Brain disease classification prediction method based on brain-like calculation

Abstract

The invention provides a brain disease classification prediction method based on brain-like calculation, aims at solving the problem that a brain network cannot be effectively combined with a brain-like calculation method, and is effectively applied to diagnosis and classification of brain diseases. On one hand, the high-order multi-mode brain network constructed by the method integrates brain structure and function information and structure-function coupling information, information loss caused by independent analysis of a structure network and a function network is overcome, multi-scale understanding of a brain mechanism and an abnormal mode is realized, and extraction of pathological features of people with brain diseases is facilitated. In addition, the brain disease classification prediction application model based on the brain-like calculation method provided by the invention combines the characteristics of GCN and SNN, obtains relatively high accuracy in classification prediction of cognitive memory impairment diseases, shows the potential of brain-like calculation in the field of brain science application, and provides a new technical means for diagnosis and intervention of brain diseases.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and brain science based on brain networks and brain-like computing, specifically to a brain disease classification and prediction method and prediction model based on brain-like computing. Background Technology

[0002] With the increasing aging of society, the incidence of cognitive impairment diseases such as Alzheimer's disease and mild cognitive impairment is rising, placing a heavy burden on society and families. Currently, the prevention and treatment of these diseases require a deeper understanding of the brain's cognitive and memory mechanisms. A typical method in brain mechanism research is the construction and analysis of brain networks. Brain networks can effectively reveal structural and functional information about the brain. In brain disease research, spiking neural networks have unique advantages due to their interpretability, low energy consumption, high precision, low latency, and neuromorphic hardware compatibility; however, their architectural characteristics make them difficult to integrate with brain networks, an important tool. These limitations lead to a less in-depth understanding of brain mechanisms such as cognitive memory, and also limit the accuracy and reliability of brain disease diagnosis. Summary of the Invention

[0003] This invention provides a brain disease classification and prediction method based on brain-like computing. The technical problem to be solved is: to address the issue that brain networks cannot be effectively combined with brain-like computing methods, this invention proposes an application model that combines graph convolutional networks and spiking neural networks, and effectively applies it to the diagnosis and classification of brain diseases.

[0004] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0005] A brain disease classification and prediction method based on neuromorphic computing is provided, comprising the following steps:

[0006] Step 1: Construction and Analysis of High-Order Multimodal Brain Networks

[0007] Based on magnetic resonance imaging data, canonical correlation analysis was used to construct brain structural and functional networks and extract structure-function coupling. A high-order multimodal brain network was constructed by fusing information from the brain structural network, brain functional network, and structure-function coupling based on a multi-layer network architecture.

[0008] Step 2: Application of brain disease classification and prediction combining multimodal brain networks and brain-like computing

[0009] This paper proposes a brain disease classification and prediction application model that combines multimodal brain networks and brain-inspired computing methods. Graph Convolutional Networks (GCNs) are used to extract effective representations of the multimodal brain networks, and spiking neural networks (SNNs) are combined for classification prediction. The constructed multimodal brain networks are used as input features of the model. Through feature extraction, encoding transformation, spiking embedding aggregation, and graph classification modules, a three-class classification prediction of Alzheimer's disease, mild cognitive impairment, and healthy status is achieved, thus providing auxiliary support for the diagnosis of brain diseases.

[0010] According to the brain disease classification and prediction method based on brain-like computing described in this application, the construction of the multimodal brain network in step 1 includes four steps: node definition, intra-layer relation definition, inter-layer relation definition, and network representation. The node representation of the multimodal brain network represents the brain regions of interest in the study, using 90 cerebral regions and 26 cerebellar regions defined by the AAL116 template. To better map the connection between structure and function, the same node definition is used for the structural layer and the functional layer of the multimodal brain network. Intra-layer relations refer to the connections between nodes within a layer. For the structural layer, the intra-layer relation is defined as the absolute value of the correlation between gray matter densities in brain regions. For the functional layer, the intra-layer relation is defined as the absolute value of the correlation between time series of blood oxygenation signals in brain regions. Thus, the structural layer is equivalent to a non-sparsed unimodal brain structural connectivity network, and the functional layer is equivalent to a non-sparsed unimodal brain functional connectivity network, denoted as SC network and FC network, respectively. For inter-layer relations, the structure-function coupling strength is defined. Specifically, the structure-function coupling of each brain region is defined by the following formula.

[0011] SFC i =|Pearson(SC) i FC i )|

[0012] Where i represents the brain region number, SC i and FC i SFC represents the structural and functional connectivity vectors of brain region i, respectively. i This refers to the structure-function coupling of brain region i; all structure-function coupling values ​​here are positive, meaning they range from 0 to 1. After calculating the coupling value for each brain region, a diagonal matrix is ​​constructed to obtain the whole-brain structure-function coupling network, i.e., the SFC network. SFC quantifies the statistical dependence between brain region structure and function, and can represent the correlation between structure and function. The multimodal network is represented using a superadjacency matrix; the superadjacency matrix is ​​represented by Super_Matrix.

[0013]

[0014] According to the brain disease classification and prediction method based on brain-like computing described in this application, the method is characterized in that, in step 2, a brain disease classification and prediction application model based on brain-like computing is constructed, and the model includes a feature extraction module, an encoding conversion module, a pulse embedding aggregation module, and a graph classification module.

[0015] (1) Feature extraction module

[0016] The feature extraction module is used to extract the embedded features of each node in the graph. As input to the model, each subject will construct a multimodal brain network. The MRI data of 30 subjects selected from the Human Connectome Project will be preprocessed to obtain their gray matter density volume. For each subject S in the subject group, a group of 31 subjects will be formed using the data of the 30 reference subjects and the data of S, and the structure matrix SC will be constructed. After the adjacency matrix of the multimodal network is input into the GCN, the most direct two-layer graph convolution will be used to aggregate the node and its neighbor information, as shown below.

[0017]

[0018] X is the feature matrix formed by the embeddings of all nodes in the graph. For the angle matrix, Let H be the adjacency matrix, and H be the new node feature matrix, which incorporates neighbor information.

[0019] (2) Encoding conversion module

[0020] The main purpose of this module is to perform encoding conversion, including converting node embeddings into pulse signals and pulse signals into embedding vectors;

[0021] The node embedding is converted into an impulse signal using Bernoulli sampling; the following equation shows a Bernoulli distribution; for an independent random variable X, it takes only two values, 0 and 1; p represents the probability of success in a single experiment;

[0022] P(X=k)=p k ·(1-p) 1-k

[0023] For a node embedding h i ={h i,1 ,h i,2 ,...,h i,n The following formula is used for feature encoding; where o i,j This represents the encoded pulse signal, while h i,j It follows a Bernoulli distribution; thus, the information is transformed from continuous numerical values ​​of node embeddings into binary values ​​that follow a Bernoulli distribution.

[0024] p(o i，j )~Bernoulli(hi，j )

[0025] For h i A single sampling constitutes a j-dimensional vector o. i The input is fed into a fully connected matrix of dimension j×n_class; where n_class represents the number of neurons, which is also the number of categories in the graph classification; o i After passing through a fully connected network, the simulated output voltage value is input into the neuron; the neuron goes through the charging, ignition, and reset process, and outputs an n_class dimensional embedding vector representing the ignition status of each neuron; a value of 1 indicates that the neuron is ignited, and a value of 0 indicates that it is not ignited;

[0026] (3) Pulse conversion coding module

[0027] In the encoding conversion module, for h i A single sampling will output an embedding vector; for h i Sample T times to obtain T embedding vectors; aggregate these T vectors to form a pulse time-frequency vector; output this vector to the graph classification module as a new node embedding for node i; for each node i in the graph, after passing through the encoding conversion module and the pulse embedding aggregation module, a new node embedding vector will be generated, thereby updating the node features of the entire graph once.

[0028] (4) Graph Classification Module

[0029] The main purpose of this module is to output the classification of the graph; at this point, the node feature matrix H of the graph has been updated; average pooling is performed on each node vector of H, as shown below;

[0030]

[0031] After obtaining the global embedding representation of the graph, a fully connected layer with dimension j×n_class is output, and the classification result is output after passing through a softmax layer.

[0032] prediction_class = argmax i softmax(g i )

[0033] The loss function used in this module is cross-entropy loss;

[0034]

[0035] Where N is the number of samples, C is the number of categories, and y i,j It is the true label of the i-th sample in the j-th category. It is the predicted probability of the i-th sample in the j-th category.

[0036] The beneficial effects of this invention are:

[0037] 1. The constructed high-order multimodal brain network integrates brain structural and functional information as well as structure-function coupling information, overcoming the information loss caused by analyzing structural and functional networks separately. It enables a multi-scale understanding of brain mechanisms and abnormal patterns, which helps to extract pathological characteristics of people with brain diseases.

[0038] 2. The proposed brain disease classification and prediction application model based on brain-inspired computing methods combines the characteristics of GCN and SNN, and has achieved high accuracy in the classification and prediction of cognitive and memory disorders. It demonstrates the potential of brain-inspired computing in the field of brain science applications and provides a new technical means for the diagnosis and intervention of brain diseases. Attached Figure Description

[0039] Figure 1 The process of constructing multimodal brain networks;

[0040] Figure 2 Application model for classifying and predicting brain diseases. Detailed Implementation

[0041] The following will be combined with the appendix Figures 1-2 This invention will be described in detail, and the technical solutions in the embodiments of this invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0042] This invention provides a brain disease classification and prediction method based on neuromorphic computing. The technical problem it aims to solve is the inability to effectively combine brain networks with neuromorphic computing methods. It proposes an application model combining graph convolutional networks and spiking neural networks, which can be effectively applied to the diagnosis and classification of brain diseases. The specific implementation is as follows;

[0043] 1. Construction and analysis of multimodal brain networks

[0044] Based on a multi-layer network architecture, a high-order multimodal brain network is constructed, such as... Figure 2As shown. The construction of a multimodal brain network includes four steps: node definition, intra-layer relation definition, inter-layer relation definition, and network representation. The nodes of the multimodal brain network represent the Region of Interest (ROI) being studied, using 90 cerebral regions and 26 cerebellar regions defined by the AAL116 template. To better map the connection between structure and function, the same node definition is used for both the structural and functional layers of the multimodal brain network. Intra-layer relations refer to the connections between nodes within a layer. For the structural layer, intra-layer relations are defined as the absolute value of the correlation between gray matter densities in brain regions; for the functional layer, intra-layer relations are defined as the absolute value of the correlation between time series of blood oxygenation signals in brain regions. Thus, the structural layer is equivalent to a non-sparsed unimodal structural connectivity network, and the functional layer is equivalent to a non-sparsed unimodal functional connectivity network, denoted as SC network and FC network, respectively. For inter-layer relations, the structure-function coupling strength is used. Specifically, the structure-function coupling of each brain region is defined by the following formula.

[0045] SFC i =|Pearson(SC) i FC i )|

[0046] Where i represents the brain region number, SC i and FC i SFC represents the structural and functional connectivity vectors of brain region i, respectively. i This represents the structure-function coupling of brain region i. For interpretability and ease of subsequent calculations, the structure-function coupling values ​​here are all positive, ranging from 0 to 1. After calculating the coupling value for each brain region, a diagonal matrix is ​​constructed to obtain the whole-brain structure-function coupling network, i.e., the SFC network. SFC quantifies the statistical dependence between brain region structure and function, representing the correlation between structure and function. The multimodal network is represented using a superadjacency matrix. The superadjacency matrix is ​​denoted by Super_Matrix.

[0047]

[0048] 2. Application Model for Classification and Prediction of Brain Diseases

[0049] Construct a brain disease classification and prediction application model based on brain-inspired computing methods. For example... Figure 2 As shown, the model includes a feature extraction module, an encoding conversion module, a pulse embedding aggregation module, and a graph classification module.

[0050] (1) Feature extraction module

[0051] The feature extraction module is used to extract the embedded features of each node in the graph. As input to the model, each subject constructs a multimodal brain network. This raises the issue that while a fully connected (FC) network can be extracted for each individual subject, only a single structural (SC) network can be extracted for each subject group. A collaborative structural network strategy is employed to address this problem. Specifically, MRI data from 30 subjects selected from the Human Connectome Project (HCP) are preprocessed to obtain their gray matter density volumes. For each subject S in the subject group, a structural matrix SC is constructed using data from the 30 reference subjects and the data from S, totaling 31 subjects. It is worth noting that the carefully selected 30 subjects have a similar average age and gender ratio to the experimental group.

[0052] After the adjacency matrix of the multimodal network is input into the GCN, the most direct two-layer graph convolution is used to aggregate the node and its neighbor information, as shown below.

[0053]

[0054] X is the feature matrix formed by the embeddings of all nodes in the graph. For the angle matrix, Let H be the adjacency matrix. H is a new node feature matrix that incorporates neighbor information.

[0055] (2) Encoding conversion module

[0056] The main purpose of this module is to perform encoding conversion, including converting node embeddings into pulse signals and pulse signals into embedding vectors.

[0057] The node embedding is converted into an impulse signal using Bernoulli sampling. The following equation represents a Bernoulli distribution. For an independent random variable X, it takes on only two values: 0 and 1. p represents the probability of success in a single experiment.

[0058] P(X=k)=p k ·(1-p) 1-k

[0059] For a node embedding h i ={h i,1 ,h i,2 ,...,h i,n The following formula is used for feature encoding. Where o i,j This represents the encoded pulse signal, while h i,j It follows a Bernoulli distribution. In this way, the information is transformed from continuous numerical values ​​of node embeddings into binary values ​​that follow a Bernoulli distribution.

[0060] p(o i，j)~Bernoulli(h i,j )

[0061] For h i A single sampling constitutes a j-dimensional vector o. i The input is fed into a fully connected matrix of dimension j×n_class. Here, n_class represents the number of neurons, which is also the number of categories in the graph classification. i The simulated output voltage value is fed into the neuron after passing through a fully connected network. The neuron undergoes a charging, ignition, and reset process, outputting an n_class dimensional embedding vector representing the ignition status of each neuron. A value of 1 indicates that the neuron is ignited, and a value of 0 indicates that it is not ignited.

[0062] (3) Pulse conversion coding module

[0063] In the encoding conversion module, for h i Each sampling will output an embedding vector. For h i Sampling is performed T times to obtain T embedding vectors. These T vectors are then aggregated to form a pulse time-frequency vector. This vector is output to the graph classification module as a new node embedding for node i. For each node i in the graph, after passing through the encoding conversion module and the pulse embedding aggregation module, a new node embedding vector is generated, thereby updating the node features of the entire graph.

[0064] (4) Graph Classification Module

[0065] The primary purpose of this module is to output a graph classification. At this point, the node feature matrix H of the graph has been updated. Average pooling is then performed on each node vector of H, as shown below.

[0066]

[0067] After obtaining the global embedding representation of the graph, a fully connected layer with dimension j×n_class is output, and the classification result is output after passing through a softmax layer.

[0068] prediction_class = argmax i softmax(g i )

[0069] This module uses cross-entropy loss as its loss function.

[0070]

[0071] Where N is the number of samples, C is the number of categories, and y i,j It is the true label of the i-th sample in the j-th category. It is the predicted probability of the i-th sample in the j-th category.

[0072] This application presents a brain disease classification and prediction method based on brain-inspired computing. This patented technology constructs a high-order multimodal brain network (integrating structural, functional, and structure-function coupling information) and combines it with a brain-inspired computing model of graph convolutional networks (GCN) and spiking neural networks (SNN) to achieve accurate classification and prediction of brain diseases such as Alzheimer's disease and mild cognitive impairment. Its innovation lies in overcoming the limitations of traditional brain network analysis, providing an efficient and reliable solution for practical applications. This method can be widely applied in the following fields to assist in brain health management and disease intervention: This method has significant application value in the following areas:

[0073] 1. Clinical medical auxiliary diagnostic field:

[0074] In general hospitals or neurology clinics, this method can serve as an intelligent assistive tool, helping doctors quickly identify individuals at risk for brain diseases. For example, combined with standard brain imaging equipment (such as MRI), the system automatically constructs a multimodal brain network and performs a three-category prediction (Alzheimer's disease, mild cognitive impairment, and healthy status), providing objective evidence for diagnostic decisions. This not only shortens screening time but also improves the accuracy of early detection, making it particularly suitable for resource-constrained medical environments. (The document continues...) Figure 2 The model architecture shown (brain disease classification and prediction application model) can be directly deployed in hospital information systems for seamless integration.

[0075] 2. Community health monitoring and screening:

[0076] This method is applicable to large-scale brain health screening programs in community health centers or elderly care facilities. Data is collected using portable brain imaging technology, and the system utilizes structure-function coupled networks (SNNs) to capture subtle pathological features, combined with the low-energy characteristics of SNNs for edge computing. This makes it possible to efficiently conduct age-related cognitive decline screenings in community settings, without relying on high-end equipment or specialized personnel, significantly reducing public health costs and promoting preventative interventions for brain diseases.

[0077] 3. Neuroscience research and drug development:

[0078] In the field of pharmaceutical research and development, this method can serve as a dynamic evaluation tool to track the efficacy of neuromodulation therapies or new drugs. For example, based on the characteristics of pulse time-frequency vectors, the system can quantify changes in brain networks before and after intervention, providing quantifiable biomarkers for drug efficacy. This not only accelerates the clinical trial process but also promotes basic research on the mechanisms of brain diseases, contributing to the development of personalized medicine. (The document contains...) Figure 1 The multimodal brain network construction process shown provides core support for data standardization.

[0079] Summary of application value:

[0080] All of the above applications benefit from the core advantages of this patent: the high-order multimodal brain network enables multi-scale brain mechanism analysis, while the GCN-SNN fusion model ensures high accuracy and low energy consumption. These applications not only expand the practical boundaries of brain science and technology but also provide more timely and convenient services for patients with brain diseases, fully demonstrating the translational potential of brain-inspired computing in the medical and health field. In the future, with the popularization of brain imaging technology, this method is expected to be implemented in more scenarios, promoting the progress of precision medicine.

[0081] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A brain disease classification and prediction method based on brain-inspired computing, characterized in that, Includes the following steps: Step 1: Construction and analysis of high-order multimodal brain networks; Based on magnetic resonance imaging data, canonical correlation analysis was used to construct brain structural networks and brain functional networks, and structure-function coupling was extracted. Based on a multi-layered network architecture, a high-order multimodal brain network is constructed by integrating brain structural networks, brain functional networks, and structure-function coupling information. Step 2: Application of brain disease classification and prediction combining multimodal brain networks and brain-like computing; This paper proposes a brain disease classification and prediction application model that combines multimodal brain networks and brain-like computing methods. Graph Convolutional Networks (GCNs) are used to extract effective representations of the multimodal brain networks, and spiking neural networks (SNNs) are combined for classification prediction. The constructed multimodal brain networks are used as input features of the model. Through feature extraction, encoding transformation, spiking embedding aggregation, and graph classification modules, a three-class classification prediction of Alzheimer's disease, mild cognitive impairment, and healthy status is achieved, thus providing auxiliary support for the diagnosis of brain diseases.

2. The brain disease classification and prediction method based on brain-like computing according to claim 1, characterized in that, Step 1, the construction of the multimodal brain network, includes four steps: node definition, intra-layer relation definition, inter-layer relation definition, and network representation. The nodes of the multimodal brain network represent the brain regions of interest, using 90 cerebral regions and 26 cerebellar regions defined by the AAL116 template. To better map the connection between structure and function, the structural and functional layers of the multimodal brain network use the same node definition. Intra-layer relations refer to the connections between nodes within a layer. For the structural layer, intra-layer relations are defined as the absolute value of the correlation between gray matter densities in brain regions; for the functional layer, intra-layer relations are defined as the absolute value of the correlation between time series of blood oxygenation signals in brain regions. Thus, the structural layer is equivalent to a non-sparsed unimodal brain structural connectivity network, and the functional layer is equivalent to a non-sparsed unimodal brain functional connectivity network, denoted as SC network and FC network, respectively. For inter-layer relations, the structure-function coupling strength is defined. Specifically, the structure-function coupling of each brain region is defined by the following formula. SFC i =|Pearson(SC i ,FC i )| Where i represents the brain region number, SC i and FC i SFC represents the structural and functional connectivity vectors of brain region i, respectively. i This refers to the structure-function coupling of brain region i; all structure-function coupling values ​​here are positive, meaning they range from 0 to 1. After calculating the coupling value for each brain region, a diagonal matrix is ​​constructed to obtain the whole-brain structure-function coupling network, i.e., the SFC network. SFC quantifies the statistical dependence between brain region structure and function, and can represent the correlation between structure and function. The multimodal network is represented using a superadjacency matrix; the superadjacency matrix is ​​represented by Super_Matrix.

3. The brain disease classification and prediction method based on brain-like computing according to claim 1, characterized in that, Step 2 involves constructing a brain disease classification and prediction application model based on brain-like computing methods. The model includes a feature extraction module, an encoding conversion module, a pulse embedding aggregation module, and a graph classification module. (1) Feature extraction module The feature extraction module is used to extract the embedded features of each node in the graph; As input to the model, each subject will construct a multimodal brain network. MRI data of 30 subjects will be selected from the Human Connectome Project, and data preprocessing will be performed to obtain their gray matter density volume. For each subject S in the subject group, a structure matrix SC will be constructed using the data of the 30 reference subjects and the data of S, totaling 31 subjects. After the adjacency matrix of the multimodal network is input into the GCN, the most direct two-layer graph convolution will be used to aggregate the node and its neighbor information, as shown below. X is the feature matrix composed of all node embeddings in the graph, D is the angle matrix, A is the adjacency matrix, and H is the new node feature matrix, which integrates neighbor information. (2) Encoding conversion module The main purpose of this module is to perform encoding conversion, including converting node embeddings into pulse signals and pulse signals into embedding vectors; The node embedding is converted into an impulse signal using Bernoulli sampling; the following equation shows a Bernoulli distribution; for an independent random variable X, it takes only two values, 0 and 1; p represents the probability of success in a single experiment; P(X=k)=p k ·(1-p) 1-k For a node embedding h i ={h i,1 ,h i,2 ,...,h i,n The following formula is used for feature encoding; where o i,j This represents the encoded pulse signal, while h i,j It follows a Bernoulli distribution; thus, the information is transformed from continuous numerical values ​​of node embeddings into binary values ​​that follow a Bernoulli distribution. p(o i,j )~Bernoulli(h i，j ) For h i A single sampling constitutes a j-dimensional vector o. i The input is fed into a fully connected matrix of dimension j×n_class; where n_class represents the number of neurons, which is also the number of categories in the graph classification; o i After passing through a fully connected network, the simulated output voltage value is input into the neuron; the neuron goes through the charging, ignition, and reset process, and outputs an n_class dimensional embedding vector representing the ignition status of each neuron; a value of 1 indicates that the neuron is ignited, and a value of 0 indicates that it is not ignited; (3) Pulse conversion coding module In the encoding conversion module, for h i A single sampling will output an embedding vector; for h i Sample T times to obtain T embedding vectors; aggregate these T vectors to form a pulse time-frequency vector; output this vector to the graph classification module as a new node embedding for node i; for each node i in the graph, after passing through the encoding conversion module and the pulse embedding aggregation module, a new node embedding vector will be generated, thereby updating the node features of the entire graph once. (4) Graph Classification Module The main purpose of this module is to output the classification of the graph; at this point, the node feature matrix H of the graph has been updated; average pooling is performed on each node vector of H, as shown below; After obtaining the global embedding representation of the graph, a fully connected layer with dimension j×n_class is output, and the classification result is output after passing through a softmax layer. prediction_class=argmax i softmax(g i ) The loss function used in this module is cross-entropy loss; Where N is the number of samples, C is the number of categories, and y i,j It is the true label of the i-th sample in the j-th category. It is the predicted probability of the i-th sample in the j-th category.

Images