A cognitive activity prediction method integrating network topology and physical constraints

Abstract

The present invention discloses a cognitive activity prediction method integrating network topology and physical constraints, comprising the following steps: (1) constructing a brain network adjacency matrix based on resting-state functional magnetic resonance imaging data and a functional connectivity method; (2) calculating the physical distance between brain regions according to a brain template to obtain a spatial constraint matrix; (3) constructing a functional asymmetric path connection considering the functional hierarchical differences between brain regions; (4) calculating the shortest path between brain regions based on the functional asymmetric path connection matrix and considering physical space constraints to predict cognitive brain activity; (5) executing the model by the leave-one-out method to obtain the predicted brain activity of each brain region, and finally calculating the similarity between the predicted and actual brain activations. This method not only considers the topological properties of the brain network, but also combines the spatial geometry and functional hierarchical characteristics of the brain. By verifying this method on a real data set, the experimental results show that the present invention significantly improves the accuracy of cognitive activity prediction.

Description

Technical Field

[0001] The present invention is applicable to the fields of functional magnetic resonance imaging and brain network analysis, and provides a cognitive activity prediction method integrating brain network topology and physical constraints. Background Art

[0002] The human brain is one of the most complex systems in the world. As a core system, it supports humans to adaptively complete various tasks. Studying brain structure and function, further developing the brain, and protecting the brain are of far-reaching significance, and have always been a hot topic for exploration by many fields and researchers. Among them, abstracting the brain into a sparse complex network to explore its functions is an important direction. The nodes in the network represent the brain regions, and the edges represent the connection strength between the functional brain regions. Significant progress has also been made in using the topological structure information of the brain network to analyze the working principles of the brain and further apply it in actual clinical practice.

[0003] At present, magnetic resonance imaging has been widely used as a non-invasive brain imaging technology, providing an effective tool for exploring brain function and structure. Magnetic resonance imaging technology has become an important tool for many researchers because it can explore whole-brain activity and has good temporal and spatial resolution. In order to explore the process of the human brain performing various tasks, researchers have designed various experimental paradigms to let subjects perform specific tasks to explore task-induced cognitive processes. Brain activity induced when performing specific tasks is generally called task-state brain activation, which is mainly used to characterize the response of the human brain when completing specific tasks. In addition, compared with the task state, the brain activity of the subjects when they are not performing any tasks is called resting-state brain activity. There are also a lot of studies on using resting-state magnetic resonance imaging data to construct brain networks and explore the organizational characteristics of brain function.

[0004] However, in most current studies, the resting state and the task state are usually explored separately, and the relationship between the two is not deeply explored. The activity flow model proposed in recent years considers the brain network constructed by resting state data as a basic path, combined with the task state, to predict the activation of specific brain areas, and has achieved good prediction results. It also provides a way to combine the resting state and the task state. However, in the prediction, its prediction only considers the functional connection of a pair of brain areas, fails to fully consider the functional hierarchy of the brain, and equates the functional influence between brain areas, that is, the constructed brain network is symmetrical, which is inconsistent with current neuroscience theory. In addition, the physical structural characteristics of the brain itself are not taken into account in the prediction. This physical space constraint may also become an important factor in the prediction model. Summary of the invention

[0005] The purpose of the present invention is to provide a cognitive activity prediction method that integrates network topology and physical constraints to address the deficiencies of the prior art. Considering the hierarchical organization of the brain, the functional connections between brain regions are reconstructed, and the asymmetric functional connections obtained represent the different mutual influences between brain regions. Then consider the physical constraint characteristics of the brain itself, that is, the physical distance between each brain region. Finally, the two characteristics are combined to complete the prediction of brain region activation.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A cognitive activity prediction method integrating network topology and physical constraints comprises the following steps:

[0008] Step 1: Preprocess the resting-state MRI and construct functional connections. For task-state MRI, calculate brain activation according to different task paradigms.

[0009] Step 2: For the constructed functional connections, use the Dijkstra algorithm to calculate the shortest distance between any two points and form a shortest distance connection matrix;

[0010] Step 3: Calculate the physical distance between each brain region through the corresponding brain partitioning template;

[0011] Step 4: Based on the distance connection matrix calculated in step 2, the different influences between brain regions are calculated using its own functional topological structure to form an asymmetric path connection matrix with direction;

[0012] Step 5: Predict cognitive activation of different tasks based on the network topology and physical constraints between brain regions.

[0013] To optimize the above technical solutions, the specific measures taken also include:

[0014] Furthermore, in step 1, the resting state magnetic resonance imaging is divided using a 264 brain region template, the time series of the corresponding brain regions are extracted, and the functional network is constructed based on the time series using the Pearson correlation coefficient. At this time, each sample will form a fully connected weighted network, and for a given task, the significant fluctuations of each brain region during task execution compared to the baseline level are calculated according to the task paradigm design and the general linear model (GLM), that is, for a given task, a 1 x 264 vector is formed.

[0015] Furthermore, in step 2, for the functional connections constructed in the resting state, the shortest distance between any brain regions is calculated according to the connection strength, and a distance matrix SP is formed.

[0016] Furthermore, in step 3, according to the brain region division template in step 1, the physical distance between any two brain regions is measured using the Euclidean distance, that is, the three-dimensional spatial coordinates of brain regions i and j in the template are and , then the physical distance between the two brain regions is:

[0017]

[0018] The physical distance matrix D of any pair of brain regions in the whole brain is calculated in sequence.

[0019] Furthermore, in step 4, according to the path connection matrix constructed in the second step, for the path connection constructed using each sample, an asymmetric network is constructed according to the topological properties of each brain region, and for the brain region i in the sample path connection SP, its path connection to the brain region j is calculated as follows:

[0020]

[0021] The first term on the right side of the above formula represents the scaling factor of the original path connection according to the topological properties of the brain region, and the second term represents the original path connection. Specifically, is the asymmetric path connection from brain region i to j, Represented as the original path connection from brain region i to brain region t, is the number of connections that brain region i has with all other brain regions in the connectivity matrix. Similarly, represents the asymmetric path connection from brain region j to brain region i, and Not the same.

[0022] Furthermore, in step 5, according to the physical distance calculated above and the asymmetric path connection, the activation of the brain area i of the sample on a certain task can be predicted by the following formula:

[0023]

[0024] in, represents the predicted activation value of brain region i, represents the activation of brain area j except brain area i, is the asymmetric path connection from brain region i to brain region j, It indicates the Euclidean distance between brain regions i and j in the standard physical space calculated in step 3.

[0025] The present invention is based on neuroscience theory, combines the multi-level theory of brain region function with physical characteristics, and fully explores the information transmission characteristics of the brain. On real data sets, the proposed prediction model has higher prediction ability than existing methods.

[0026] The present invention provides a cognitive activity prediction method that integrates network topology and physical constraints, which has the following effects: by constructing a functional network, the shortest path connection is calculated, and according to the different connection weight distributions of each brain region, the asymmetric path connections of different brain regions are reconstructed, and further combined with the physical space distribution of different brain regions, the network topology and physical constraints are combined to more accurately express the transmission of cognitive information and enhance the performance of the model in predicting different tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of the calculation of the example brain area network topology provided by the present invention;

[0028] Figure 2 A schematic diagram of calculating the physical space constraints between brain regions in the example provided by the present invention;

[0029] Figure 3 It is a curve diagram of the prediction accuracy of the present invention under different sparsity on the Main dataset;

[0030] Figure 4 It is a curve diagram of the prediction accuracy PredictionAccuracy of the present invention under different sparsities on the Replication dataset;

[0031] Figure 5 This is a curve diagram of the prediction accuracy PredictionAccuracy of the present invention under different sparsities on the Reduced dataset. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is now further described in detail in conjunction with the accompanying drawings and embodiments.

[0033] The present invention proposes a cognitive activity prediction method integrating network topology and physical constraints, which mainly includes the following steps:

[0034] Step 1: Preprocess the resting-state functional imaging data, construct functional connections, and calculate brain activation based on different task paradigms for the task-state functional imaging data;

[0035] Step 2: For the functional connection calculated in step 1, the shortest path is calculated by Dijkstra algorithm, and the shortest path connection is obtained;

[0036] Step 3: Calculate the physical distance between each brain region through the corresponding brain partition template;

[0037] Step 4: Based on the path connection calculated in step 2, the different influences between brain regions are calculated using its own functional topological structure to form an asymmetric path connection matrix with direction;

[0038] Step 5: Predict brain activation for different tasks based on network topology characteristics and physical constraints between brain regions.

[0039] In step 1, each sample is preprocessed by standardization to obtain the corresponding time series. For each sample, the functional connectivity constructed using Pearson is represented as a symmetric adjacency matrix ,in represents the brain network adjacency matrix of the sth person, M is the number of samples, represents the connection weight between brain regions i and j of the sth person. To further verify the robustness of the proposed method, the constructed functional brain network was sparsed using different thresholds, and functional networks with different densities were obtained for subsequent analysis and prediction.

[0040] Furthermore, according to the corresponding task paradigm, the activation of the corresponding brain area is calculated using the general linear model, which is a vector of the number of corresponding brain areas, ,in Corresponding to the activation of the i-th brain region in the s-th sample.

[0041] In step 2, according to the functional connectivity matrix, the inverse of the connection weight is converted into a path step length, and the Dijkstra algorithm is used to obtain the shortest path distance between any brain regions, and the inverse is further taken to obtain the shortest path connection. ,in represents the brain network adjacency matrix of the sth person, M is the number of samples, represents the connection weight between brain regions i and j of the sth person.

[0042] In step 3, for all samples, the physical distance between any brain regions is calculated using the following formula based on the standard brain template to obtain the physical distance matrix:

[0043]

[0044] Based on physical distance, the information transmission between brain regions is further constrained, see Figure 2 Calculations showing physical spatial constraints between brain regions.

[0045] In step 4, according to the individual path connection matrix calculated above, for the sample s path connection SP, for brain region i, its path connection to brain region j is as follows: Figure 1 The calculation of functional asymmetric path connectivity between brain regions is shown as follows:

[0046]

[0047] The first term on the right side of the above formula represents the scaling factor of the original path connection according to the topological properties of the brain region, and the second term represents the original path connection. Specifically, is the asymmetric path connection from brain region i to brain region j, Represented as the original path connection from brain region i to brain region t, is the number of connections that brain region i has with all other brain regions in the connection matrix. represents the asymmetric path connection from brain region j to brain region i, and Not the same.

[0048] According to the physical and functional characteristics proposed above, the individual brain network is better characterized, further improving the model's performance in predicting brain region activation when performing different tasks. Experiments were conducted on three sub-datasets of the real public human connectome (Main dataset, Replication dataset and Reduced dataset) to verify the effectiveness of the proposed method.

[0049] The technical solution of the present invention will be further described in detail below in conjunction with application examples:

[0050] In a specific example of the present invention, the effectiveness of the proposed method is evaluated on three public fMRI datasets. Table 1 gives the basic information of these data.

[0051] Table 1: Statistics of subjects in the dataset

[0052]

[0053] The fMRI data used in the experiment were obtained from the Human Connectome Project (HCP, https: / / db.humanconnectome.org / data / projects / HCP_1200) with minimal preprocessing. Subsequently, the fMRI data of the resting state and task state were further preprocessed using the SPM12 toolbox (http: / / www.fil.ion.ucl.ac.uk / spm) and REST (https: / / www.nitrc.org / projects / rest) software. For the resting state fMRI data, the first 10 time points of each run were discarded to account for the influence of participants' adaptation to the experimental environment. Linear regression was used to remove motion estimates and cerebrospinal fluid and white matter signals. Linear trend removal, bandpass filtering (0.01-0.08Hz), and spatial smoothing were then performed. The brain was divided into 264 brain regions using the Power264 template, and time series were extracted. The functional brain network was constructed using the Pearson correlation coefficient. For the task state fMRI data, linear regression was used to remove motion estimates, and then a 4mm Gaussian filter was used for spatial smoothing. According to different task paradigm designs, the activation values ​​of different brain regions were calculated using the general linear model.

[0054] The original resting-state brain network constructed is a fully connected symmetric network, which may contain false connections. In brain network research, multiple thresholds are generally set to sparse the network to obtain multiple sparse networks. In this experiment, weak connections are removed in a certain proportion, leaving stronger connections. And based on the sparse network, its functional asymmetric network is calculated. The physical distance between brain regions is calculated based on the brain region division template, and the Euclidean distance is used here. Finally, based on the activity flow model, the functional asymmetric network and physical distance are used to predict brain activation under different tasks, and the similarity with the real brain activation is calculated as the prediction accuracy.

[0055] Table 2, Table 3 and Table 4 respectively show the prediction results of the proposed cognitive activity prediction method integrating network topology and physical constraints (Combined) on three data sets, and compare it with the traditional activity flow prediction (Baseline), and also compare it with the proposed single feature combined method (network topology-FE, physical constraint-PE). Specifically, the method proposed in the present invention shows a significant improvement in prediction performance on different network densities of the three data sets, further confirming the effectiveness of the proposed method.

[0056] Table 2: Prediction performance of all methods on the Main dataset

[0057] Method MAC AUC Significance Baseline 0.51 0.51 - FE 0.57 0.57 p < 0.0001 PE 0.59 0.59 p < 0.0001 Combined 0.61 0.61 p < 0.0001

[0058] Table 3: Prediction performance of all methods on the Replication dataset

[0059] Method MAC AUC Significance Baseline 0.54 0.54 - FE 0.59 0.59 p < 0.0001 PE 0.60 0.61 p < 0.0001 Combined 0.63 0.63 p < 0.0001

[0060] Table 4: Prediction performance of all methods on the Reduced dataset

[0061] Method MAC AUC Significance Baseline 0.47 0.47 - FE 0.53 0.53 p < 0.0001 PE 0.57 0.58 p < 0.0001 Combined 0.59 0.59 p < 0.0001

[0062] MAC (Mean Accuracy) is the average prediction accuracy under all network densities, AUC (Area UnderCurve) is the area under the curve, and Significance is the significance of the prediction accuracy of the current method relative to the baseline method. All the above results demonstrate the effectiveness of the proposed brain activity prediction method that integrates network topology and physical constraints. In addition, it can be seen that when the physical and network topology characteristics are introduced separately, significant prediction performance improvements are also achieved, further proving that the two proposed characteristics have significant improvements and important roles in brain activity prediction. Figure 3 Represents the prediction on Maindataset, Figure 4 Represents the prediction on the Replication dataset, Figure 5 It represents the prediction on the Reduced dataset, and shows the changes in prediction accuracy under different network sparsity on the three datasets. It can be seen that the proposed method is better than the baseline under the same sparsity, which verifies the effectiveness and robustness of the proposed method.

[0063] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

[0064] It should be noted that the terms such as "upper", "lower", "left", "right", "front", "back", etc. cited in the invention are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments in their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

Claims

1. A cognitive activity prediction method integrating network topology and physical constraints, characterized in that: The following steps are involved: Step 1: Preprocess the resting state and task state functional magnetic resonance imaging data, construct the functional connection between brain regions, and calculate brain activation based on the task paradigm for the task state functional imaging data; Step 2: For the functional connection in step 1, the shortest path is calculated by Dijkstra algorithm, and the shortest path connection is obtained; Step 3: Calculate the physical distance between each brain region through the corresponding brain partition template; for each pair of brain regions i and j, calculate the Euclidean distance D of the node according to the spatial three-dimensional coordinates of the registration template ij ; Form the physical distance matrix D: In the formula, (x i ,y i ,z i ) and (x j ,y j ,z j ) correspond to the spatial coordinates of nodes i and j respectively; Step 4: Calculate the asymmetric path network based on the topological attributes of each node for the constructed shortest path network; for the shortest path network constructed using each sample, construct an asymmetric network based on the topological attributes of each node. For node i in the shortest path connection SP of sample s, its path connection to node j is calculated as follows: The first term on the right side of the above formula represents the scaling factor of the shortest path length according to the node topology attribute, and the second term represents the shortest path length; specifically, is the asymmetric path connection from node i to node j, Represented as the original path connection from node i to node t, C i is the number of connections that node i has with all other nodes in the connection matrix; similarly, represents the asymmetric path connection from node j to node i, and are not the same; Represents the connection weight between node i and node j of sample s; Step 5: Calculate the activation of each node in the task state according to the physical distance and asymmetric function of the node; specifically, according to the calculated physical distance and asymmetric path connection, the activation of brain area i of the sample on a certain task is predicted by the following formula: Among them, E i represents the predicted activation value of node i, A j Indicates that j nodes other than i nodes are activated, SC ij is the asymmetric path connection from node i to node j, D ij Represents the Euclidean distance between nodes i and j in the standard physical space.

2. The cognitive activity prediction method integrating network topology and physical constraints as claimed in claim 1, characterized in that: The step one is specifically as follows: the resting state and task state functional magnetic resonance imaging data are divided into 264 brain regions according to the template; the resting state data are processed into a time series of corresponding length, and the functional brain network is constructed using the Pearson correlation coefficient. The brain network constructed at this time is a symmetrical weighted network; the task state functional magnetic resonance imaging data are preprocessed and the brain region activation is estimated using the general linear model GLM.

3. The cognitive activity prediction method integrating network topology and physical constraints as claimed in claim 1, characterized in that: In the second step, for the created original functional connection, the Dijkstra algorithm is used to calculate the shortest distance between any two points to form a shortest distance connection matrix SP.

Images