阅读说明：本技术 基于时空注意力机制的精神分裂症患者甄别方法 (Schizophrenia patient discrimination method based on space-time attention mechanism ) 是由 李平 王标 余兰兰 黄罗杰 黄金诚 唐国根 黄睿 于 2021-08-25 设计创作，主要内容包括：本发明公开了一种基于时空注意力机制的精神分裂症患者甄别方法,属于智慧医疗领域,该甄别方法运用了结合注意力机制的动态脑功能网络分析方法对每个观察对象的fMRI信号数据进行处理分析；首先,基于fMRI信号值,通过网络建模的方法构建出动态脑功能网络；其次,针对现有的动态脑功能网络分析方法往往忽略了不同时间和脑区对于动态脑功能网络状态的不同作用与影响,构建结合时空注意力机制的卷积神经网络模型；再通过模型训练的方式,甄别出精神分裂症患者,与此同时得到不同脑区不同时间窗的重要性,并捕获正常人群和精神分裂症患者人群的各个时间点各个脑区的时变特性。本发明方法相比于现有的模型方法,具有高的召回率。(The invention discloses a schizophrenic patient discrimination method based on a space-time attention mechanism, which belongs to the field of intelligent medical treatment and is used for processing and analyzing fMRI signal data of each observed object by using a dynamic brain function network analysis method combined with the attention mechanism; firstly, constructing a dynamic brain function network by a network modeling method based on an fMRI signal value; secondly, aiming at the fact that the existing dynamic brain function network analysis method usually ignores different functions and influences of different time and brain areas on the state of the dynamic brain function network, a convolutional neural network model combining a space-time attention mechanism is constructed; and screening the schizophrenia patients in a model training mode, obtaining the importance of different time windows of different brain areas, and capturing the time-varying characteristics of the brain areas at various time points of normal people and the schizophrenia patient groups. Compared with the existing model method, the method has high recall rate.)

1. A schizophrenia patient screening method based on a space-time attention mechanism is characterized by comprising the following steps:

step 1: acquiring a sequence signal value of each brain area of each observation object, wherein the sequence length is T;

step 2: acquiring time window characteristics of each brain area at T moments by adopting a sliding window mode, calculating the correlation of each brain area at each moment by utilizing a Pearson correlation coefficient, and constructing a dynamic brain function network with a time scale of T;

and step 3: performing principal component analysis dimensionality reduction operation on the acquired n x n dimensional brain function network at the T moments to reduce the obtained n x n dimensional brain function network to n x d dimensional brain function network; then, the vector is spread to T vectors with 1 x (n x d) dimensions, and the vector is reconstructed into T again by means of a sliding window_w1 (T/T)_wN d) dimensional vector; wherein, T_wFor the number of time windows, each time window contains T/T_wA brain function network, wherein n represents the number of brain areas, and d represents the characteristic dimension of the brain areas;

and 4, step 4: introducing a space-time attention mechanism, and learning the weight of each time window and each brain area in a self-adaptive manner according to the respective characteristics of the time window and the brain area; by means of a space-time attention mechanism, will T_wBrain function network aggregation within each time window;

and 5: building a convolutional neural network model; adopting convolution kernel of 1 x (n x d) to carry out convolution operation on the dynamic brain function network characteristics, and adopting convolution kernel of 1 x 1 to bear pooling action to capture time characteristics of the convolution kernel;

step 6: sending the features after the convolution pooling into a full-connection layer for dimensionality reduction;

and 7: and finishing classification by adopting a classifier.

2. The method for schizophrenic patient screening based on spatiotemporal attention mechanism according to claim 1, wherein in step 4, the weight of each time window is calculated as follows:

1) flattening each network snapshot feature in the time window to a vector with dimensions of 1 x (n x d);

2) mapping each flattened network feature to a space Q for dimensionality reduction;

3) after the dimensionality reduction is finished, splicing a plurality of features in the same time window together to serve as the final time window feature;

4) by means of a weight vectorAfter each high-dimensional time window characteristic is changed into a scalar quantity, an activation function is used for activation operation, and the weight value alpha of each time window is obtained through calculation by means of a SoftMax function_i(ii) a The specific calculation is shown as the following formula;

wherein the content of the first and second substances,is a characteristic of each network after tiling within the jth time window,is the feature of each network after being flattened in the ith time window, | | is the splice symbol, and σ is an activation function.

3. The method for discriminating schizophrenia patients based on spatiotemporal attention mechanism as set forth in claim 2, wherein the weight of each brain region is calculated as follows in step 4:

1) mapping the characteristics of each brain area in each time window into a space U;

2) splicing the characteristics of all the moments of each brain area together to serve as the characteristics of each brain area on the whole time sequence;

3) by means of a weight vectorMapping each high-dimensional brain region feature into a scalar, and calculating by means of a SoftMax function to obtain a weight value of each brain regionThe specific calculation is shown as the following formula;

wherein the content of the first and second substances,features of the brain region o at 1 to T moments, respectively;are characteristic of the brain region q at 1 to T moments, respectively.

4. The method for screening schizophrenia patient based on spatiotemporal attention mechanism as set forth in claim 2, wherein the activation function σ in step 4 is an activation function LeakyReLU.

5. The method for schizophrenia patient screening based on spatiotemporal attention mechanism as set forth in claim 3, wherein in step 5, the objective loss function of the convolutional neural network model in combination with attention mechanism isL₁Cross entropy loss function, and introduce F norm of matrix to prevent model training overfitting, after introducing, the final loss function L of the model is:

wherein, V_LC represents the number of categories of the observed objects; y is_v，kIs an indicator function, if the observed object v belongs to class k, then y_v，k1, otherwise 0; p is a radical of_v，kIs the probability that the observed object v belongs to the class k, λ is the regular term coefficient, and W is the parameter that needs training and learning in the model.

6. The method for schizophrenia patient screening based on spatiotemporal attention mechanism as claimed in claim 3, wherein the convolutional neural network module is replaced with a recurrent neural network module in the step 5.

7. The method for schizophrenia patient screening based on spatiotemporal attention mechanism as set forth in claim 1, wherein the classifier employs a SoftMax function in step 7.

Technical Field

The invention belongs to the field of intelligent medical treatment, and particularly relates to a method for discriminating schizophrenia patients based on a space-time attention mechanism.

Background

Schizophrenia, a disorder of mental illness, seriously affects the cognitive function and mental health of patients. At present, the treatment of the patients with schizophrenia brings huge economic burden to a plurality of patients, also costs huge manpower in society, and reduces the life expectancy of the patients with schizophrenia by 10-25 years. And a plurality of scientific research workers analyze the fMRI signal value of each observed object by means of a dynamic network analysis method, and develop pathogenesis research while screening schizophrenia patients.

The dynamic brain function network analysis method for the dynamic brain function network may be classified into a dynamic brain function network analysis method based on statistical learning and a dynamic brain function network analysis method based on presentation learning. The dynamic brain function network analysis method based on representation learning is usually based on a mode of building a neural network model, aims at tasks (such as screening brain disease patients), and captures the characteristics of the dynamic brain function network when training the model. Dynamic brain function network analysis methods based on representation learning often analyze the differences between normal populations and corresponding populations of brain disease patients based on various features obtained from model training. Compared with a statistical method for capturing characteristics biased to a user-defined mode or a fixed mode, the learning-based dynamic brain function network analysis method can better adapt to different application scenes, and extracts corresponding time-varying characteristics of different dynamic brain function networks.

The learning-based dynamic brain function network analysis method can be divided into two subclasses, and the first method is to construct a dynamic brain function network and learn the characteristics of the dynamic brain function network by using a neural network framework. In another method, the connection of the dynamic brain function network is also obtained by building a neural network model and training.

In the first category of methods, Kam et al propose a new convolutional neural network model. They first construct a plurality of brain function connection networks by considering static and dynamic brain function connections, then decompose the constructed networks into a plurality of groups of a plurality of static brain function networks using an improved dimension reduction method, and quantify the dynamics of the brain function networks using voxel variances in the dynamic brain function networks. And finally, by sequentially covering the brain areas and inputting the brain areas into the multi-channel convolutional neural network model, the attention of each brain area is obtained while model training is completed.

Fan et al propose an end-to-end deep learning model that combines convolutional neural networks and Long-Short term Memory networks (LSTM) to simultaneously capture the spatiotemporal features of the brain Network's functional connection sequences. The model uses convolution layers with different convolution kernel sizes to learn the features of the dynamic brain function network under different scales, but the features captured by the model are more standing on the macro structure of network topology connection, and the space-time characteristics of the dynamic brain function network are not considered from the microstructure of time and brain areas.

Suk et al are different from the first method and belong to the second method. They first designed a Deep Auto-Encoder (Deep Auto-Encoder) to discover the hierarchical nonlinear functional relationship between regions, thereby converting the region features into an embedding space based on a complex functional network. Based on a given functional feature embedding, they use a Hidden Markov Model (HMM) to estimate the features of the dynamic brain functional network from internal states. By establishing a generation model with an HMM, the state of the dynamic brain function network of each observed object is obtained, and the label of the observed object is determined.

Jie et al also obtained the connectivity of the dynamic brain function network while diagnosing brain diseases by means of deep model training. They first defined a new weighted correlation kernel (wc-kernel) to measure the correlation between brain regions, by which the contribution of different time points is characterized by learning weighting factors in a data-driven manner. Specifically, a convolutional layer is first defined to construct a dynamic brain function network using wc-kernel. Then, another three convolutional layers are defined, and local (brain area-specific), global (brain network-specific) and temporal features are sequentially extracted from the constructed dynamic brain function network to diagnose the brain diseases.

Similarly, Azavedo et al learn the spatio-Temporal features of the brain in an end-to-end manner by combining the Graph Neural Network (GNN) method with the Temporal Convolutional Network (TCN) method. Specifically, the relationship between brain regions is modeled by using GNN, the relationship between time sequences is modeled by using TCN, and finally, the connection of dynamic brain function networks and the relationship between time points are obtained through model training.

Huang et al propose a Hierarchical Representation Learning method (Hierarchical reconstruction Learning) of a neural network framework based on graph convolution, and their network modeling objects are different from the above-mentioned methods. Huang et al consider each network snapshot in the dynamic brain function network as a node, the characteristic of the node is the topological connection condition of each network snapshot, and the connection edges between the nodes are learned by the characteristic of the node. And then, a plurality of connected graph convolution pooling layers are utilized to learn the dynamic brain function network state characteristics under different layers, and the dynamic brain function network state characteristics are input into a classifier to distinguish the schizophrenia patient.

The two methods do not simultaneously consider different influences of the importance of time and brain areas in the dynamic brain function network on the state of the dynamic brain function network, and the second method considers the modeling of the dynamic brain function network connection, so that although the connection of the dynamic brain function network can be solved scientifically, the time complexity of the algorithm is increased to a great extent.

The existing dynamic brain function network analysis method based on representation learning usually ignores different effects and influences of different time and brain areas on the state of the dynamic brain function network; meanwhile, in a part of methods which need to learn the topological connection of the dynamic brain function network, although the network modeling mode is more scientific and reasonable, the complexity of the algorithm is increased.

Disclosure of Invention

The invention provides a dynamic brain function network analysis method based on representation learning, which is characterized in that training is carried out through a constructed model, and time-varying characteristics of different brain regions of normal people and schizophrenia patients at different moments are captured while the schizophrenia patients are screened.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows:

a schizophrenia patient screening method based on a space-time attention mechanism comprises the following steps:

step 1: acquiring a sequence signal value of each brain area of each observation object, wherein the sequence length is T;

step 2: acquiring time window characteristics of each brain area at T moments by adopting a sliding window mode, calculating the correlation of each brain area at each moment by utilizing a Pearson correlation coefficient, and constructing a dynamic brain function network with a time scale of T;

and step 3: performing PCA (Principal Component Analysis) dimension reduction operation on the acquired n x n-dimensional brain function network at T moments to reduce the dimension to n x d dimension; then, the vector is spread to T vectors with 1 x (n x d) dimensions, and the vector is reconstructed into T again by means of a sliding window_w1 (T/T)_wN d) dimensional vector; wherein, T_wFor the number of time windows, each time window contains T/T_wA brain function network, wherein n represents the number of brain areas, and d represents the characteristic dimension of the brain areas;

and 4, step 4: introducing a space-time attention mechanism, and learning the weight of each time window and each brain area in a self-adaptive manner according to the respective characteristics of the time window and the brain area; by means of a space-time attention mechanism, will T_wBrain function network aggregation within each time window;

and 5: building a convolutional neural network model; adopting convolution kernel of 1 x (n x d) to carry out convolution operation on the dynamic brain function network characteristics, and adopting convolution kernel of 1 x 1 to bear pooling action to capture time characteristics of the convolution kernel;

step 6: sending the features after the convolution pooling into a full-connection layer for dimensionality reduction;

and 7: and finishing classification by adopting a classifier.

Further, in step 4, the weight of each time window is calculated as follows:

1) flattening each network snapshot feature in the time window to a vector with dimensions of 1 x (n x d);

2) mapping each flattened network feature to a space Q for dimensionality reduction;

3) after the dimensionality reduction is finished, splicing a plurality of features in the same time window together to serve as the final time window feature;

4) by means of a weight vectorAfter each high-dimensional time window characteristic is changed into a scalar quantity, an activation function is used for activation operation, and the weight value alpha of each time window is obtained through calculation by means of a SoftMax function_i(ii) a The specific calculation is shown as the following formula;

wherein the content of the first and second substances,is a characteristic of each network after tiling within the jth time window,is the feature of each network after being tiled in the ith time window, | | is the concatenation symbol, | is an activation function, Q represents the space, in this formula is the vector space formed by a set of vectors.

Further, in step 4, the weight of each brain region is calculated as follows:

1) mapping the characteristics of each brain area in each time window into a space U;

2) splicing the characteristics of all the moments of each brain area together to serve as the characteristics of each brain area on the whole time sequence;

3) by means of a weight vectorMapping each high-dimensional brain region feature into a scalar, and calculating by means of a SoftMax function to obtain a weight value of each brain regionThe specific calculation is shown as the following formula;

whereinFeatures of the brain region o at 1 to T moments, respectively;the features of the brain region q at 1 to T moments are respectively, and U represents a space, which is a vector space formed by a group of vectors in the formula.

Further, in the step 4, the activation function σ adopts an activation function leak relu.

Further, in step 5, the objective loss function of the convolutional neural network model in combination with the attention mechanism is L₁Cross entropy loss function, and introduce F norm of matrix to prevent model training overfitting, after introducing, the final loss function of the model is:

wherein, V_LC represents the number of categories of the observed objects; y is_v，kIs an indicator function, if the observed object v belongs to class k, then y_v，k1, otherwise 0; p is a radical of_v，kIs the probability that the observed object v belongs to the class k, λ is the regular term coefficient, and W is the parameter that needs training and learning in the model.

Further, the convolutional neural network module in the step 5 is replaced by a recurrent neural network module.

Further, in step 7, the classifier employs a SoftMax function.

Compared with the prior art, the invention has the beneficial effects that: compared with the existing model, the method has high recall rate, and the model algorithm can screen out all schizophrenia patients; the model method captures the importance of different brain areas with different time windows by introducing an attention mechanism, thereby discovering the time-varying characteristics of dynamic brain function networks of different people; when the convolutional neural network module is used, the adjacent matrix of the dynamic brain function network is not directly used as the pixel matrix for feature processing, which is equivalent to continuing feature learning on the basis of keeping the topological information of the original dynamic brain function network at each moment.

Drawings

FIG. 1 is a schematic diagram of a dynamic brain function network classification model based on spatiotemporal attention.

Fig. 2 is a schematic illustration of the attention calculation of the time window.

Fig. 3 is a schematic diagram of attention calculation of a brain region.

Fig. 4 is a schematic view of the attention of the brain region of each observation target.

Fig. 5 is the frequency of occurrence of the top 1/2 ranked brain regions in the total observer population.

Fig. 6 is the frequency of important brain regions ranked top 1/2 among different populations.

Fig. 7 is a time window attention diagram of each observation object.

Fig. 8 is the average interval of the top 3 important time windows of the normal population and the schizophrenic patient population.

Fig. 9 is the average interval of the top 4 important time windows of the normal population and the schizophrenic patient population.

Fig. 10 is the average interval of the top 5 important time windows of the normal population and the schizophrenic patient population.

Figure 11 is a close-loop triangular structure trace formed by the amygdala _ left brain region of schizophrenic patient sample No. 1.

Figure 12 is a close-loop triangular structure trace formed by the amygdala _ left brain region of schizophrenic patient sample No. 2.

FIG. 13 is a close-loop triangular structure trace formed by the amygdala _ left brain region of normal human sample number 155.

Fig. 14 is a close-loop triangular structure trace formed by the amygdala _ left brain region of normal human sample number 156.

Figure 15 is a close-loop triangular structure trace formed by temporalis superior temporal gyrus right brain region of schizophrenia patient sample No. 1.

Figure 16 is a close-loop triangular structure trace formed by temporalis superior temporal gyrus right brain region of schizophrenia patient sample No. 2.

Fig. 17 is a trace of the closed-loop triangle structure formed by temporalis pole, superior temporal gyrus, right brain area of normal human sample number 155.

Fig. 18 is a trace of the closed-loop triangle structure formed by the temporalis pole, superior temporal gyrus-right brain region of normal human sample number 156.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

The invention aims to explore the time-varying characteristics of different brain areas of schizophrenia patients and normal people at different moments (namely, the pathogenesis of the schizophrenia patients is explored), provides a convolutional neural network model combined with a space-time attention mechanism, and captures the importance of the different brain areas at different moments while screening the schizophrenia patients.

Dynamic brain function network model and classification method

The invention relates to a schizophrenia patient screening method based on a space-time attention mechanism, which is shown in figure 1.

1) Acquiring a sequence signal value of each brain area of each observation object, wherein the sequence length is T;

2) acquiring time window characteristics of each brain area at T moments by adopting a sliding window mode, and calculating the correlation of each brain area at each moment by utilizing a Pearson correlation coefficient so as to construct a dynamic brain function network with a time scale of T;

3) the acquired n x n dimensional brain function network at T moments is reduced to n x d dimensions and then flattened to T1 x (n x d) dimensional vectors. And again reconstructed into T by means of a sliding window_wA (T/T)_wN d) dimensional vector, d representing the characteristic dimension of the brain region; t is_wFor the number of time windows, each time window contains T/T_wA network of individual brain functions.

4) A space-time attention mechanism is introduced, and the weight of each brain area in each time window is mainly learned in an adaptive mode according to the time window and the respective characteristics of the brain areas. And with the help of a space-time attention mechanism, will T_wAggregating a plurality of brain function networks in each time window (the characteristics of each time window of each brain area in each time window are multiplied by weight and then are subjected to an addition operation) to obtain T_w1 x (n x d) dimensional vector;

5) building a convolutional neural network model, specifically performing convolution operation on the dynamic brain function network characteristics by using a convolution kernel (the downward sliding distance is 1) of 1 x (n x d), and capturing the time characteristics by using the 1 x 1 convolution kernel (the sliding distance is 1) to play a role of pooling;

6) sending the features after the convolution pooling into a full-connection layer for dimensionality reduction;

7) classification (output probability that the observed subject belongs to normal persons and schizophrenic patients) is done using a classifier (SoftMax function).

The target loss function of the model is the cross entropy loss function L₁. To prevent the model from overfitting, the F-norm of the matrix is introduced and the final loss function L is shown as:

wherein, V_LIs a set of observation objects involved in training, and c represents the number of categories of observation objects, where c is 2. y is_v，kIs an indicator function, if the observed object v belongs to class k, then y_v，kIs 1, otherwise is 0. p is a radical of_v，kIs the probability that the observed object v belongs to the class k, λ is the regular term coefficient, and W is the parameter that needs training and learning in the model. According to the invention, a model for discriminating the schizophrenia patient is finally obtained by optimizing the loss function.

Two, space-time attention mechanism

The attention mechanism can focus on important information with high weight obtained by model training learning (the weight of each brain area in each time window needs to be considered when acquiring the dynamic brain function network characteristics), and discards unimportant information focused with low weight obtained by learning, so that the expandability and the robustness are extremely high. In the present model, different attention needs to be paid to different brain regions and different time windows in the observed object, and when dynamic brain function network features in each time window are aggregated in fig. 1, attention values are used. The premise of this model is that the brain region weight of the observed object remains unchanged for continuous observation time, and the calculation of the time window attention and the brain region attention are shown in fig. 2 and 3, respectively.

1. Time window attention

As shown in fig. 2, when calculating the attention of the time window, firstly, each network snapshot feature in the time window needs to be flattened to a 1 x (n x d) dimensional vector, and then each flattened network feature is mapped into the space Q for dimension reduction (each row vector in Q is used as a coordinate base to obtain a coordinate baseThe corresponding coordinate value under each coordinate base). After the dimension reduction operation is completed, a plurality of features in the same time window need to be spliced together to serve as a final time window feature. Finally, by means of a weight vector(will beAs a coordinate base), each high-dimensional time window feature is converted into a scalar quantity, then an activation function is used for activation operation, and a weight value alpha of each time window is obtained by calculation through a SoftMax function_i. The specific calculation is shown below.

Wherein the content of the first and second substances,is a characteristic of each network after tiling within the jth time window,the method is characterized in that the network is flattened in the ith time window, | | is a splicing symbol, sigma is an activation function, here, LeakyReLU can be selected as the activation function, the function is a variant of the ReLU function, and negative values are reserved according to a certain proportion.

2. Attention of brain region

The attention pattern for the brain region is calculated in a similar manner as for the time window. As shown in fig. 3, first, features of each brain region in each time window need to be mapped into the space U. And then, splicing the characteristics of all the moments of each brain area together to serve as the characteristics of each brain area on the whole time sequence. Finally, the same applies to a weight vectorMapping each high-dimensional brain region featureIs a scalar, and obtains the weight value of each brain area by means of calculation of a SoftMax functionThe specific calculation is shown below.

WhereinFeatures of the brain region o at 1 to T moments, respectively;are characteristic of the brain region q at 1 to T moments, respectively.

Thirdly, evaluation of classification results

The evaluation indexes commonly used by the classification model are Acc (Accuracy), P (Precision), R (Recall), and F1 values.

The definition of accuracy is shown as follows:

where m denotes the number of data samples, x_iFor one of the pieces of data, f (x)_i) Is denoted by x_iPredictive label of y_iIs data x_iThe real tag of (1).

The P, R, F1 values were calculated from TP, FN, FP, TN, and the definitions of TP, FN, FP, TN are shown in Table 1.

TABLE 1 Classification result confusion matrix

Wherein, TP represents the number of predicted positive examples in the true positive examples, FP represents the number of predicted positive examples in the true negative examples, TN represents the number of predicted negative examples in the true negative examples, and FN represents the number of predicted negative examples in the true positive examples.

The definition of accuracy, recall, and F1 values are shown below:

wherein, the accuracy rate P is the proportion of the true examples actually contained in the samples with the prediction labels as the true examples, the recall rate R is the proportion of the true examples predicted as the true examples, and the F1 value can be understood as the harmonic mean of P, R.

The model aims at screening schizophrenia patients. The model was compared to 3 baseline models, including the SDBFN-CNN model proposed by Kam et al, the HARL model proposed by Huang et al, and the wck-CNN model proposed by Jie et al. The first method is to train a deep neural network model based on a predefined dynamic brain function connection network, and the last two methods are to acquire the connection state and relevant characteristics of the dynamic brain function network in the deep neural network model training process.

TABLE 2 dynamic brain function network Classification results based on attention mechanism

The TS-CNN model in Table 2 is the model method in this chapter, and the recall (R) of this model method is highest compared with these methods, indicating that the model can distinguish all schizophrenia patients in the test set. The F1 value of the model is better than that of the SDBFN-CNN model, but is also closer to those of the HARL and wck-CNN models. The wck-CNN model has the highest F1 value, because the connection state of the dynamic brain function network in the wck-CNN model is learned by model training, each observation object can obtain a proper dynamic network topology structure for describing the correlation of each brain area according to the characteristics of the fMRI sequence information of each observation object. Of course, the algorithm complexity of the wck-CNN model is highest compared to other methods and the present model method. Overall, the present model F1 value is an acceptable result, and the present model can achieve more explanatory results (brain region attention and time window attention) than the SDBFN-CNN and hirl models.

Fourth, mechanism analysis

And analyzing the time-varying characteristics of different brain regions at different moments among different crowds by utilizing a statistical method aiming at the attention and time window importance of the brain regions obtained by model training.

1. Importance of brain region

Based on the attention of the brain region obtained by model training, the brain region is visualized, and it can be seen that partial brain region nodes (for example, the brain region No. 33 and the brain region No. 34, namely, the left island leaf and the right island leaf) have larger weights in normal people and schizophrenic people.

Based on the weight values, the occurrence frequency of the top 1/2 ranked brain region in all the observed people is counted as shown in fig. 5, so as to find out the frequent brain region as the important brain region. The 0.7 threshold is used to screen the brain areas with the frequency higher than 0.7, and the important brain areas are amygdala _ left (45), temporalis pole: temporalis superior _ right (88), temporalis pole: temporalis superior _ left (87), islets _ right (34), temporalis superior _ left (85), amygdala _ right (46), olfactory cortex _ left (17), islets _ left (33), and lenticular putamen _ right (78)9 brain areas.

Fig. 6 counts the frequency of these important brain regions ranked at top 1/2 among different populations. It can be seen from fig. 6 that there is a clear distinction between these important brain regions in the normal population of patients with nuclear schizophrenia, i.e. the frequency of the important brain regions in the normal population is higher than that in the population of patients with schizophrenia, from which it can be inferred that when normal people suffer from schizophrenia, the importance of the original important brain regions in the brain function network will be reduced.

2. Importance of time window

Fig. 7 is a visualization of the attention of time windows of the observation objects, and it can be preliminarily found that there is a difference in the time intervals of the important time windows of different observation objects, and the dark squares of the left half observation objects (numbers 1 to 154), i.e. the patient population, are significantly dispersed, while the dark squares of the right half observation objects, i.e. the normal population, are relatively dense.

In order to quantify the interval of important time windows, the invention searches that the average interval between the most important time windows is observed firstly, and the average intervals of important time windows which are ranked at the top 3, the top 4 and the top 5 are counted respectively. For example, the top five time windows are (17, 5, 1, 33, 45), respectively, then the average time interval is (5-1+17-5+33-17+45-33)/(5-1) ═ 11. From figures 8-10 it can be seen that the median of the mean interval of time windows of importance is higher for the schizophrenic patient population than for the normal population, and that the interval of time of importance is generally higher for the schizophrenic patient population as a whole than for the healthy population. The different weights of the time windows may be considered to be different in the states of the brain function network under the time windows, and the larger the interval of the important time windows is, the weaker the ability of the state transition to occur may be considered. The state transition ability of the schizophrenic patient population is weaker than that of the normal population.

In order to further and deeply analyze the difference between the normal population and the schizophrenia patient population, the topological difference of important brain areas of different populations at important time points is observed by combining the statistically obtained important brain areas. Specifically, a closed loop triangle structure formed by important brain areas in an important time window (1 time window comprises 3 time points) is tracked, if a closed loop triangle structure is formed with other important brain areas, the closed loop triangle structure formed by the other important brain areas is also tracked, and the visual description is performed by taking the amygdala _ left (45) and temporalis _ superior _ right (88) brain areas of two schizophrenia patient samples (No. 1 and No. 2) and two normal human samples (No. 155 and No. 156) as examples.

As shown in fig. 11 to 14, the large nodes are the important brain areas found, and the small nodes are the brain areas forming closed-loop triangles with the important brain areas. The closed-loop triangle structure of the motif has been proved to have certain biological significance in brain function networks. Taking the tracking of the community where the amygdala _ left brain region is located as an example, it can be found in the normal human samples No. 155 and 156 that there are many other important brain regions and the amygdala _ left brain region both forming a closed-loop triangle structure in continuous time, while the number of other important brain regions in the brain of the schizophrenic patient is relatively small, and as time evolves, the other important brain regions will be separated from the amygdala _ left brain region. For example, in the time window of importance for schizophrenia patient sample number 1 in fig. 5-10, the number of large nodes is decreasing.

Similarly, as shown in fig. 15 to 18, the same law was found by tracking the closed-loop triangle formed by the temporosuperior-right brain region, and the number of important brain regions was relatively small and decreased in the time windows of samples No. 1 and No. 2 of schizophrenic patients, and the closed-loop triangle structure was also decreased.

From the above example, it can be seen that in the dynamic brain function network, the important brain regions of the normal population are more biased to be connected with each other within the important time window, and simultaneously, more closed-loop triangle structures are formed. In contrast, the important brain regions of the patient population in the mental classification are more biased towards disconnection within the important time window, while the closed-loop triangle structure tends to be less.

TABLE 3 94 brain region numbers and their corresponding Chinese and English names studied in the description of the embodiments of the present invention

In conclusion, the method provided by the invention is trained through the constructed model, the schizophrenia patient is screened, meanwhile, the importance of different time windows of different brain areas is obtained, and the time-varying characteristics of the brain areas at various time points of the normal population and the schizophrenia patient population are captured. Compared with the existing model method, the method has high recall rate.