A brain network classification method based on deep hashing mutual learning

Abstract

The present invention discloses a brain network classification method based on deep hash mutual learning, comprising: data preprocessing and brain functional network construction; brain network data partitioning; individual feature extraction based on deep hash learning; group feature extraction based on deep hash learning; mutual learning based on hash codes; and classification based on hash codes. For the first time, the present invention considers the differences in phenotypic labels in group brain networks, constructs a group brain network relationship graph using phenotypic labels, and proposes a deep hash learning model based on GCN to extract group features of brain networks. Furthermore, considering the relationship between individual and group features of brain networks, a brain network classification method based on deep hash mutual learning is adopted to enhance feature discrimination through mutual learning between individual and group features. Compared with other methods, this method has better classification performance.

Description

Technical Field

[0001] The present invention relates to brain network classification technology in the field of brain science. It mainly realizes the interaction of individual brain network characteristics and group brain network characteristics through mutual learning between two deep hash learning models, and designs a brain network classification method based on deep hash mutual learning. Background Art

[0002] In brain science, brain functional network classification refers to the use of feature mining and analysis of human brain functional network data to automatically determine whether a subject has a neurological disorder. This technology provides an effective tool for understanding the pathogenesis of brain diseases and for early diagnosis, and therefore has significant research and application value. In recent years, resting-state functional magnetic resonance imaging (rs-fMRI) has been widely used to construct brain functional networks due to its ability to reveal patterns of spontaneous functional brain activity. The constructed brain functional network consists of nodes and edges, where nodes correspond to neurons, neuronal clusters, brain regions, or regions of interest (ROIs). Edges are typically defined as the connections between pairs of nodes, also known as functional connectivity (FC), with the FC strength representing the edge weight. Hereinafter, the term "brain network" will refer to all brain functional networks.

[0003] At present, the most popular brain network classification methods are those based on machine learning, which mainly include traditional machine learning methods and deep learning methods. Among the traditional machine learning methods, the most widely used are support vector machines (SVM) and least absolute shrinkage and selection operators (LASSO), but they are all shallow models. When faced with high-dimensional brain network data, their classification performance is poor due to insufficient feature extraction capabilities. In recent years, deep learning methods have received increasing attention in brain network classification because they can learn richer and deeper features by processing the original brain network layer by layer. In 2017, Kawahara et al. proposed a convolutional neural network (CNN) framework called BrainNetCNN, which designed edge-to-edge, edge-to-point, and point-to- Figure 3A convolutional filter extracts brain network features. Experimental results show that the brain network features extracted by the BrainNetCNN framework can be used to predict clinical neurodevelopmental outcomes. In 2018, Heinsfeld et al. developed a model based on deep neural networks (DNNs) to identify patients with autism spectrum disorder (ASD) based on brain activation patterns. Experimental results showed that the DNN model can improve the accuracy of ASD recognition and identify the brain regions that are most helpful in distinguishing ASD patients from healthy controls. In 2019, Ju et al. proposed a framework based on the stacked sparse autoencoding (SSAE) method for distinguishing normal aging from mild cognitive impairment, and achieved significant improvements in classification accuracy compared to traditional machine learning methods. In 2020, Xing et al. proposed a new brain network classification method based on adaptive multi-task CNN (AMTCNN), which achieved good results in classifying ASD patients from healthy controls. In 2021, Ji et al. proposed a new convolution kernel with an element-wise weighting mechanism (CKEW) and a CKEW-based CNN classification framework. Experimental results show that the framework can more accurately classify brain networks and identify abnormal connection patterns related to brain diseases. In 2022, Ji et al. used diagnostic and clinical phenotypic labels to construct a semantic space, and combined the proximity relationship of brain networks in the semantic space to propose a brain functional network classification method based on deep graph hashing learning (BNC-DGHL). Experimental results on three datasets show that this method can achieve better classification performance, which also demonstrates the important auxiliary role of phenotypic labels in brain network classification. However, existing methods only consider the differences in phenotypic labels between individual brain networks, and ignore the differences in phenotypic labels between group brain networks. In fact, in classification tasks, the differences in phenotypic labels between group brain networks are more relevant to the classification results. For example, the gender label of a single subject's brain network alone cannot determine whether the subject has ASD, but based on the gender label differences of a group of subjects' brain networks, it is easy to infer that males are more likely to have ASD than females. In other words, the phenotypic characteristics of the group brain network are more discriminative than the phenotypic characteristics of the individual brain network. Summary of the Invention

[0004] In response to the above-mentioned problems existing in the prior art, the present invention proposes a brain network classification method based on Deep Hashing Mutual Learning (DHML), which mainly enhances the discriminative ability of features through mutual learning between individual features and group features. Specifically, DHML first uses a deep hash learning model based on CNN to extract individual features of the brain network and map them into hash codes, in which a separable convolutional layer is designed to extract multi-level individual topological features; secondly, DHML obtains the adjacency matrix of the brain network based on the similarity relationship between phenotypic labels, and constructs a group relationship graph together with individual features; thirdly, DHML uses a deep hash learning model based on GCN (Graph Convolutional Network) to extract group features of the brain network and map them into hash codes; then, the two deep hash learning models learn from each other by minimizing the distribution difference between hash codes, thereby realizing the interaction between individual features and group features; finally, the diagnostic label of the new instance is predicted by calculating the similarity between hash codes.

[0005] The technical solution adopted by the present invention is a brain network classification method based on deep hash mutual learning, which includes the following steps:

[0006] Step 1: Data preprocessing and brain functional network construction.

[0007] The subjects' rs-fMRI data were obtained and first preprocessed using the DataProcessing Assistant for Resting-State fMRI (DPARSF). Then, the brain functional network was constructed. The specific method is as follows.

[0008] Step 1.1: Preprocess the raw rs-fMRI data using DPARSF.

[0009] (1) Delete the first 5 time points of the original rs-fMRI time series.

[0010] (2) The linear interpolation method is used to eliminate the time phase difference caused by interlayer scanning and complete the layer time correction.

[0011] (3) The data of subjects with horizontal head movement exceeding 2 mm and rotational head movement exceeding 2° were eliminated to complete the head movement correction.

[0012] (4) Perform rs-fMRI image registration, smoothing, and filtering based on T1 images.

[0013] Step 1.2: Construct brain network.

[0014] (1) Locate the location of the brain region in the cerebral cortex as the node of the brain network based on the brain map.

[0015] (2) The Pearson correlation coefficient of the inter-node rs-fMRI time series was calculated as the functional connectivity strength between brain regions, that is, the weight of the connecting edges between brain region nodes.

[0016] Step 2: Brain network data division.

[0017] The brain network dataset is divided into training set X according to the ratio of 8:1:1. t , validation set X a and the test set X e . Training set There are N brain networks in it, and the corresponding label set is represented by Y t , which contains the diagnostic label set and the phenotypic label set is d represents the diagnostic label and p represents the phenotypic label. Similarly, the validation set There are M brain networks in it, and the corresponding label set is expressed as Test set There are K brain networks in it, and the corresponding label set is expressed as

[0018] Step 3: Individual feature extraction based on deep hash learning.

[0019] Adopting CNN-based deep hashing learning model I H The topological features of individual brain networks are extracted and mapped into hash codes. Specifically, a five-layer CNN is first designed for feature extraction, and the extraction process is defined as Then a hash layer is designed for feature mapping, and the mapping process is defined as

[0020] Step 3.1, individual feature extraction.

[0021] For feature extraction, a separable CNN is used to extract topological features of individual brain networks at the edge, node, and graph levels. Specifically, L0 is the input layer that receives the brain network; L1 is a separable edge-to-edge (E2E) layer, which includes two processes: channel-by-channel E2E (depthwise E2E, E2E-DW) and point-by-point E2E (E2E-PW), for extracting edge-level topological features from the input brain network; L2 is a separable edge-to-node (E2N) layer, which includes two processes: channel-by-channel E2N (E2N-DW) and point-by-point E2N (E2N-PW), for extracting node-level features from edge-level features; L3 is a node-to-graph (N2G) layer, for extracting graph-level features from node-level features; and L4 is a fully connected layer. The detailed process is as follows.

[0022] (1) Define the feature extraction process of L0-L4 layer CNN as The individual features of the brain network IF are extracted during the training process t , the formula is:

[0023]

[0024] Among them, X t is the training set, which contains N brain networks.

[0025] (2) Define the cross entropy loss function J1 acting on the L0-L4 layer as:

[0026]

[0027]

[0028] Among them, W c and b represent the weights and biases of the separable CNN, respectively. and b q ∈b are the weights and biases of the Softmax layer, and N is the training set X t The number of brain networks included, It's X t The diagnostic label set, is the Softmax function, IF t is the individual characteristic of the brain network, ξ(IF t ) is the predicted probability of the diagnostic label of the individual feature.

[0029] Step 3.2, individual feature mapping.

[0030] For feature mapping, a hash layer is designed to first map the extracted individual features into a feature vector of length l using a hash function, where l is the predefined hash code length. The feature vector is then binarized to obtain the hash code representation corresponding to the individual topological features. The specific process is as follows.

[0031] (1) Define the feature mapping process of the hash layer as The extracted individual features IF t After the hash layer is mapped to the hash code IU t , the formula is:

[0032]

[0033] in, and b h ∈b are the weight and bias of the hash layer, σ(·) is the sigmoid function, sgn(·) is the sign function, IU t ∈{0, 1} N×l IF t The hash code represents, N represents X t The number of brain networks contained in the training set X, l is the predetermined hash code length, and each hash code is an l-dimensional vector with a value of 0 or 1. t The hash code set IU t Also known as a hash code library.

[0034] (2) Define the similarity preservation loss function J2 acting on the hash layer as:

[0035]

[0036]

[0037]

[0038] Where S∈{0,1} N×N It is a pairwise similarity matrix constructed based on diagnostic labels, N represents X t The number of brain networks included in t (i) With brain network X t (j) have the same diagnostic label, i.e. Then s ij =1, otherwise s ij =0;iu t (i) and iu t (j) are brain networks X t (i) and X t (j) Hash code representation of individual features.

[0039] Step 3.3, objective function construction.

[0040] By integrating the loss function J1 of the feature extraction part and the loss function J2 of the feature mapping part, the deep hash learning model I based on CNN H The objective function is defined as:

[0041]

[0042] Among them, W c and b are model I H The last term is a regularization term to prevent the model from overfitting during the training phase, λ1, λ2 and λ r is a hyperparameter.

[0043] Step 4: Group feature extraction based on deep hash learning.

[0044] Adopting GCN-based deep hash learning model G H Extract the topological features of the group brain network and map them into hash codes. Specifically, first construct the adjacency matrix A based on the similarity of the phenotypic labels, and then combine the individual brain network features IF extracted from step 3 t Construct a group relationship graph P, and finally use GCN to extract the group features of the brain network and map them into hash codes through the hash layer. The GCN extraction process is defined as The mapping process is defined as

[0045] Step 4.1: Construct the adjacency matrix A of the group brain network.

[0046] (1) Calculate the Mahalanobis distance between brain network phenotype label vectors. The calculation formula is as follows:

[0047]

[0048] Where D = {d ij |i, j = 1, ..., N} is the Mahalanobis distance, which can eliminate the scale differences between different phenotypic labels. Brain network x t (i) and x t (j) phenotypic label vector, C∈R N×N is the covariance matrix, N represents X t The number of midbrain networks, C -1 is the inverse of the covariance matrix.

[0049] (2) Define the adjacency matrix of the group brain network:

[0050]

[0051] stα∈(0,1)

[0052] Among them, A={a ij |i, j = 1, ..., N} is the adjacency matrix, a ij =1 indicates two brain networks x t (i) and x t (j) There is a connection between ij =0 means brain network x t (i) and x t (j) are not adjacent nodes, α is used to control the number of connecting edges between brain networks, D is the Mahalanobis distance, and max(D) is the maximum value of D.

[0053] Step 4.2: Construct the group brain network relationship graph P.

[0054] Taking a single brain network {x t (1), x t (2), ..., x t (N)} is a node, A is an adjacency matrix, which is used to determine whether there is an edge between nodes. t For the initial attribute characteristics of the node, define the group relationship graph P(A, IF t ), where A is the adjacency matrix representing the graph structure, IF t is the feature matrix representing the node attributes.

[0055] Step 4.3, group feature extraction.

[0056] For feature extraction, GCN is used to extract features from the group brain network. Specifically, L′0 is the input layer that receives the group relationship graph P; L′1-L′2 are graph convolution layers that aggregate features of adjacent nodes. The specific process is as follows.

[0057] (1) Define the feature extraction process of L′0-L′2 layer GCN as The group feature GF of the brain network is extracted during the training process t , the formula is:

[0058]

[0059] Among them, X t is the training set, which contains N brain networks.

[0060] (2) Define the cross entropy loss function J3 acting on the L′0-L′2 layer as:

[0061]

[0062] Among them, W g is the weight of GCN, N is X tThe number of brain networks in For X t The diagnostic label set, is the Softmax function, GF t is the group characteristic of the brain network, The predicted probability of the diagnostic label for the group feature.

[0063] Step 4.4, group feature mapping.

[0064] For feature mapping, a hash layer is designed to map the extracted group features into a hash code representation of length l through a hash function. The specific process is as follows.

[0065] (1) Define the feature mapping process of the hash layer as The extracted group features GF t After the hash layer is mapped to the hash code GU t , the formula is:

[0066]

[0067] Among them, σ(·) is the sigmoid function, and sgn(·) is the sign function. t is the group characteristic GF t The hash code representation of , each hash code is an l-dimensional vector with a value of 0 or 1, where l is a predetermined hash code length.

[0068] (2) Define the similarity preservation loss function J4 acting on the hash layer as:

[0069]

[0070]

[0071] Where S∈{0,1} N×N is the pairwise similarity matrix constructed based on the diagnostic labels, s ij ∈S, N represents X t The number of brain networks included in t (i) and gu t (j) are brain networks X t (i) and X t (j) Hash code representation of group characteristics.

[0072] Step 4.5, objective function construction.

[0073] By integrating the loss function J3 of the feature extraction part and the loss function J4 of the feature mapping part, the deep hash learning model G based on GCN H The objective function is defined as:

[0074] J(W g )=λ3J3+λ4J4

[0075] Among them, W g For model G H The weights of , λ3 and λ4 are hyper parameters.

[0076] Step 5: Mutual learning based on hash codes.

[0077] In order to make the two deep hash learning models I H and G H During the training process, they collaboratively learn and explore each other's learned features, and define the maximum average difference as the loss function of the mutual learning process:

[0078]

[0079]

[0080] Among them, IU t Is an individual characteristic IF t The hash code of GU t is the group characteristic GF t The hash code representation of δ(·) is the Gaussian kernel function, and η is the standard deviation parameter that controls the range of δ(·). Therefore, the deep hash learning model I based on CNN H The objective function is updated as:

[0081]

[0082] Among them, W c and b are model I H The weights and biases of λ1, λ2, λ m and λ r is a hyperparameter.

[0083] GCN-based deep hash learning model G H The objective function is updated as:

[0084] J(W g )=λ3J3+λ4J4+λ′ m J m (14)

[0085] Among them, W g For model G H The weights, λ3, λ4 and λ′ m is a hyperparameter. The two models are jointly optimized by learning from each other until the objective function converges.

[0086] Step 6: Classification based on hash codes.

[0087] After mutual learning steps, the trained I H After the model is built, use the validation set X a To verify the classification performance of the model and further adjust the model parameters. The specific process is as follows.

[0088] (1) Obtain the validation set X a The hash code represents IU a , the calculation formula is as follows:

[0089]

[0090] in, For I H The feature extraction function of the model, For I H The feature mapping function of the model.

[0091] (2) The hash code IU of the training set t (Formula (3)) is used as the hash code library to calculate IU a Each hash code iu a (i)With IU t The distance between all hash codes in IU t Find the hash code with the smallest distance and use its label as the predicted label The calculation formula is expressed as:

[0092]

[0093] in, Represents the training set X t The true diagnostic label of N is the training set X t The number of brain networks included, M is X a The number of brain networks included in It is the XOR logical operation.

[0094] (3) When the loss function of the DHML method reaches convergence and the prediction accuracy of the validation set no longer improves, the model and hash code library at this time are retained for testing the brain network X e Classification of diagnostic labels.

[0095] Compared with the prior art, the present invention has the following advantages:

[0096] (1) This paper takes into account the differences in phenotypic labels in group brain networks for the first time, uses phenotypic labels to construct a group brain network relationship diagram, and proposes a deep hash learning model based on GCN to extract group characteristics of brain networks. Compared with other methods, the extracted features contain richer group difference information and have stronger discrimination ability.

[0097] (2) The present invention takes into account the relationship between individual characteristics and group characteristics of brain networks, and adopts a brain network classification method based on deep hash mutual learning to enhance the feature discrimination ability through mutual learning between individual characteristics and group characteristics. Compared with other methods, the classification performance is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 It is the overall framework of the DHML method.

[0099] Figure 2 It is a separable E2E convolution layer.

[0100] Figure 3 is a separable E2N convolutional layer. DETAILED DESCRIPTION

[0101] The invention will be further described below with reference to the accompanying drawings and examples.

[0102] The overall framework of the brain network classification method based on deep hash mutual learning is as follows Figure 1 shown.

[0103] Step 1: Based on different brain map divisions, O is used to represent the number of nodes in the brain region, and the constructed brain network is represented as an O×O matrix.

[0104] Step 2: training set X t , validation set X a and the test set X e It contains N, M and K brain networks respectively, N:M:K=8:1:1.

[0105] Step 3, individual feature extraction based on deep hash learning Figure 1 As shown in module a). For the feature extraction part, L0 is the input layer that receives the brain network, where O represents the number of brain regions and the input brain network is represented as an O×O matrix; L1 is a separable E2E layer, which includes two processes, E2E-DW and E2E-PW, and is used to extract edge-level topological features from the input brain network; Figure 2 As shown in Figure 1, E2E-DW first convolves the input matrix with the column convolution kernel O×1, then convolves with the row convolution kernel 1×O, and finally adds the output convolution results; E2E-PW convolves the result of E2E-DW with 32 1×1 convolution kernels; L2 is a separable E2N layer, which includes two processes, E2N-DW and E2N-PW, and is used to extract node-level features from edge-level features; as shown in Figure 1 Figure 3As shown in the figure, E2N-DW first convolves the result of the L1 layer with 32 O×1 column convolution kernels and transposes them, then convolves with 32 1×O row convolution kernels, and finally adds the two convolution results. E2N-PW convolves the result of E2N-DW with 64 1×1 convolution kernels. L3 is the N2G layer, which contains 128 O×1 convolution kernels and is used to extract graph-level features from node-level features. L4 is a fully connected layer, containing 96 1×1 convolution kernels, which obtains individual features of the brain network. For the feature mapping part, a hash layer is included to map the extracted individual features of the brain network into binary hash code representations.

[0106] Step 4: Group feature extraction based on deep hash learning Figure 1 As shown in module b). Taking a single brain network {x t (1), x t (2), ..., x t (N)} is a node, A is an adjacency matrix, which is used to determine whether there is an edge between nodes. t For the initial attribute characteristics of the node, define the group relationship graph P(A, IF t ), where A is the N×N adjacency matrix representing the graph structure, IF t is an N×96 feature matrix representing node attributes, where N is the training set X t The number of brain networks included; the group relationship graph P first passes through GCN to obtain the brain network of the brain network, and then passes through the hash layer to obtain the hash code representation of the group characteristics.

[0107] Step 5, mutual learning based on hash code Figure 1 As shown in module c).

[0108] Step 6, classification based on hash code Figure 1 Taking the three brain atlases of the ABIDE I dataset (AAL, Dosenbach160, and CC200) as an example, the classification performance comparison with other algorithms is shown in Table 1.

[0109] Table 1 Classification performance comparison results

[0110]

Claims

1. A brain network classification method based on deep hash mutual learning, characterized by: The discriminative ability of features is enhanced through mutual learning between individual features and group features. First, a CNN-based deep hash learning model is used to extract individual features of the brain network and map them into hash codes, and a separable convolutional layer is designed to extract multi-level individual topological features. Secondly, the adjacency matrix of the brain network is obtained based on the similarity relationship between phenotypic labels, and a group relationship graph is constructed together with individual features. Thirdly, a GCN-based deep hash learning model is used to extract group features of the brain network and map them into hash codes. Then, the two deep hash learning models learn from each other by minimizing the distribution difference between hash codes, thereby realizing the interaction between individual and group features. Finally, the diagnostic label of the new instance is predicted by calculating the similarity between hash codes.

2. The brain network classification method based on deep hash mutual learning according to claim 1 is characterized in that The following steps are involved: Step 1: Data preprocessing and brain functional network construction; To obtain rs-fMRI data, we first preprocessed it using the resting-state functional magnetic resonance processing tool, and then constructed the brain functional network; the specific method is as follows: Step 1.1: Preprocess the raw rs-fMRI data using DPARSF; (1) Delete the first 5 time points of the original rs-fMRI time series; (2) Use linear interpolation to eliminate the time phase difference caused by interlayer scanning and complete layer time correction; (3) Eliminate the data of subjects with horizontal head movement exceeding 2 mm and rotational head movement exceeding 2° to complete head motion correction; (4) Registration, smoothing, and filtering of rs-fMRI images based on T1 images; Step 1.2, construct brain network; (1) Locate the locations of brain regions in the cerebral cortex as nodes of the brain network based on brain maps; (2) Calculate the Pearson correlation coefficient of the inter-node rs-fMRI time series as the functional connectivity strength between brain regions, that is, the weight of the connection edge between brain region nodes; Step 2: brain network data division; The brain network dataset is divided into training set X according to the ratio of 8:1:

1. t , validation set X a and the test set X e ; Training set There are N brain networks in it, and the corresponding label set is represented by Y t , which contains the diagnostic label set and the phenotypic label set is d represents the diagnostic label, p represents the phenotypic label; similarly, the validation set There are M brain networks in it, and the corresponding label set is expressed as Test set There are K brain networks in it, and the corresponding label set is expressed as Step 3: Individual feature extraction based on deep hash learning; Adopting CNN-based deep hashing learning model I H The topological features of individual brain networks are extracted and mapped into hash codes. Specifically, a five-layer CNN is designed for feature extraction. The extraction process is defined as Then a hash layer is designed for feature mapping, and the mapping process is defined as Step 3.1, individual feature extraction; For feature extraction, a separable CNN is used to extract the topological features of individual brain networks from the edge level, node level, and graph level respectively; L0 is the input layer that receives the brain network; L1 is a separable edge-to-edge layer, which includes two processes, channel-by-channel E2E and point-by-point E2, for extracting edge-level topological features from the input brain network; L2 is a separable edge-to-node layer, which includes two processes: channel-by-channel E2N and point-by-point E2N, and is used to extract node-level features from edge-level features; L3 is the node-to-graph layer, which is used to extract graph-level features from node-level features; L4 is the fully connected layer; the specific process is as follows; (1) Define the feature extraction process of L0-L4 layer CNN as The individual features of the brain network IF are extracted during the training process t , the formula is: Among them, X t is the training set, which contains N brain networks; (2) Define the cross entropy loss function J1 acting on the L0-L4 layer as: Among them, W c and b represent the weights and biases of the separable CNN, respectively. and b q ∈b are the weights and biases of the Softmax layer, and N is the training set X t The number of brain networks included, It's X t The diagnostic label set, is the Softmax function, IF t is the individual characteristic of the brain network, ξ(IF t ) is the predicted probability of the diagnostic label of the individual feature; Step 3.2, individual feature mapping; For the feature mapping part, a hash layer is designed to first map the extracted individual features into a feature vector of length l through a hash function, where l is the predefined hash code length. The feature vector is then binarized to finally obtain the hash code representation corresponding to the individual topological features. The specific process is as follows; (1) Define the feature mapping process of the hash layer as The extracted individual features IF t After the hash layer is mapped to the hash code IU t , the formula is: in, and b h ∈b are the weight and bias of the hash layer, σ(·) is the sigmoid function, sgn(·) is the sign function, IU t ∈{0,1} N×l IF t The hash code represents, N represents X t The number of brain networks contained in the training set X is , l is the predetermined hash code length, and each hash code is an l-dimensional vector with a value of 0 or 1; t The hash code set IU t Also known as a hash code library; (2) Define the similarity preservation loss function J2 acting on the hash layer as: Where S∈{0,1} N×N It is a pairwise similarity matrix constructed based on diagnostic labels, N represents X t The number of brain networks included in t (i) With brain network X t (j) have the same diagnostic label, i.e. Then s ij =1, otherwise s ij =0;iu t (i) and iu t (j) are brain networks X t (i) and X t (j) Hash code representation of individual characteristics; Step 3.3, objective function construction; By integrating the loss function J1 of the feature extraction part and the loss function J2 of the feature mapping part, the deep hash learning model I based on CNN H The objective function is defined as: Among them, W c and b are model I H The weights and biases of ; the last one is the regularization term to avoid overfitting of the model during the training phase, λ1, λ2 and λ r is a hyperparameter; Step 4: Group feature extraction based on deep hash learning; Adopting GCN-based deep hash learning model G H Extract the topological features of the group brain network and map them into hash codes; specifically, first construct the adjacency matrix A based on the similarity of the phenotypic labels, and then combine the individual brain network features IF extracted from step 3 t Construct a group relationship graph P, and finally use GCN to extract the group features of the brain network and map them into hash codes through the hash layer. The GCN extraction process is defined as The mapping process is defined as Step 4.1, construct the adjacency matrix A of the group brain network; (1) Calculate the Mahalanobis distance between brain network phenotype label vectors. The calculation formula is as follows: Where D = {d ij |i,j=1,…,N} is the Mahalanobis distance, which can eliminate the scale differences between different phenotypic labels; Brain network x t (i) and x t (j) phenotypic label vector, C∈R N×N is the covariance matrix, N represents X t The number of midbrain networks, C -1 is the inverse of the covariance matrix; (2) Define the adjacency matrix of the group brain network: stα∈(0,1) Among them, A={a ij |i,j=1,…,N} is the adjacency matrix, a ij =1 indicates two brain networks x t (i) and x t (j) There is a connection between ij =0 means brain network x t (i) and x t (j) are not adjacent nodes, α is used to control the number of connecting edges between brain networks, D is the Mahalanobis distance, and max(D) is the maximum value of D; Step 4.2, constructing the group brain network relationship graph P; Taking a single brain network {x t (1),x t (2),…,x t (N)} is a node, A is an adjacency matrix, which is used to determine whether there is an edge between nodes. t For the initial attribute characteristics of the node, define the group relationship graph P(A,IF t ), where A is the adjacency matrix representing the graph structure, IF t is the feature matrix representing the node attributes; Step 4.3, group feature extraction; For feature extraction, GCN is used to extract the features of the group brain network. Specifically, L'0 is the input layer that receives the group relationship graph P; L'1-L'2 are graph convolution layers used to aggregate the features of adjacent nodes. The specific process is as follows: (1) Define the feature extraction process of L'0-L'2 layer GCN as The group feature GF of the brain network is extracted during the training process t , the formula is: Among them, X t is the training set, which contains N brain networks; (2) Define the cross entropy loss function J3 acting on the L'0-L'2 layer as: Among them, W g is the weight of GCN, N is X t The number of brain networks in For X t The diagnostic label set, is the Softmax function, GF t is the group characteristic of the brain network, The predicted probability of the diagnostic label for the group feature; Step 4.4, group feature mapping; For the feature mapping part, a hash layer is designed to map the extracted group features into a hash code representation of length l through a hash function; the specific process is as follows; (1) Define the feature mapping process of the hash layer as The extracted group features GF t After the hash layer is mapped to the hash code GU t , the formula is: Among them, σ(·) is the sigmoid function, sgn(·) is the sign function; GU t is the group characteristic GF t The hash code representation of , each hash code is an l-dimensional vector with a value of 0 or 1, where l is a predetermined hash code length; (2) Define the similarity preservation loss function J4 acting on the hash layer as: Where S∈{0,1} N×N is the pairwise similarity matrix constructed based on the diagnostic labels, s ij ∈S, N represents X t The number of brain networks included in t (i) and gu t (j) are brain networks X t (i) and X t (j) Hash code representation of group characteristics; Step 4.5, objective function construction; By integrating the loss function J3 of the feature extraction part and the loss function J4 of the feature mapping part, the deep hash learning model G based on GCN H The objective function is defined as: J(W g )=λ3J3+λ4J4 Among them, W g For model G H The weight of , λ3 and λ4 are hyperparameters; Step 5: Mutual learning based on hash codes; In order to make the two deep hash learning models I H and G H During the training process, they collaboratively learn and explore each other's learned features, and define the maximum average difference as the loss function of the mutual learning process: Among them, IU t Is an individual characteristic IF t The hash code of GU t is the group characteristic GF t The hash code representation of δ(·) is the Gaussian kernel function, and η is the standard deviation related parameter that controls the range of δ(·); therefore, the deep hash learning model I based on CNN H The objective function is updated as: Among them, W c and b are model I H The weights and biases of λ1,λ2,λ m and λ r is a hyperparameter; GCN-based deep hash learning model G H The objective function is updated as: J(W g )=λ3J3+λ4J4+λ' m J m (14) Among them, W g For model G H The weights of λ3, λ4 and λ' m is a hyperparameter; the two models are jointly optimized by learning from each other until the objective function converges; Step 6, classification based on hash code; After mutual learning steps, the trained I H After the model is built, use the validation set X a To verify the classification performance of the model and further adjust the model parameters; the specific process is as follows; (1) Obtain the validation set X a The hash code represents IU a , the calculation formula is as follows: in, For I H The feature extraction function of the model, For I H The model's feature mapping function; (2) The hash code IU of the training set t (Formula (3)) is used as the hash code library to calculate IU a Each hash code iu a (i)With IU t The distance between all hash codes in IU t Find the hash code with the smallest distance and use its label as the predicted label The calculation formula is expressed as: in, Represents the training set X t The true diagnostic label of N is the training set X t The number of brain networks included, M is X a The number of brain networks included in It is the XOR logical operation; (3) When the loss function of the DHML method reaches convergence and the prediction accuracy of the validation set no longer improves, the model and hash code library at this time are retained for testing the brain network X e Classification of diagnostic labels.

Images