The meaning is to construct a thermal kernel based on the Euclidean distance, and the closer the distance, the closer the similarity to 1.

[0087] In this embodiment, the calculation method of similarity measurement is selected as dot product similarity.

[0088] In steps S4 and S5, the regularization term is added to the total loss function during training:

[0089] ;

[0090] ;

[0091] Wherein: is the cross-entropy loss, is the number of categories; is the real label, represented by one-hot encoding, only the target category corresponds to 1, and the other positions are 0; is the predicted probability of the model, that is, the output value of ; is a learnable parameter to control the influence of the regularization term; is the regularization loss defined by the regularizer, which should be applied in vector form , wherein is the thermal kernel weight.

[0092] The learning rate is set to 0.0001, because the numerical value of BOLD signal is relatively small, all order of magnitude. The batch size is set to 16, and the number of subjects is small. The appropriate batch size and small learning rate can ensure the stability of the training.

[0093] When training to the last epoch, stop training, and output the test accuracy of this round and the label, and sequentially embed the label of all patient nodes in the group graph. Each patient in the healthy group is regarded as a positive sample to make it gather and close in the feature space, while each patient in the patient group is regarded as a negative sample to make it disperse and far away in the feature space; each patient in the patient group is regarded as a positive sample to make it gather and close in the feature space, while each patient in the healthy group is regarded as a negative sample to make it disperse and far away in the feature space, and finally obtain a group graph .

[0094] To verify the effectiveness of the model proposed in the application, a two-flow spatio-temporal brain network analysis method embedded with population prior was tested and evaluated on two public neuroimaging classification datasets, ABIDE and MDD, respectively. The performance indicators of the model constructed by the method are shown in the following table. Figure 3 As shown in the table, on the ABIDE dataset, the model of the application achieved classification accuracies (ACC) of 70.64% and 75.00% on the NYU116 and Leuven subsets, respectively, AUCs of 70.17% and 75.42%, respectively, and balanced accuracies (BACs) of more than 69%, and stable F1-scores, indicating that the model has good recognition performance for autism samples. Figure 4

[0095] On the MDD dataset, the model showed good classification accuracy and generalization ability in multiple site data, such as achieving an AUC of 85.39% and a BAC of 78.99% on the SITE9 dataset, and an F1-score of 79.47%, indicating that the model has good discrimination ability for identifying depression samples.

[0096] ABIDE and MDD are two different datasets, and the model of the application performs well on both datasets, which shows that the model structure design has good universality and adaptability, and has significant practical value in actual clinical prediction tasks.

[0097] Of course, the above description is not limited to the above examples, and the technical features not described in the application can be implemented by or using existing technology, which will not be described here; the above examples and drawings are only used to illustrate the technical solutions of the application and are not limited to the application, and the ordinary skilled in the art should understand that the changes, modifications, additions or substitutions made by the ordinary skilled in the art within the essential scope of the application do not deviate from the purpose of the application, and should also belong to the protection scope of the claims of the application.​

Claims

1. A two-stream spatiotemporal brain network analysis method embedded with group priors, characterized by: The following steps are involved: S1. Preprocessing of collected brain functional images; S2. Based on the automatic anatomical labeling atlas, the brain is divided into several brain regions, and the average time series of each brain region is extracted and expressed as ,in Indicates the Time series of brain regions, represents the length of the time series, Indicates the number of brain regions; S3. The connection weights are calculated by Pearson correlation coefficient to construct the edges of the brain map as follows: For brain areas and , and its Pearson correlation coefficient: ； in: It is the brain area The time series is z-score standardized to make the mean 0 and the standard deviation 1: ； in: For brain area At the time point The signal value of and brain regions The mean and standard deviation of , a total of Secondary correlation coefficient, output symmetrical correlation matrix ; S4. Define a graph isomorphism network under spatial features. Use the correlation matrix R output in step S3 and the original data X as the input of the graph isomorphism network, and output the label Z1 that has learned the spatial features: S5. Collect BOLD signal input of pre-processed brain functional images Give Model, obtains the label Z2 that learns the time feature; S6, respectively perform dimensionality reduction aggregation on the label Z1 of the learned spatial features obtained in step S4 and the label Z2 of the learned temporal features obtained in step S5 to obtain a label Z that integrates the spatial features and temporal features in the original data; S7. Build a group diagram , input the fused label Z as a node into the group graph , the group attraction graph is used to update the features of node Z to obtain a new label Z_cat embedded with group priors, thereby realizing the classification and recognition of brain functional images.

2. The dual-stream spatiotemporal brain network analysis method embedded with population priors according to claim 1 is characterized by: In step S5, the selective mechanism mapping is used to obtain 、 and , with the state space model SSM formula, using parameters Through zero-order hold and After discretization, we get and input Multiply and discretize to get and original state Multiply and add these two terms to get the new state , and finally the new state and Multiply to get the output : ； ； in: , , is the identity matrix, and the discretized SSM is equivalent to the following convolution: ； 。 3. The dual-stream spatiotemporal brain network analysis method embedded with population priors according to claim 1 is characterized by: In steps S5 and S6, the regularization term is added to the total loss function during training to define a regularizer: ； During training, the regularization term is added to the overall loss function: ； ； in: is the standard cross entropy loss, is the number of categories; is the real label, represented by one-hot encoding, where only the position corresponding to the target category is 1 and the other positions are 0; is the label probability predicted by the model, i.e. The output value of It is a learnable parameter used to control the influence of the regularization term; is the regularized loss defined by the regularizer, which should be in vector form ,in It is thermonuclear rights.

4. The dual-stream spatiotemporal brain network analysis method embedded with population priors according to claim 1 is characterized by: In step S7, the label Z obtained according to the spatiotemporal dual-stream model is embedded into the patient node. Each patient in the healthy group is regarded as a positive sample, so that they are clustered together in the feature space. At the same time, each patient in the patient group is regarded as a negative sample, so that they are dispersed and pushed away in the feature space. Each patient in the patient group is regarded as a positive sample, so that they are clustered together in the feature space. At the same time, each patient in the healthy group is regarded as a negative sample, so that they are sparse and pushed away in the feature space. Finally, a group graph G consisting of two groups of clusters is obtained. p .

5. The dual-stream spatiotemporal brain network analysis method embedded with population priors according to claim 1 is characterized by: In step S7, the group graph The edge is constructed as the Top-k neighborhood method, which selects the k neighbor nodes with the highest similarity measure of the input node to connect the bidirectional Top-k edges.

6. The dual-stream spatiotemporal brain network analysis method embedded with population priors according to claim 1, characterized in that: In step S4, the graph isomorphism network is: ； ； in: is a brain network constructed using Pearson correlation. For the Layer Node Initialize the eigenvector of is a fully connected layer; For learnable parameters or set a fixed scalar; Is with the node A set of adjacent nodes, It is an aggregation node Neighbor node information.

7. The dual-stream spatiotemporal brain network analysis method embedded with population priors according to claim 1 is characterized by: In step S1, the DPARSF toolbox of Matlab was used to preprocess the brain functional map imaging to remove the front At each time point, the image is corrected for head motion artifacts and time layers, and the effects of ventricular and white matter signals and head motion high-order effects are removed. The corrected image is registered to the standard space. Temporal bandpass filtering to reduce the effects of heartbeat and respiration.

Images