Mental disease identification method based on pulse structure-functional brain network coupling

Abstract

The invention provides a mental disease identification method based on pulse structure-functional brain network coupling, belongs to the technical field of intelligent auxiliary medical diagnosis, and effectively solves the technical problem of neglected neurobiological mechanism between structural connection and functional connection in the traditional mental disease diagnosis process. According to the technical scheme, firstly, functional and structural brain networks are extracted from functional magnetic resonance imaging and diffusion tensor imaging; then, information feature maps of the two brain networks are extracted through a BrainNetCNN; then constructing a pulse coupling neural network to learn a brain structure-function coupling mechanism so as to obtain pulse structure-function coupling; and finally, inputting the obtained coupling information into a classification layer to obtain a disease recognition result, and training and optimizing the result by using a cross entropy loss function. The beneficial effects of the invention are that the method is helpful for deep understanding of neurological mechanisms of mental diseases, and has wide application prospects in clinical application.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent assisted medical diagnosis, and in particular to a method for identifying mental illness based on pulse structure-function brain network coupling. Background Art

[0002] Mental illness refers to a disease in which the brain is dysfunctional under the influence of various biological, psychological and social environmental factors, resulting in varying degrees of disorders in mental activities such as cognition, emotion, will and behavior. A person with a mental illness may have different brain structures and functions, or different neurochemical reactions. In addition, most studies have confirmed that patients with mental illness have enlarged ventricles and atrophied gray matter in their brains. It can be seen that the complexity of mental illness mainly focuses on changes in brain structure and function.

[0003] The brain is a complex network composed of the interaction of structural and functional connections. Structural connectivity mainly refers to the white matter fibers that connect gray matter areas in the brain, and functional connectivity is defined as the statistical dependencies between different brain regions. The relationship between structural and functional connectivity is called structure-function coupling, which plays an important role in the diagnosis of mental illness.

[0004] At present, traditional structure-function coupling methods mainly aim to find statistical associations between topological patterns of structural and functional connections, lacking principled system-level integration to characterize the coupling mechanisms of how neuronal populations communicate and produce significant brain functions on top of structural connection groups. Secondly, most existing structure-function coupling deep learning models do not take into account the natural and complex connections between structure and function, and ignore the neurobiological mechanisms between structural and functional connections.

[0005] The paper Learning tensor-based features for whole-brain fMRI classification proposed to use multilinear principal component analysis to study the structure-function coupling mechanism. The structural brain network matrix and the functional brain network matrix are combined into a three-dimensional tensor, and then the feature vector is extracted through multilinear principal component analysis. Finally, the support vector machine is used for classification to obtain the accuracy. This method only considers the statistical association between structural connection and functional connection, and cannot fully capture the complex interaction and dynamic changes between brain structure and functional connection.

[0006] The paper Learning to model the relationship between brain structural and functional connectomes uses a supervised graph representation learning framework based on graph convolutional networks to study the structure-function coupling mechanism. The encoder learns node-level structural connection embedding and generates a whole-brain graph representation to reconstruct the functional brain network. Finally, a multi-objective loss function is used to jointly reconstruct the functional brain network and learn the discriminant graph representation of the structure-function coupling mapping for graph-level classification. This method ignores the neurobiological mechanism between structural and functional connections and does not fully consider the biological characteristics of brain connections and the neural transmission process.

[0007] Therefore, using the pulse transmission mechanism in spiking neurons to simulate the action potential transmission in the brain can better study the brain structure-function coupling mechanism and significantly improve the accuracy of diagnosing mental illness. Summary of the invention

[0008] In view of the technical problem of the neurobiological mechanism between structural connection and functional connection which is often overlooked in the traditional diagnosis of mental illness, the purpose of the present invention is to provide a method for identifying mental illness based on pulse structure-function brain network coupling to improve the accuracy of mental illness identification.

[0009] The inventive concept of the present invention is that the structural-functional coupling of the brain plays an important role in the diagnosis of mental illness. It involves the relationship between the structural connection and the functional connection of the brain. In the diagnosis of mental illness, the analysis of structural-functional coupling can provide important information about how the brain network supports or deviates from normal function. At present, most structural-functional coupling analysis methods do not comprehensively consider the natural and complex connection between structural connection and functional connection, and ignore the neurobiological mechanism between structural connection and functional connection. Spiking neural network is a computational model inspired by biological nervous system. The pulse transmission mechanism in spiking neural network can simulate the interaction between structural connection and functional connection in the brain. Therefore, based on the spiking neural network, a pulse coupled neural network is proposed, which uses the dense interconnection of spiking neurons and synaptic weights to simulate the interaction between structural connection and functional connection in the brain to obtain the spiking structure-functional coupling, which provides a new perspective for the study of structure-functional coupling.

[0010] In order to better achieve the above-mentioned invention object, the present invention is implemented by the following technical solution: a method for identifying mental illness based on pulse structure-function brain network coupling, the specific steps are as follows:

[0011] A. Functional brain networks were constructed by functional magnetic resonance imaging (fMRI). The fMRI data were processed by spm12 in the DPARSF toolbox. The initial functional time series were slice time acquisition corrected, realigned and normalized to the EPI template. Then a detrending process was performed to remove the source of pseudo-variance, namely six head motion parameters, the average signal of CSF and white matter, and the global brain signal. The time series were band-pass filtered at 0.01-0.08 Hz. The resulting volume had 240 time points, which were divided into 90 regions of interest using the automatic anatomical labeling atlas AAL. Structural brain networks were constructed by diffusion tensor imaging DTI. DTI data were processed using the PANDA suite and DTI distortion correction was performed using the FSL toolbox. Then, fiber images were obtained using a deterministic tracking method using TrackVis. Anatomical regions were defined using the AAL convention based on the registered T1 images of each subject.

[0012] B. Use BrainNetCNN to obtain information feature maps of structural brain networks and functional brain networks. α=1,2,where A 1 Represents the adjacency matrix of the structural brain network, A 2 The adjacency matrix represents the functional brain network, N corresponds to the node set of the brain region. In order to obtain the complex features in the two brain network data, edge-to-edge convolution E2E and edge-to-node convolution E2N operations are used to obtain the feature maps of the two brain networks;

[0013] C. Use the Leaky-Integrate-and-Fire pulse neuron model to build a pulse coupled neural network. The structural feature map λ after E2E and E2N convolution operations sc and the functional characteristic graph λ fc Repeat T times as input and perform implicit pulse encoding through internal pulse neurons, where T is the time step, and then use pulse coupled neural network to learn the structure-function coupling mechanism from the structure and function feature graph to find the pulse structure-function coupling;

[0014] D. The obtained pulse structure-function coupling is input into the classification layer, the predicted category is determined by the pulse output sequence, and the cross entropy loss between the predicted category and the true category is used as the loss function for training, thereby obtaining a neural network model that can identify the subtle relationship between healthy and disease connection groups based on the pulse structure-function coupling.

[0015] As the mental illness identification method based on pulse structure-function brain network coupling provided by the present invention, the specific steps of step B are as follows:

[0016] a. Edge-to-edge E2E convolution, which aims to extract edge features from the original adjacency matrix. It uses the topological structure of the network to perform filtering operations by combining the weights of the edges of shared nodes. The formula is as follows:

[0017]

[0018] in, and is the learning weight, f n Represents the nonlinear activation function of the nth filter. The feature maps of the structural brain network and the functional brain network after E2E convolution are and Where C1 / C1′ is the number of filters in the E2E convolution operation;

[0019] b. Edge-to-node E2N convolution aims to aggregate edge features into node features. It extracts information from the neighboring edges of each node and generates a representation of each node. The formula is as follows:

[0020]

[0021] in, is the learning weight, (m,n) represents each pair of input and output feature maps of each layer, and the feature map after E2N convolution is recorded as and Where C2 / C2' is the number of filters in the E2N convolution operation;

[0022] As the mental illness identification method based on pulse structure-function brain network coupling provided by the present invention, the specific steps of step C are as follows:

[0023] a. Use the Leaky-Integrate-and-Fire (LIF) spiking neuron model to build a pulse coupled neural network. The LIF spiking neuron can be described by a differential function:

[0024]

[0025] Where τ represents the time constant, U(t) represents the membrane potential of the postsynaptic neuron, and I(t) represents the total input received from the presynaptic neuron ensemble at any given time. After solving this differential equation, the dynamics of the LIF neuron layer is governed by the following formula:

[0026] U[t]=βU[t-1]+X[t]-S out [t-1]θ (17)

[0027] Where t is the time step, X[t] is the spatial input information, U[t] is the membrane potential, θ is the threshold, and β = e -1 / τis the attenuation factor, S out [t]∈{0,1} is the output pulse generated by the neuron, and a pulse is generated if the membrane potential exceeds the threshold:

[0028]

[0029] If activated, the reset term subtracts the threshold θ from the membrane potential. Otherwise, the reset term has no effect;

[0030] b. The structural feature map after E2E and E2N convolution operations and functional feature maps As the input of the pulse coupled neural network, where T is the time step, it converts λ 1 and λ 2 Repeat T times as input:

[0031]

[0032] Where D is the feature dimension, is the spiking neuron layer, U′ sc and U′ fc It is through , so they are a binarized impulse tensor;

[0033] c. Through And the linear layer Liner(·) can get the structure and function impulse tensor:

[0034]

[0035] Among them, X sc represents the structured brain network spike tensor, X fc represents the functional brain network spike tensor,

[0036] d.X sc and X fc They are all binary pulse tensors, which contain important characteristic information of structural brain networks and functional brain networks. They simulate the interaction between structural connections and functional connections in the brain through dense interconnection of pulse neurons and synaptic weights to obtain pulse structure-function coupling:

[0037]

[0038] in, is the Hadmard product, SP sc-fc It is the spike structure-function coupling obtained by learning structural and functional brain networks in the form of spike trains, which enables us to identify subtle patterns of network dysfunction between healthy and disease connectomes.

[0039] As the mental illness identification method based on pulse structure-function brain network coupling provided by the present invention, the specific steps of step D are as follows:

[0040] a. Order Denotes the spikes emitted by each output neuron over time, which is a time-varying vector where m is the number of output categories. The pulse count for each output neuron can be obtained by Summing to obtain:

[0041]

[0042] b. The largest count The index of corresponds to the predicted class y:

[0043]

[0044] c. For binary classification tasks, the cross entropy loss is calculated as follows:

[0045]

[0046] Among them, L c is the cross entropy loss function, and y is the true category.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. Compared with most previous structure-function coupling analysis models, the method for identifying mental illness based on pulse structure-function brain network coupling proposed in the present invention not only fully considers the multimodal interaction between structural connection and functional connection, but also further studies the neurobiological mechanism between structural connection and functional connection.

[0049] 2. In view of the complex data characteristics of structural and functional brain networks, the pulse-coupled neural network proposed in the present invention provides a powerful analysis tool that can capture and simulate the dynamic interactions and time dependencies in brain networks. By simulating the pulse activity of neurons, it can more accurately reflect the complexity of the brain at the structural and functional levels, thus providing a new perspective for understanding how the brain achieves specific functions through its structure.

[0050] 3. The research on pulse structure-function coupling proposed in the present invention helps to better understand the pathophysiological mechanisms of brain diseases. It not only significantly improves the accuracy of identifying mental illnesses, but also assists doctors in diagnostic analysis to achieve efficient treatment efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention but do not constitute a limitation of the present invention.

[0052] Figure 1 This is the overall framework diagram of the mental illness identification method based on pulse structure-function brain network coupling of the present invention;

[0053] Figure 2 This is a model diagram of the mental illness identification method based on pulse structure-function brain network coupling of the present invention. DETAILED DESCRIPTION

[0054] In order to deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with embodiments and drawings. The embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0055] Example 1

[0056] See also Figure 1 , Figure 2 This embodiment is specifically implemented through the following technical solutions:

[0057] like Figure 1 and Figure 2 As shown in the figure, the method for identifying mental illness based on pulse structure-function brain network coupling has the following specific steps:

[0058] A. Functional brain networks were constructed by functional magnetic resonance imaging (fMRI). The fMRI data were processed by spm12 in the DPARSF toolbox. The initial functional time series were slice time acquisition corrected, realigned and normalized to the EPI template. Then a detrending process was performed to remove the source of pseudo-variance, namely six head motion parameters, the average signal of CSF and white matter, and the global brain signal. The time series were bandpass filtered at 0.01-0.08Hz. The resulting volume had 240 time points, which were divided into 90 regions of interest using the automatic anatomical labeling atlas AAL. Structural brain networks were constructed by diffusion tensor imaging (DTI). DTI data were processed using the PANDA suite and DTI distortion correction was performed using the FSL toolbox. Then, fiber images were obtained using a deterministic tracking method using TrackVis. Anatomical regions were defined using the AAL convention based on the registered T1 images of each subject.

[0059] The above model was tested using a psychiatric brain map dataset, which was acquired by a 3.0T Siemens Trio MRI scanner. The DTI scan parameters were as follows: repetition time = 6100ms, interval = 93ms, flip angle = 90°, voxel size = 0.94×0.94×3mm 3fMRI scan parameters were as follows: repetition time = 2000 ms, interval = 30 ms, flip angle = 90°, voxel size = 3.75 × 3.75 × 3.75 mm 3 .

[0060] B. Use BrainNetCNN to obtain information feature maps of structural brain networks and functional brain networks. α=1,2,where A 1 Represents the adjacency matrix of the structural brain network, A 2 Represents the adjacency matrix of the functional brain network, and N corresponds to the node set of the brain region. In order to obtain the complex features in the two brain network data, the edge-to-edge convolution E2E and edge-to-node convolution E2N operations proposed by Kawahara et al. are used to obtain the feature maps of the two brain networks; the specific steps are as follows:

[0061] a. Edge-to-edge E2E convolution, which aims to extract edge features from the original adjacency matrix. It uses the topological structure of the network to perform filtering operations by combining the weights of the edges of shared nodes. The formula is as follows:

[0062]

[0063] in, and is the learning weight, f n Represents the nonlinear activation function of the nth filter. The feature maps of the structural brain network and the functional brain network after E2E convolution are and Where C1 / C1′ is the number of filters in the E2E convolution operation;

[0064] b. Edge-to-node E2N convolution aims to aggregate edge features into node features. It extracts information from the neighboring edges of each node and generates a representation of each node. The formula is as follows:

[0065]

[0066] in, is the learning weight, (m,n) represents each pair of input and output feature maps of each layer, and the feature map after E2N convolution is recorded as and Where C2 / C2' is the number of filters in the E2N convolution operation;

[0067] The number of samples in the mental illness brain map dataset is 217, and the decision-making attributes are divided into two categories, namely, patients with confirmed mental illness and normal people. The connection matrices of the structural brain network and the functional brain network are and The E2E convolutional layer consists of 32 1×90 and 32 90×1 filters, and the generated structure and function feature maps are and The E2N convolutional layer consists of 64 1×90×32 filters, and the generated structural and functional feature maps are and

[0068] C. Use the Leaky-Integrate-and-Fire (LIF) pulse neuron model to build a pulse coupled neural network. The structural feature map λ after E2E and E2N convolution operations sc and the functional characteristic graph λ fc Repeat T times as input and perform implicit pulse encoding through internal pulse neurons, where T is the time step. Then use pulse coupled neural network to learn structure-function coupling mechanism from structure and function feature graph to find pulse structure-function coupling; the specific steps are as follows:

[0069] a. Use the Leaky-Integrate-and-Fire (LIF) spiking neuron model to build a pulse coupled neural network. The LIF spiking neuron can be described by a differential function:

[0070]

[0071] Where τ represents the time constant, U(t) represents the membrane potential of the postsynaptic neuron, and I(t) represents the total input received from the presynaptic neuron ensemble at any given time. After solving this differential equation, the dynamics of the LIF neuron layer is governed by the following formula:

[0072] U[t]=βU[t-1]+X[t]-S out [t-1]θ (30)

[0073] Where t is the time step, X[t] is the spatial input information, U[t] is the membrane potential, θ is the threshold, and β = e -1 / τ is the attenuation factor, S out [t]∈{0,1} is the output pulse generated by the neuron, and a pulse is generated if the membrane potential exceeds the threshold:

[0074]

[0075] If activated, the reset term subtracts the threshold θ from the membrane potential, otherwise, the reset term has no effect;

[0076] b. The structural feature map after E2E and E2N convolution operations And functional feature map As the input of the pulse coupled neural network, where T is the time step, it converts λ 1 and λ 2 Repeat T times as input:

[0077]

[0078] Where D is the feature dimension, is the spiking neuron layer, U′ sc and U′ fc It is through The obtained, therefore, is a binarized impulse tensor;

[0079] c. Through And the linear layer Liner(·) can get the structure and function impulse tensor:

[0080]

[0081] Among them, X sc represents the structured brain network spike tensor, X fc represents the functional brain network spike tensor,

[0082] d.X sc and X fc They are all binary pulse tensors, which contain important characteristic information of structural brain networks and functional brain networks. They simulate the interaction between structural connections and functional connections in the brain through dense interconnection of pulse neurons and synaptic weights to obtain pulse structure-function coupling:

[0083]

[0084] in, is the Hadmard product, SP sc-fc It is the spike structure-function coupling obtained by learning structural and functional brain networks in the form of spike trains, which enables us to identify subtle patterns of network dysfunction between healthy and disease connectomes.

[0085] A fully connected pulse coupled neural network was designed with 90×64 neurons in the input layer, 800 to 1200 neurons in the hidden layer, and 2 output neurons for classification, and a dropout of 0.5 was used for each layer. The decay factor of the LIF neurons was β = 0.9, and the time step T = {32, 64, ..., 256}.

[0086] D. Input the obtained pulse structure-function coupling into the classification layer, determine the predicted category through the pulse output sequence, and use the cross entropy loss between the predicted category and the true category as the loss function for training, and obtain a neural network model that identifies the subtle relationship between the healthy and disease connection groups based on the pulse structure-function coupling. The specific steps are as follows:

[0087] a. Order Denotes the spikes emitted by each output neuron over time, which is a time-varying vector where m is the number of output categories. The pulse count for each output neuron can be obtained by Summing to obtain:

[0088]

[0089] b. The largest count The index of corresponds to the predicted class y:

[0090]

[0091] c. For binary classification tasks, the cross entropy loss is calculated as follows:

[0092]

[0093] Among them, L c is the cross entropy loss function, and y is the true category.

[0094] Using the five-fold cross-validation method, the model was trained and tested independently for five rounds, with different training sets, test sets, and validation sets divided in each round. In these rounds, the highest accuracy recorded was 97.13%, and the comprehensive average accuracy was 92.68%. These data show that the model can significantly improve the accuracy of mental illness identification.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying mental illness based on pulse structure-function brain network coupling, characterized in that: The following steps are involved: S1. Functional brain networks were constructed by functional magnetic resonance imaging (fMRI). The fMRI data were processed by spm12 in the DPARSF toolbox. The initial functional time series were corrected for slice time acquisition, realigned, and normalized to the EPI template. Then a detrending process was performed to remove the source of pseudo-variance, and the time series were band-pass filtered at 0.01-0.08 Hz. The resulting volume had 240 time points, which were divided into 90 regions of interest using the automatic anatomical labeling atlas AAL. Structural brain networks were constructed by diffusion tensor imaging DTI. DTI data were processed using the PANDA suite and DTI distortion correction was performed using the FSL toolbox. Then, fiber images were obtained using a deterministic tracking method using TrackVis. Anatomical regions were defined using the AAL convention based on the registered T1 images of each subject. S2, use BrainNetCNN to obtain the information feature map of the structural brain network and functional brain network, let α=1,2, where A 1 Represents the adjacency matrix of the structural brain network, A 2 The adjacency matrix represents the functional brain network, N corresponds to the node set of the brain region. In order to obtain the complex features in the two brain network data, edge-to-edge convolution E2E and edge-to-node convolution E2N operations are used to obtain the feature maps of the two brain networks. S3. A pulse coupled neural network is constructed using the Leaky-Integrate-and-Fire pulse neuron model. The structural feature map λsc and the functional feature map λfc after E2E and E2N convolution operations are repeated T times as input, and implicit pulse encoding is performed through internal pulse neurons, where T is the time step. Then, the pulse coupled neural network is used to learn the structure-function coupling mechanism from the structural and functional feature maps to obtain the pulse structure-function coupling. S4. The obtained pulse structure-function coupling is input into the classification layer, the predicted category is determined by the pulse output sequence, and the cross entropy loss between the predicted category and the true category is used as the loss function for training, so as to obtain a neural network model that can identify the subtle relationship between healthy and disease connection groups based on the pulse structure-function coupling.

2. The method for identifying mental illness based on pulse structure-function brain network coupling according to claim 1, characterized in that: The specific steps of using BrainNetCNN to obtain two brain network information feature maps in step S2 are as follows: S21, edge-to-edge E2E convolution, aims to extract edge features from the original adjacency matrix. It utilizes the topological structure of the network and performs filtering operations by combining the weights of the edges of shared nodes. The formula is as follows: in, and is the learning weight, fn represents the nonlinear activation function of the nth filter, and the feature maps of the structural brain network and the functional brain network after E2E convolution are recorded as and Where C1 / C1′ is the number of filters in the E2E convolution operation; S22, edge-to-node E2N convolution, aims to aggregate edge features into node features. It extracts information from the neighboring edges of each node and generates a representation of each node. The formula is as follows: in, is the learning weight, (m,n) represents each pair of input and output feature maps of each layer, and the feature map after E2N convolution is recorded as and where C2 / C2' is the number of filters in the E2N convolution operation.

3. The method for identifying mental illness based on pulse structure-function brain network coupling according to claim 1, characterized in that: The specific steps of constructing the pulse coupled neural network described in step S3 are as follows: S31. Use the Leaky-Integrate-and-Fire spike neuron model to build a pulse coupled neural network, and use differential functions to describe the LIF spike neuron: Where τ represents the time constant, U(t) represents the membrane potential of the postsynaptic neuron, and I(t) represents the total input received from the presynaptic neuron ensemble at any given time. After solving this differential equation, the dynamics of the LIF neuron layer is governed by the following formula: U[t]=βU[t-1]+X[t]-S out [t-1]θ (4) Where t is the time step, X[t] is the spatial input information, U[t] is the membrane potential, θ is the threshold, β = e- 1τ is the attenuation factor, Sout[t]∈{0,1} is the output pulse generated by the neuron, and a pulse is generated if the membrane potential exceeds the threshold: If activated, the reset term subtracts the threshold θ from the membrane potential, otherwise, the reset term has no effect; S32, the structural feature map after E2E and E2N convolution operations And functional feature map As the input of the pulse coupled neural network, where T is the time step, it converts λ 1 and λ 2 Repeat T times as input: Where D is the feature dimension, is the spike neuron layer, Us′c and U′fc are The obtained, therefore, is a binarized impulse tensor; S33, through And the linear layer Liner(·) to get the structure and function impulse tensor: Among them, Xsc represents the structural brain network pulse tensor, Xfc represents the functional brain network pulse tensor, S34, Xsc and Xfc are all binary spike tensors, which contain important characteristic information of structural and functional brain networks. They simulate the interaction between structural and functional connections in the brain through dense interconnection of spike neurons and synaptic weights to obtain spike structure-function coupling: in, is the Hadmard product, SP sc-fc It is the spike structure-function coupling obtained by learning structural and functional brain networks in the form of spike trains, which is used to identify subtle patterns of network dysfunction between healthy and disease connection groups.

4. The method for identifying mental illness based on pulse structure-function brain network coupling according to claim 1, characterized in that: The specific steps of determining the prediction category from the pulse output sequence described in step S4 are as follows: S41. Order represents the pulses emitted by each output neuron over time, which is a time-varying vector, where m is the number of output categories, let The spike count for each output neuron is obtained by summing Z[t] over T time steps: S42, the largest count The index of corresponds to the predicted class y: S43. For binary classification tasks, the cross entropy loss is calculated as follows: Among them, Lc is the cross entropy loss function and y is the true category.

Images