CN107967686B - An epilepsy recognition device combining dynamic brain network and long-term memory network - Google Patents

Abstract

The invention discloses an epilepsy recognition device combining a dynamic brain network and a long-short-term memory network, and belongs to the technical field of biomedical image pattern recognition. The present invention firstly calculates the brain network and dynamic brain network of functional magnetic resonance imaging of epilepsy patients and normal control groups. Then, for the brain network, the F-score feature selection algorithm was used to calculate the F value of each functional connection in the brain network, and the features of the dynamic brain network corresponding to the F value greater than 0.06 were taken out as the input of the long and short-term memory network. Next, a long-term and short-term memory network structure is established. The input is the selected dynamic brain network feature, and the output is the sample label. The label of epilepsy patients is 1, and the label of normal people is 0. Finally, the stochastic gradient descent algorithm is used to optimize the parameters in the network, and after continuous training, the identification task between epilepsy patients and normal people is finally completed. The invention combines the dynamic brain network and the long-term memory network for the first time to carry out the task of auxiliary diagnosis of epilepsy.

Description

An epilepsy recognition device combining dynamic brain network and long-term memory network

technical field

The invention belongs to the technical field of biomedical image pattern recognition, in particular to the construction of a dynamic brain network pattern recognition framework for functional magnetic resonance images.

Background technique

Epilepsy, commonly known as "horn wind", is a common and dangerous neurological disease. According to the statistics of the World Health Organization, there are about 50 million people with epilepsy, and 400,000 new people with epilepsy every year. Previous studies have found that there are many reasons for the disease, such as congenital factors, brain lesions, systemic or systemic diseases. In addition, the incidence of epilepsy is also age-related.

At present, for some epilepsy diseases, such as idiopathic tonic-clonic epilepsy (GTCS), in routine examination, the brain structure of epilepsy patients does not have significant organic lesions, and some other physiological and metabolic functions of patients are also different from those of ordinary people. In the same way, the correct pathogenesis and accurate lesions have not been found so far. Therefore, improving the level of diagnosis and finding out the focal area is the guarantee for the treatment and prediction of epileptic seizures.

Functional magnetic resonance imaging (fMRI) is a new non-invasive imaging technology developed in recent years to study brain function, and it is a powerful tool to study human brain cognitive thinking activities. The application of fMRI can study the pathogenesis of the overall development of epilepsy, which has good clinical practicability. Moreover, studying fMRI images through pattern recognition algorithms can not only diagnose diseases with high accuracy at the individual level, but also better locate epilepsy foci.

SUMMARY OF THE INVENTION

By analyzing the brain images, the present invention overcomes the problem that the brain images of epilepsy patients cannot be well identified in the prior art, and designs a device for identifying whether the patients suffer from epilepsy.

The technical solution of the present invention is an epilepsy recognition device combining a dynamic brain network and a long-short-term memory network, the device includes a data input interface, a data storage, and a data processor, and the data on the data storage can be processed and executed to achieve the following steps;

Step 1: Step Brain Network Computation:

Step 11: Obtain the resting-state functional magnetic resonance signals of the patient's brain at M time nodes, and perform preprocessing including time correction, head movement correction, and registration on the functional magnetic resonance signals; The fMRI data are processed as follows;

Step 12: According to the template of 246 brain regions, calculate the average signal of each brain region, and get 246 average signals;

Step 13: Calculate the Pearson correlation coefficient between the average signals of each brain area, and get a matrix composed of 246*246 Pearson correlation coefficients, which is a brain network;

Step 2: Calculate the dynamic brain network;

Arrange the brain networks of the M time nodes obtained in step 1 in order, set the sliding window length K and step size of the time nodes, and calculate the Pearson correlation coefficient matrix of the K brain networks in the window each time during the sliding window process. After the window is completed, H 246*246 Pearson correlation coefficient matrices are obtained, and these Pearson correlation coefficient matrices are dynamic brain networks;

Step 3: Feature extraction on the brain network;

The feature extraction matrix is used to extract the features of the dynamic brain network, and H 246*246 feature matrices are obtained. The feature extraction matrix extracts the data of a specific position in the Pearson correlation coefficient matrix obtained in step 2, and the remaining position data is set to 0, set Non-zero contains E non-zero data, and the non-zero data is taken out to form a characteristic matrix of H*E;

Step 4: Input the H 246*246 feature matrices extracted in step 3 into the trained long-term memory network, and judge the patient's condition according to the output of the long-term memory network.

Further, in the step 4, select the first R rows of data of the H*E feature matrix obtained in step 3, and then slide one row each time, extract the next R row data each time, and after the sliding ends, extract all the data. The data is used as the input of the long and short-term memory network, and R is determined according to the actual situation.

Further, the method for feature extraction matrix in the step 2 is:

Step 2.1: Obtain all sample data in the data storage, including normal samples and epilepsy samples, use the methods of Step 1 and Step 2 to obtain the brain networks of all samples at a time node, and obtain Q sample brain networks, where Q is the total number of samples;

Step 2.2: Extract the data of the first position (1, 1) in all sample brain networks to form a functional connection vector x;

Step 2.3: Calculate the F value of the first position using the following formula;

表示特征向量x的所有样本的均值；

表示特征向量中，属于正常样本子集的均值；

表示特征向量中，属于癫痫样本子集的均值；n₀表示样本中正常样本的个数；n₁表示样本中癫痫样本的个数；

表示正常样本子集中第i个样本的功能连接值；

表示癫痫样本子集中第i个样本的功能连接值；in,

Represents the mean of all samples of the feature vector x;

Represents the mean value of the normal sample subset in the feature vector;

Represents the mean of the epilepsy sample subset in the feature vector; n ₀ represents the number of normal samples in the sample; n ₁ represents the number of epilepsy samples in the sample;

represents the functional connectivity value of the ith sample in the subset of normal samples;

represents the functional connectivity value of the ith sample in the epilepsy sample subset;

Step 2.4: Use the method of step 2.3 to calculate the F value of all other positions, and obtain an F value matrix with a size of 246*246;

Step 2.5: Set the feature extraction threshold, set each position element in the F-value matrix greater than the threshold to 1, and set the element less than or equal to the threshold to 0, and finally obtain the feature extraction matrix.

Further, the step size of the long and short-term memory network in the 4 is 3, the number of layers is 8, the size of the input layer is 1*E, and the output layer is 1 value; The feature matrix of all samples is used to input the feature matrix of each sample into the long and short-term memory network, and the stochastic gradient descent algorithm is used for training, and the parameters in the long and short-term memory network are updated. Finally, the parameters in the long and short-term memory network are basically stable. time memory network.

The invention combines the method of dynamic functional connection and long-short-term memory network, uses the long-short-term memory network to simulate the dynamic characteristics of the brain network of epilepsy patients, and provides an effective epilepsy diagnosis method based on functional magnetic resonance imaging.

Description of drawings

Fig. 1 is the flow chart of the present invention;

2 is a structural diagram of a long-short-term memory network (LSTM) of the present invention;

3 is a graph showing the relationship between different iteration times and classification accuracy rates of the present invention;

Figure 4 shows the number of iterations, training loss values and test loss values.

Detailed ways

An epilepsy recognition method combining dynamic brain network and long-term memory network, including the following steps:

A. Calculation of brain network:

Step A1: Scan the resting-state fMRI signals of epilepsy patients and normal people; perform preprocessing on the fMRI signals, wherein the preprocessing includes time correction, head movement correction, and registration.

Step A2: For the processed fMRI data, according to the template of 246 brain regions, calculate the average signal of each brain region, and 246 average signals will be obtained.

Step A3: Calculate the Pearson correlation coefficient between the average signals of each brain region, and a matrix composed of 246*246 Pearson correlation coefficients will be obtained. This matrix is the brain network.

B. Computation of Dynamic Brain Networks:

Step B1: Scan the resting-state fMRI signals of epilepsy patients and normal people. The functional magnetic resonance signals are preprocessed, wherein the preprocessing includes time correction, head movement correction, and registration.

Step B2: Using the processed fMRI data, according to the template of 246 brain regions, calculate the average signal of each brain region, and 246 average signals will be obtained.

Step B3: Calculate the Pearson correlation coefficient between the average signals of each brain region using the sliding window method. The preprocessed data of each person has 245 time points. During the calculation, each sliding window is set to a window width of 50 time points, and the overlap of two adjacent sliding windows is 90%. As a result, each person will calculate a matrix of 40 246*246 Pearson correlation coefficients.

C.F-score performs feature selection on brain networks:

Step C1: The training sample contains 100 subjects (50 epilepsy patients and 50 normal people). Each subject will get a total of 100 246*246 two-dimensional brain network matrices and A one-dimensional label vector y of size 100*1, where 1 represents epilepsy patients and 0 represents normal subjects.

Step C2: Take out the functional connectivity value of the first position in the subject's brain network matrix, and the coordinates of the first position are (1, 1). The functional connection vector x of all subjects in the first position will be obtained, and its size is 100*1. For the functional connection vector x and label vector y of all subjects in the first position, the F value of the first position is calculated according to the F-score formula. The F-score formula is as follows:

表示特征向量x的所有样本的均值；

表示特征向量中，属于正常被试样本子集的均值；

表示特征向量中，属于癫痫患者样本子集的均值；n0表示样本中正常被试的个数；n₁表示样本中癫痫患者的个数；

表示正常被试子集中第i个样本的功能连接值；

表示癫痫患者子集中第i个样本的功能连接值；in,

Represents the mean of all samples of the feature vector x;

Represents the mean value of the normal sample subset in the feature vector;

Represents the mean value of the epilepsy patient sample subset in the feature vector; n0 represents the number of normal subjects in the sample; n1 represents the number _of epilepsy patients in the sample;

represents the functional connectivity value of the ith sample in the normal subject subset;

represents the functional connectivity value of the ith sample in the subset of epilepsy patients;

Step C3: Repeat step C2 for each position of the brain network matrix, and an F-value matrix with a size of 246*246 will be obtained.

Step C4: Extract the position information W corresponding to the F value greater than 0.06 in the F value matrix. The data used in this patent, greater than 0.06, contains 469 functional connections. According to the position information W, the connection value in the three-dimensional dynamic brain network matrix is extracted, and a matrix of size 40*469 is obtained, where 40 represents the number of sliding windows of the dynamic brain network. For each subject, after steps C1, C2, and C3, a total of 100 dynamic feature matrices of 40*469 will be obtained.

D. Long and short-term memory network training:

Step D1: According to Step C4, each subject gets a dynamic feature matrix of 40*469. For each subject, the first three rows of the dynamic feature matrix are taken out as the input features of the long-short-term memory network, the size of which is 3*469. Then slide one row in turn and take out three rows until every row in the 40*469 dynamic feature matrix has been taken. Each subject will get 38 long and short-term memory network input features of size 3*469. The training set contains 100 subjects, and a total of 3800 long and short-term memory network input features of size 3*469 will be obtained.

Step D2: For the long-short-term memory network model, the step size is 3 and the number of layers is 8. The input layer size is 1*469. The output layer is 1 value representing the label.

Step D3: Treat 3800 long-short-term memory network input features with a size of 3*469 as 3800 new sample data. If the new sample data is from epilepsy patients, the label is 1; if it is from normal subjects, the label is 0. Finally, a label vector of size 3800*1 is obtained.

Step D4: Input 3800 new sample data with a size of 3*469 and the corresponding labels into the long-term memory network in turn, use the stochastic gradient descent algorithm for training, and update the parameters in the long-term memory network.

Step D5: Repeat step D4 continuously, and finally the parameters in the long and short-term memory network are basically stable.

E. Long and short-term memory network test:

Step E1: According to all the steps in A and B, calculate 100 test samples with a size of 40*246*246 three-dimensional dynamic brain network matrix.

Step E2: According to the position information W in Step C4, each of the 100 test samples will obtain a dynamic feature matrix with a size of 40*469.

Step E3: According to Step D1, each test subject will get 38 long-term memory network input features with a size of 3*469. Inputting 38 test long-short-term memory network input features into the previously trained model will yield 38 label predictions. Among the 38 label prediction values, if the number of 1s is greater than the number of 0s, the test subject predicts the label to be 1; if the number of 1s is less than or equal to the number of 0s, the subject predicts the label to be 0 .

Step E4: Step E3 is repeated for each test subject, and the predicted label of each subject will be obtained, and then check whether it matches the standard label. Finally, the classification accuracy rate is obtained, and the classification accuracy rate of the data used in this patent can reach 80%.

Claims (4)

1. an epilepsy recognition device of a joint dynamic brain network and a long-term memory network, the device comprises a data input interface, a data storage, a data processor, and the data on the data storage is processed to implement the following steps; Step 1: Step Brain Network Computation: Step 11: Obtain the resting-state functional magnetic resonance signals of the patient's brain at M time nodes, and perform preprocessing including time correction, head movement correction, and registration on the functional magnetic resonance signals; The fMRI data are processed as follows; Step 12: According to the template of 246 brain regions, calculate the average signal of each brain region, and get 246 average signals; Step 13: Calculate the Pearson correlation coefficient between the average signals of each brain area, and get a matrix composed of 246*246 Pearson correlation coefficients, which is a brain network; Step 2: Calculate the dynamic brain network; Arrange the brain networks of the M time nodes obtained in step 1 in order, set the sliding window length K and step size of the time nodes, and calculate the Pearson correlation coefficient matrix of the K brain networks in the window each time during the sliding window process. After the window is completed, H 246*246 Pearson correlation coefficient matrices are obtained, and these Pearson correlation coefficient matrices are dynamic brain networks; Step 3: Feature extraction on the brain network; The feature extraction matrix is used to extract the features of the dynamic brain network, and H 246*246 feature matrices are obtained. The feature extraction matrix extracts the data of a specific position in the Pearson correlation coefficient matrix obtained in step 2, and the remaining position data is set to 0, set Non-zero contains E non-zero data, and the non-zero data is taken out to form a characteristic matrix of H*E; Step 4: Input the H 246*246 feature matrices extracted in step 3 into the trained long-term memory network, and judge the patient's condition according to the output of the long-term memory network. 2. the epilepsy recognition device of a kind of joint dynamic brain network and long-term memory network as claimed in claim 1, it is characterized in that the data on the described data storage is processed by step 4 and selects the H*E that step 3 obtains. The first R rows of data in the feature matrix, and then slide one row each time, and extract the next R row data each time. After the sliding ends, use all the extracted data as the input of the long-term memory network, and R is determined according to the actual situation. 3. the epilepsy recognition device of a kind of joint dynamic brain network and long-term memory network as claimed in claim 1, it is characterized in that the data on described data storage is processed the method for feature extraction matrix in described step 2 is: Step 2.1: Obtain all sample data in the data storage, including normal samples and epilepsy samples, use the methods of Step 1 and Step 2 to obtain the brain networks of all samples at a time node, and obtain Q sample brain networks, where Q is the total number of samples; Step 2.2: Extract the data of the first position (1,1) in all sample brain networks to form a functional connection vector x; Step 2.3: Calculate the F value of the first position using the following formula;

in,

Represents the mean of all samples of the feature vector x;

Represents the mean value of the normal sample subset in the feature vector;

Represents the mean of the epilepsy sample subset in the feature vector; n ₀ represents the number of normal samples in the sample; n ₁ represents the number of epilepsy samples in the sample;

represents the functional connectivity value of the ith sample in the subset of normal samples;

represents the functional connectivity value of the ith sample in the epilepsy sample subset; Step 2.4: Use the method of step 2.3 to calculate the F value of all other positions, and obtain an F value matrix with a size of 246*246; Step 2.5: Set the feature extraction threshold, set each position element in the F-value matrix greater than the threshold to 1, and set the element less than or equal to the threshold to 0, and finally obtain the feature extraction matrix. 4. the epilepsy identification device of a kind of joint dynamic brain network and long-term memory network as claimed in claim 1, is characterized in that the data on described data storage is processed by described step 4, the step size of long-term memory network is 3. The number of layers is 8, the size of the input layer is 1*E, and the output layer is 1 value; adopt the methods of steps 1, 2, and 3 to calculate the feature matrix of all samples in the memory, and use the feature matrix of each sample to input the length of time. The memory network is trained using the stochastic gradient descent algorithm, and the parameters in the long and short-term memory network are updated. Finally, the parameters in the long and short-term memory network are basically stable, and the trained long and short-term memory network is obtained.

Images