CN116912576A - Self-adaptive graph convolution brain disease classification method based on brain network higher-order structure - Google Patents

Abstract

The invention discloses a brain network higher-order structure-based self-adaptive graph convolution brain disease classification method, belonging to the technical field of brain disease classification; the method solves the problems of neglecting higher-order structure information, high consumption of calculation resources and the like in the prior art, wherein the size of a convolution kernel depends on super-parameter selection, and comprises the following steps: acquiring an attempted image data set for preprocessing; constructing a higher-order brain function network and a corresponding NBS matrix, performing feature aggregation on brain regions with different neighbor numbers in the higher-order brain function network by adopting a self-adaptive mechanism, and updating feature characterization of the brain regions to obtain node embedding; designing a backbone network, embedding the self-adaptive graph convolution module into the backbone network, and carrying out dimension reduction and classification on the feature characterization by the obtained edge feature through a classifier module, so as to further reduce the learnable parameters and calculated amount of the model; the invention is applied to brain disease classification.

Description

Self-adaptive graph convolution brain disease classification method based on brain network higher-order structure

Technical Field

The invention belongs to the technical field of image processing, and relates to a self-adaptive graph convolution brain disease classification method based on a brain network high-order structure.

Background

With the development of magnetic resonance imaging technology, resting-state functional magnetic resonance imaging gradually develops into an effective method, which can capture minute abnormal brain changes related to brain diseases. The rs-fMRI obtains a functional connection network by calculating the correlation between a spatial distance region in the functional network and a cross-functional network, can reflect the functional tissue of the brain, and has the potential of becoming a clinical neuroscience biomarker.

The traditional deep learning model regards the brain function network as mesh data, ignoring its inherent topology, and therefore performs poorly on part of the classification task. In contrast, the atlas model is able to extract spatial features from the connected group neighborhood, thereby capturing more carefully the topological features of the brain network. However, they have problems of information guidance for neglecting higher order structures, selection of convolution kernel size dependent super parameters, high consumption of computing resources, and the like.

Disclosure of Invention

The invention provides a self-adaptive graph convolution brain disease classification method based on a brain network high-order structure, which aims to solve the problems that the prior art ignores high-order structure information, the convolution kernel size depends on super-parameter selection, the consumption of computing resources is high and the like.

The technical scheme adopted by the invention is that the self-adaptive graph convolution brain disease classification method based on the brain network higher-order structure comprises the following steps:

step S1: acquiring the data of the tried image, preprocessing the image data, and obtaining preprocessed image data;

step S2: dividing the preprocessed image data into areas, extracting the time sequence of each area, calculating the pearson correlation coefficient of the time sequence of each area, and constructing a tested brain function network;

step S3: binarizing the constructed brain function network, performing motif detection on the binarized brain function network, and constructing a higher-order matrix of the brain function network based on the motif to obtain a higher-order brain function network;

step S4: building an overall network structure, wherein the overall network structure comprises an adaptive graph rolling module, a classifier module and a softmax function;

step S5: after the overall network structure is built, a high-order brain function network is input, the model is trained and updated, and a self-adaptive graph convolution brain function network classification model based on the high-order network is obtained.

Further, in step S1, the specific steps of the preprocessing include: the first 10 time points are removed, time point corrections, head motion corrections, spatial normalization, smoothing, filtering and removing covariates.

Further, in the step S3, constructing a higher-order matrix of the brain function network specifically includes:

performing motif detection on a binarized brain function network, and obtaining a motif frequency spectrum by calculating the occurrence frequency of each type of three-node motif in the network;

comparing the obtained frequency spectrum with frequency spectrums in 1000 proxy random networks; wherein the random network is a zero model, and the same number of nodes and edges and the same degree distribution are reserved;

determining the frequency of the motif in the random network, and giving the overexpression of the motif M in G by the z-score:

wherein ,for the frequency of occurrence of the motif M in the real network, < >> and />For the mean and variance of its frequency of occurrence in 1000 proxy random networks, when Z _M >1.96, motif M is defined as statistically significant.

Constructing a higher-order matrix W according to the confirmed motif M _M The entry (i, j) is the number of co-occurrences of nodes i and j in motif M, (W) _M ) _ij Number of instances of M containing nodes i and j.

Further, in the step S5, a higher-order brain function network is input to train and update the model, which specifically includes:

initializing trainable parameters in a network, constructing an NBS matrix of a higher-order brain function network, and inputting the constructed NBS matrix into the network in batches; and constructing a loss function and calculating loss according to the predicted value and the real label, and using an optimization algorithm to perform back propagation to update parameters of the network until the loss is not reduced within a certain range, and saving the network parameters as a final model.

Further, constructing an NBS matrix of the higher-order brain function network specifically includes:

carrying out multiple comparison correction analysis on each side in the two groups of tested higher-order brain function networks to obtain a p value of each side; setting a p value, reserving edges smaller than the p value, and deleting edges larger than the p value.

Further, in the step S4, the adaptive graph rolling module includes:

the first Mask layer performs Mask processing on the input NBS difference matrix, and only retains neighbor information of the nodes, and the specific calculation process is as follows:

x _i ＝a _i ·FC _i

wherein ,represents the ith row of the p matrix, +.>Line i, representing functional connection, will x _i Compressed into Where k is the length of the node representation vector;

the second layer EdgeConv layer is used for carrying out edge convolution operation on each node and the neighbors of the nodes in the matrix; inputting [16, c,116], outputting [16,2c,116, k ], wherein c is the number of channels, k is the number of neighbors of each node, different nodes have different numbers of neighbors, and the edge convolution layer adaptively adjusts the k value according to the number of the neighbors of the different nodes; extracting the characteristics of the edges, carrying out 2-dimensional convolution on the edges, inputting [16,2c,116, k ], and outputting [16,64,116, k ]; carrying out ReLU operation on the edge characteristics; the specific calculation process is as follows:

wherein xi is the characteristic of the central node i, the global characteristic of the node is represented, and xj is the characteristic of the connected neighbor node j; the node value difference xj-xi represents the local characteristics of the node; h theta represents the trainable weight theta and h is a nonlinear activation function;

and the third layer is a MaxPooling layer, and the maximum pooling operation is carried out on the node characteristics after aggregation.

Further, in the step S4, the classifier module includes:

a first layer: one-dimensional convolution layer conv1d, in_channels=roi _s Out_channels=64, taking a 1×1 convolution kernel, and using BN;

a second layer: one-dimensional convolution layer Conv1d, in_channels=kernel [ -1], out_channels=32, with a 1×1 convolution kernel, and using BN;

and the third layer is a full-connection layer, in_channels=32×64, and out_channels=128, and a 1×1 convolution kernel is adopted.

Further, in the step S4, the overall network structure is classified by a Softmax function, and the class probability calculation formula is as follows:

where pi represents the probability that sample i is predicted to be a positive class.

The invention has the beneficial effects that: the invention discloses a brain network higher-order structure-based self-adaptive graph convolution brain disease classification method, which classifies by constructing a higher-order matrix to obtain good classification results. Compared with an optimal model, the method has the advantages of greatly reducing the quantity of parameters and calculated quantity, further reducing the calculation consumption and having potential edge deployment advantages.

Drawings

FIG. 1 is a flow chart of the adaptive graph convolution brain disease classification method based on the higher-order structure of the brain network of the present invention.

FIG. 2 is a schematic diagram of a higher order brain function network constructed in accordance with the present invention.

FIG. 3 is a block diagram of an adaptive graph convolution module constructed in accordance with the present invention.

FIG. 4 is a schematic diagram of an edge convolution layer constructed in accordance with the present invention.

FIG. 5 is a block diagram of a multi-layered perceptron classification module constructed in accordance with the present invention.

Fig. 6 is a general network configuration diagram according to the present invention.

Detailed Description

In the following, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

As shown in fig. 1-6, the invention provides a self-adaptive graph convolution brain disease classification model based on a brain function network high-order structure, which captures high-order topological features of the brain function network through a self-adaptive convolution module, simultaneously focuses on brain interval heterogeneity, and maps the extracted features onto a label by using a one-dimensional convolution layer, thereby realizing classification of the brain function network.

The specific network model of the present invention is shown in fig. 6, and the specific implementation steps are as follows:

step one, higher order network construction

Data preprocessing, acquisition of ADHD is attempted like a dataset, preprocessing the image data using the DPABI toolbox. Dividing the brain to be tested into regions by using an AAL brain map, extracting a time sequence of each region, calculating a pearson correlation coefficient for the time sequence of each region, and constructing a brain function network to be tested;

the method comprises the steps of constructing a high-order network, binarizing the constructed brain function network, arranging correlation coefficients of the brain function network from large to small in order to prevent false connection and existence of isolated nodes in the network, setting the first 30% value as 1, and setting the rest value as 0. Performing motif detection on the binary functional network, finding out the motif overexpressed in the functional network, and constructing a high-order structural matrix of the brain functional network based on the motif;

and constructing an NBS matrix, carrying out network-based statistical test on two groups of higher-order brain function networks, setting a p value, reserving a difference edge smaller than the p value, and deleting a difference edge larger than the p value.

Step two, network architecture design

The self-adaptive graph convolution module is designed, mask processing is carried out on the matrix, then edge convolution operation is adopted to fuse the local information and the global information which are originally embedded, the model depth is reduced, the learning parameters of the model are compressed, meanwhile, the problem of heterogeneity between brain regions is considered, the self-adaptive mechanism is adopted to carry out feature aggregation on brain regions with different neighbor numbers in the higher-order brain function network, and the feature characterization of the brain regions is updated to obtain node embedding.

An adaptive graph convolution module:

and the first Mask layer is used for carrying out Mask processing on the input NBS difference matrix and only retaining the neighbor information of the node. The specific calculation process is as follows:

x _i ＝a _i ·FC _i

wherein ,represents the ith row of the p matrix, +.>Representing the ith row of functional connections. To save computing resources and memory, x is set to _i Compression to->Where k is the length of the node representation vector.

And the second layer is an EdgeConv layer, and performs edge convolution operation on each node and the neighbors of the nodes in the matrix. The method comprises the steps of inputting [16, c,116], outputting [16,2c,116, k ], wherein c is the channel number, k is the neighbor number of each node, different nodes have different neighbor numbers, and the edge convolution layer can adaptively adjust the k value according to the neighbor node numbers of the different nodes. Features of the edges are extracted, the edges are convolved in 2 dimensions, input [16,2c,116, k ], output [16,64,116, k ]. And carrying out ReLU operation on the edge feature. The specific calculation process is as follows:

wherein xi is the characteristic of the central node i, the global characteristic of the node is represented, and xj is the characteristic of the connected neighbor node j. The node value difference xj-xi represents the local characteristics of the node. h theta denotes a trainable weight theta, a nonlinear activation function h, implemented by a multi-layer perceptron and a corrective linear unit (Rectified linear unit, reLU).

And the third layer is a MaxPooling layer, and finally, the maximum pooling operation is carried out on the node characteristics after aggregation.

And (3) designing a classifier module, namely performing dimension reduction and classification on the feature characterization by using the edge features obtained by the self-adaptive graph convolution module through the classifier module. Meanwhile, to prevent overfitting, an L2 regularization loss function is added to the classifier layer.

And a classifier module:

a first layer: one-dimensional convolution layer conv1d, in_channels=roi _s Out_channels=64, taking a 1×1 convolution kernel, and using BN;

a second layer: one-dimensional convolution layer Conv1d, in_channels=kernel [ -1], out_channels=32, with a 1×1 convolution kernel, and using BN;

the first two layers have the main functions of reducing the feature dimension and the calculated amount;

and the third layer is a full-connection layer, in_channels=32×64, and out_channels=128, and a 1×1 convolution kernel is adopted.

The overall network structure comprises an adaptive graph rolling module, a classifier module and a Softmax function, and is specifically as follows:

the first layer is a Mask layer, mask processing is carried out on the input NBS difference matrix, and only neighbor information of the nodes is reserved;

and the second layer is an EdgeConv layer, and performs edge convolution operation on each node and the neighbors of the nodes in the matrix. The method comprises the steps of inputting [16, c,116], outputting [16,2c,116, k ], wherein c is the channel number, k is the neighbor number of each node, different nodes have different neighbor numbers, and the edge convolution layer can adaptively adjust the k value according to the neighbor node numbers of the different nodes. Features of the edges are extracted, the edges are convolved in 2 dimensions, input [16,2c,116, k ], output [16,64,116, k ]. Carrying out ReLU operation on the edge characteristics;

the third layer is a MaxPooling layer, and finally, the maximum pooling operation is carried out on the node characteristics after aggregation;

fourth layer: one-dimensional convolution layers conv1d, in_channels=rois, out_channels=64, with 1×1 convolution kernels, and using BN;

fifth layer: one-dimensional convolution layer Conv1d, in_channels=kernel [ -1], out_channels=32, with a 1×1 convolution kernel, and using BN;

and the sixth layer is a full-connection layer, in_channels=32×64, and out_channels=128, and a 1×1 convolution kernel is adopted.

Finally, classification is performed by a Softmax function.

Step three, network structure parameter design

The preliminary experiments of the invention find that the classification accuracy of the network is not high when the traditional space diagram convolution extracts the characteristics of the neighbor information of the nodes, and the characteristics of each node cannot be effectively extracted due to the variable node size of the brain function network and the possibility that each node has different numbers of neighbors. Finally, the self-adaptive graph rolling module is determined through experiments, so that the high-order topological characteristic of the brain function network can be captured in a self-adaptive mode, the heterogeneity of brain regions is focused more, and excellent classification performance is achieved. In addition, through multiple experiments, the MLP layer is replaced by the one-dimensional convolution layer on the basis of ensuring accurate classification, so that the learnable parameters and calculated amount of the model are further reduced.

Step four, experimental process and result analysis

The network training is performed when the threshold value of the NBS matrix with the higher-order structure is set to be p=0.05. Model training is performed by using an Adam optimizer, and model weights are updated through an adaptive learning rate. When the verification accuracy stops increasing, a gradual learning rate decay is used, with a minimum learning rate of 1e-6. After each training phase, the model is evaluated on the verification dataset and the model parameters are saved only if a better verification accuracy is obtained. To avoid overfitting, L2 regularization was used, and different L2 values (0.1, 0.01, 0.001, and 0.0001) were used in the experiment, taking the highest accuracy in the range as the final performance of the model. The number of the tested articles placed in each batch in the training is 16, 20 rounds of training are performed, and the target loss function uses BCELoss.

The Adam optimizer formula is:

m _t ＝μ*m _t-1 +(1-μ)*g _t

wherein the first two formulas are a first moment estimate and a second moment estimate of the gradient, respectively; the latter two formulas are corrections to the first order second moment estimate. The last part of the formula is a dynamic constraint on the learning rate n, and has a definite range.

The BCELoss formula is:

where yi represents the label of sample i and pi represents the probability that sample i is predicted to be a positive class.

The L2 regularization formula is as follows:

where wi is a parameter of the classifier layer.

And finally, uniformly setting the batch size to be 16, and setting the super parameters according to the optimal parameters of the original model. The whole project is trained for 30 rounds, and an early stop method is adopted, so that the model with the lowest tie loss and highest accuracy on the test set is stored. For the cGCN model, only the RNN layer at the end of the model is removed. For the SVM and the logistic regression model, the iteration times are uniformly set to 1000 times, and other super parameters remain default. As shown in table 1, the network of the present invention has the highest accuracy, and has great advantages in accuracy and the number of parameters compared to other models.

TABLE 1 comparison of accuracy, sensitivity, specificity, AUC, parameter quantity and calculation of the brain network classification task using the network model of the present invention with the existing network

The self-adaptive graph rolling module provided by the invention focuses on heterogeneity of different brain regions in neural activity, effectively captures high-order topological features of a brain function network, and realizes better classification results. The self-adaptive graph convolution essentially endows different brain regions with different numbers of neighbors, namely, provides self-adaptive convolution neighborhood for different brain regions, and avoids the defect that all brain regions are endowed with convolution neighborhood with the same size in the prior work. Compared with the existing model, the self-adaptive graph convolution model provided by the invention can pay attention to brain section heterogeneity, and has a first-class accuracy rate on ADHD disease classification results.

The classifier module provided by the invention replaces the MLP layer with one-dimensional convolution, so that the learnable parameters and calculated amount of the model are further reduced. The experimental result shows that the parameter compression version of the model is slightly higher than the BrainNetCNN in parameter quantity and calculated quantity, but the parameter quantity and calculated quantity are reduced by nearly one order of magnitude compared with the latest work cGCN, and the classification accuracy is better than the latest work.

The foregoing description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims (8)

1. The self-adaptive graph convolution brain disease classification method based on the brain network higher-order structure is characterized by comprising the following steps of:

step S1: acquiring the data of the tried image, preprocessing the image data, and obtaining preprocessed image data;

step S2: dividing the preprocessed image data into areas, extracting the time sequence of each area, calculating the pearson correlation coefficient of the time sequence of each area, and constructing a tested brain function network;

step S3: binarizing the constructed brain function network, performing motif detection on the binarized brain function network, and constructing a higher-order matrix of the brain function network based on the motif to obtain a higher-order brain function network;

step S4: building an overall network structure, wherein the overall network structure comprises an adaptive graph rolling module, a classifier module and a softmax function;

step S5: after the overall network structure is built, a high-order brain function network is input, the model is trained and updated, and a self-adaptive graph convolution brain function network classification model based on the high-order network is obtained.

2. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 1, wherein in step S1, the specific step of preprocessing includes: the first 10 time points are removed, time point corrections, head motion corrections, spatial normalization, smoothing, filtering and removing covariates.

3. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 1, wherein in the step S3, constructing a higher-order matrix of a brain function network specifically includes:

performing motif detection on a binarized brain function network, and obtaining a motif frequency spectrum by calculating the occurrence frequency of each type of three-node motif in the network;

comparing the obtained frequency spectrum with frequency spectrums in 1000 proxy random networks; wherein the random network is a zero model, and the same number of nodes and edges and the same degree distribution are reserved;

determining the frequency of the motif in the random network, and giving the overexpression of the motif M in G by the z-score:

wherein ,for the frequency of occurrence of the motif M in the real network, < >> and std/>For the mean and variance of its frequency of occurrence in 1000 proxy random networks, when Z _M >1.96, motif M is defined as statistically significant.

Constructing a higher-order matrix W according to the confirmed motif M _M The entry (i, j) is the number of co-occurrences of nodes i and j in motif M, (W) _M ) _ij Number of instances of M containing nodes i and j.

4. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 1, wherein in the step S5, a higher-order brain function network is input, and the model is updated in training, specifically comprising:

initializing trainable parameters in a network, constructing an NBS matrix of a higher-order brain function network, and inputting the constructed NBS matrix into the network in batches; and constructing a loss function and calculating loss according to the predicted value and the real label, and using an optimization algorithm to perform back propagation to update parameters of the network until the loss is not reduced within a certain range, and saving the network parameters as a final model.

5. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 4, wherein constructing an NBS matrix of a higher-order brain function network specifically comprises:

carrying out multiple comparison correction analysis on each side in the two groups of tested higher-order brain function networks to obtain a p value of each side; setting a p value, reserving edges smaller than the p value, and deleting edges larger than the p value.

6. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 4, wherein in the step S4, the adaptive graph convolution module includes:

the first Mask layer performs Mask processing on the input NBS difference matrix, and only retains neighbor information of the nodes, and the specific calculation process is as follows:

x _i ＝a _i ·FC _i

wherein ,represents the ith row of the p matrix, +.>Line i, representing functional connection, will x _i Compression to-> Where k is the length of the node representation vector;

the second layer EdgeConv layer is used for carrying out edge convolution operation on each node and the neighbors of the nodes in the matrix; inputting [16, c,116], outputting [16,2c,116, k ], wherein c is the number of channels, k is the number of neighbors of each node, different nodes have different numbers of neighbors, and the edge convolution layer adaptively adjusts the k value according to the number of the neighbors of the different nodes; extracting the characteristics of the edges, carrying out 2-dimensional convolution on the edges, inputting [16,2c,116, k ], and outputting [16,64,116, k ]; carrying out ReLU operation on the edge characteristics; the specific calculation process is as follows:

x′ _i ＝max _j:(i,j)∈ε h _θ (x _i ||x _j -x _i )

x′ _i ＝max _{j＝(i,j)∈ε} (ReLU((x _i ||x _j -x _i )))

wherein xi is the characteristic of the central node i, the global characteristic of the node is represented, and xj is the characteristic of the connected neighbor node j; the node value difference xj-xi represents the local characteristics of the node; h theta represents the trainable weight theta and h is a nonlinear activation function;

and the third layer is a MaxPooling layer, and the maximum pooling operation is carried out on the node characteristics after aggregation.

7. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 6, wherein in the step S4, the classifier module includes:

a first layer: one-dimensional convolution layer conv1d, in_channels=roi _s Out_channels=64, taking a 1×1 convolution kernel, and using BN;

a second layer: one-dimensional convolution layer Conv1d, in_channels=kernel [ -1], out_channels=32, with a 1×1 convolution kernel, and using BN;

and the third layer is a full-connection layer, in_channels=32×64, and out_channels=128, and a 1×1 convolution kernel is adopted.

8. The brain network higher-order structure-based adaptive graph convolution brain disease classification method according to claim 7, wherein in the step S4, the overall network structure is classified by a Softmax function, and the class probability calculation formula is as follows:

where pi represents the probability that sample i is predicted to be a positive class.