Nervous system disease evaluation method and system based on brain network diagram

Abstract

The invention provides a nervous system disease evaluation method and system based on a brain network graph, and the method comprises the steps: obtaining brain graph data of a target object, the brain graph data comprising time sequence data and a time-varying connection matrix of M brain regions in the brain of the target object; based on the brain map data of the target object, sliding window analysis is carried out, and a brain region feature sequence is determined; and inputting the brain region feature sequence into a nervous system disease evaluation model, and determining an evaluation index of the target object in a specified nervous system disease, the specified nervous system disease including one of schizophrenia, depression and Alzheimer's disease. On the basis of brain map data, sliding window analysis is carried out to realize multi-scale dynamic feature extraction, on the basis of model design and on the basis of taking bidirectional LSTM as a core, a full connection layer is introduced to carry out cross-brain region interactive modeling, and on the basis of loss function design, a sample imbalance factor is fully considered to improve the assessment accuracy of a specified nervous system disease.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical engineering and artificial intelligence, in particular, to a neural system disease evaluation method and system based on a brain network graph. BACKGROUND

[0002] In recent years, with the popularization of functional magnetic resonance imaging (fMRI) technology, further application based on brain functional connectivity network (such as abnormal detection, neural system disease risk assessment, mental health state prediction, etc.) has also made more progress. For the risk assessment of neural system diseases, the existing technology mainly adopts the following two schemes:

[0003] (1) Static functional connectivity analysis:

[0004] By calculating the average functional connectivity matrix (such as Pearson correlation coefficient) of the whole brain region in the whole period of time, the network topology attributes (such as node centrality, modularity index) are extracted as classification features for evaluation. A representative study such as Smith et al. (2015) uses support vector machine (SVM) to classify the static connection mode of Alzheimer's disease patients.

[0005] (2) Shallow dynamic network modeling:

[0006] Some studies attempt to introduce sliding window technology to construct dynamic functional connectivity (dFC) and use traditional machine learning models for classification. For example, Damaraju et al. (2014) use k-means clustering to identify the dynamic connection state of schizophrenia patients.

[0007] But the existing technical solutions of these two routes have some defects that cannot be ignored: the main problem of the static functional connectivity analysis scheme is the lack of time dynamics. The neural system diseases (such as the seizure period, the cognitive fluctuation of schizophrenia) are often accompanied by the time-varying nature of functional connectivity state, and static analysis cannot capture the key dynamic characteristics. The main problem of the shallow dynamic network modeling scheme is the limitation of feature expression and the neglect of long-range dependence. The hand-designed statistical features (such as state transition frequency) are difficult to comprehensively represent the complex spatiotemporal pattern of brain region connection state, and the traditional model (such as random forest) used cannot model the time evolution relationship between windows. Therefore, the accuracy of these two routes of technology has a lot of room for improvement. Moreover, both of these two schemes are difficult to effectively deal with the problem of sample imbalance, which is not conducive to improving the accuracy of evaluation. SUMMARY

[0008] The purpose of the embodiment of the present application is to provide a brain network graph-based nervous system disease evaluation method and system, which is based on brain graph data, performs sliding window analysis to realize multi-scale dynamic feature extraction, and in the model design, introduces a full connection layer for cross-brain region interaction modeling based on a bidirectional LSTM as the core, and in the loss function design, fully considers the sample imbalance factor to improve the evaluation accuracy for the specified nervous system disease.

[0009] To achieve the above purpose, the embodiment of the present application is realized by the following way:

[0010] The embodiment of the present application provides a brain network graph-based nervous system disease evaluation method, comprising: obtaining brain graph data of a target object, wherein the brain graph data comprises time series data and a time-varying connection matrix of M brain regions in the brain of the target object; performing sliding window analysis based on the brain graph data of the target object to determine brain region feature sequences; and inputting the brain region feature sequences into a nervous system disease evaluation model to determine an evaluation index of the target object for a specified nervous system disease, wherein the specified nervous system disease comprises one of schizophrenia, depression, and Alzheimer's disease.

[0011] Further, the brain graph data of the target object is generated based on a specific brain graph generation strategy, and the time series data of each brain region is calculated based on the weighted average of the time series data of each voxel in the brain region.

[0012] Further, based on the brain graph data of the target object, the sliding window analysis is performed to determine the brain region feature sequences, comprising: setting the length of the sliding window to 30 seconds and the step to 10 seconds, and determining K window data in total; for each window data: calculating the correlation coefficient of each specified brain region and other brain regions in the kth window data to obtain a correlation coefficient matrix corresponding to the kth window data, wherein the number of specified brain regions S is not more than the total number M of all brain regions in the brain graph data; and combining the correlation coefficient matrices corresponding to the K window data into a brain region feature sequence in the order of the window.

[0013] Further, the correlation coefficient of each specified brain region and other brain regions in the kth window data is calculated to obtain a correlation coefficient matrix corresponding to the kth window data, comprising: for the specified brain region i in the kth window data, i∈[1,S], calculating the correlation coefficient of the specified brain region i and other brain regions j:

[0014]

[0015] wherein, indicates the ΔT node lag correlation coefficient of the specified brain region i and other brain regions j in the kth window data, ΔT∈[0,3], which is determined based on the specified nervous system disease, T kthe number of time nodes of the kth window data, the signal value of the specified brain region i at time node t in the window data, the signal average of the specified brain region i at each time node in the window data, the signal value of the other brain region j at time node (t+ΔT) in the window data, the signal average of the other brain region j at each time node in the window data; the correlation coefficient matrix of the S specified brain regions and the (M-1) other brain regions in the kth window data is combined into an S×(M-1) matrix as the correlation coefficient matrix corresponding to the kth window data.

[0016] Further, the neural system disease evaluation model comprises an input layer, a feature extraction layer, a full connection layer and an output layer, the input layer is connected with the feature extraction layer, the feature extraction layer adopts S parallel bidirectional LSTM networks, each bidirectional LSTM network is connected with the full connection layer, the full connection layer is connected with the output layer, the input layer is used for receiving the input brain region feature sequence, and the brain region feature sequence is split into K correlation coefficient matrices, and the sequence K is processed: each correlation coefficient matrix is split into S (M-1) column vectors according to rows, and is input into the S bidirectional LSTM networks of the feature extraction layer after transposition; each bidirectional LSTM network of the feature extraction layer is used for feature extraction based on the input data, parallel processing of brain region data is realized, and the extracted features are input into the full connection layer; the full connection layer is used for full connection processing of the input features, completing feature fusion and nonlinear mapping, realizing cross-brain region interaction modeling, and inputting the fused features into the output layer; the output layer is used for determining the evaluation index of the target object in the specified neural system disease based on the fused features and outputting.

[0017] Further, the loss function is designed as:

[0018]

[0019] Wherein, L is the loss function, n is the sample number, y l is the sample label, taking 0 or 1, 0 representing healthy and 1 representing suffering from a specified neural system disease, is the predicted label, is the sample balance weight, is the focal adjustment factor, is the accuracy of the predicted probability.

[0020] Further, the sample balance weight satisfies:

[0021]

[0022] n=n h +n p ,

[0023] wherein, α y＝1 represents the weight of the sick label, n h is the number of healthy labels, n is the total number of sample labels, i.e. the number of samples, α y＝0 represents the weight of the healthy label, n p is the number of sick labels.

[0024] Further, the focus adjustment factor satisfies:

[0025]

[0026] wherein, γ base is a basic adjustment factor, is a constant, β is an adjustment coefficient, is a constant, is the validation set accuracy of the current model on the category label.

[0027] Further, the output layer adopts a sigmoid function.

[0028] The embodiment of the application further provides a nervous system disease evaluation system based on a brain network diagram, comprising: a data acquisition module, used to acquire brain diagram data of a target object, wherein the brain diagram data comprises time series data and a time-varying connection matrix of M brain regions in the brain of the target object; a feature processing module, used to perform sliding window analysis based on the brain diagram data of the target object, and determine a brain region feature sequence; and a disease evaluation module, used to input the brain region feature sequence into a nervous system disease evaluation model, and determine an evaluation index of the target object on a specified nervous system disease, wherein the specified nervous system disease comprises one of schizophrenia, depression and Alzheimer's disease.

[0029] Beneficial effects:

[0030] 1. The scheme is based on brain graph data (including time series data of M brain regions in the brain of the target object and time-varying connection matrix), sliding window analysis (using a sliding window strategy with a 30-second window length and a 10-second step) is performed, the dynamic evolution pattern of functional connection can be captured in the time dimension by combining the DT time delay correlation coefficient, multi-scale dynamic feature extraction is realized, and brain region feature sequences are obtained. The DT time delay can be adjusted according to the type of nervous system disease to be evaluated (for example, Alzheimer's disease needs to combine time delay to analyze the time property of brain region connection state, which usually has the characteristic of delayed information transmission, so as to reflect on the time delay of the connection state), so as to adapt to the evaluation of different specified nervous system diseases. In the design of the nervous system disease evaluation model, bidirectional LSTM is used as the core, and a fully connected layer is introduced to model the cross-brain region interaction. Based on the extraction of local dynamic characteristics of each brain region, the feature mapping of the fully connected layer can be used to capture the collaborative abnormalities between brain regions. Compared with traditional static functional connection analysis and shallow dynamic network modeling scheme, when evaluating the specified nervous system diseases (schizophrenia, depression, Alzheimer's disease), the scheme captures short-time dynamic characteristics through sliding window, and uses bidirectional LSTM to model long-time dependence, thereby overcoming the limitations of static analysis and shallow dynamic analysis, having higher sensitivity, and thus being beneficial to improving the accuracy of nervous system disease evaluation.

[0031] 2. In order to optimize the problem of sample imbalance, in the design of the loss function, the traditional cross-entropy is used as the basis, the double weight adjustment mechanism is introduced, the class balance weight is used to dynamically adjust according to the real-time batch sample distribution, and the small sample and healthy / patient sample imbalance problem is relieved; at the same time, the adaptive focus factor is used to adjust the attention to difficult samples (such as early Alzheimer's disease) based on the validation set accuracy. Then, when the classification of a certain class is difficult, the adaptive focus factor can be increased through the classification accuracy on the validation set, so as to pay more attention to the difficult-to-classify samples, and vice versa, if the sample classification is easy, the adaptive focus factor can be reduced through the classification accuracy on the validation set, so as to reduce the attention paid to the easy-to-classify samples, and avoid over-adjustment. In addition, the accuracy of the prediction probability is also added to the base adjustment, which is also helpful to deal with the imbalance of the samples, and the weight of the prediction correct is reduced and the weight of the prediction error is increased, so that the sample imbalance problem in the evaluation of the specified nervous system disease can be effectively dealt with, thereby improving the classification accuracy and the accuracy of the evaluation of the specified nervous system disease.

[0032] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0034] Figure 1 The flow chart of the neural system disease evaluation method based on the brain network graph provided in the embodiments of the present application.

[0035] Figure 2 The schematic diagram of the neural system disease evaluation model.

[0036] Figure 3 The schematic diagram of the neural system disease evaluation system based on the brain network graph.

[0037] Icon: 10-neural system disease evaluation model; 11-input layer; 12-feature extraction layer; 13-full connection layer; 14-output layer; 20-neural system disease evaluation system; 21-data acquisition module; 22-feature processing module; 23-disease evaluation module. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described below in combination with the drawings in the embodiments of the present application.

[0039] Please refer to Figure 1 , Figure 1 The flow chart of the neural system disease evaluation method based on the brain network graph provided in the embodiments of the present application. In this embodiment, the neural system disease evaluation method based on the brain network graph can include steps S10, S20 and S30.

[0040] The neural system disease evaluation method of the brain network graph in this embodiment mainly relies on the brain map data of the target object to be realized, and in order to improve the recognition accuracy in the construction process of the brain map, the brain map data generated based on the specific brain map generation strategy is adopted in this embodiment, and the time series data of each brain region is calculated by the weighted average value of the time series data of each voxel in the brain region based on the functional connection. In order to facilitate the understanding of this scheme, the generation scheme of the brain map data is briefly introduced here.

[0041] Firstly, the brain connection data of the target object is acquired, the brain connection data including fMRI data and sMRI data, then the brain connection data is preprocessed (head motion correction, time layer correction, denoising and spatial smoothing processing, etc.), and is registered (registered to the standard space) to obtain the registered brain map data. Then, based on the set brain atlas, the brain map data is divided into multiple brain regions, and the time series data of each brain region is determined. Then, based on the time series data of each brain region, the brain region connection matrix is determined (for example, the Pearson correlation coefficient is used to measure the connection strength between brain regions), wherein the non-diagonal line elements in the brain region connection matrix reveal the connection strength between brain regions. Finally, based on the brain region of the brain map data and the brain region connection matrix, the brain network diagram of the target object can be generated.

[0042] Since the core point of this specific brain map generation method is that the time series data of each voxel in the brain region is calculated based on the weighted average value of the functional connection, a little more detailed introduction is made here:

[0043] In the process of dividing the brain map data into multiple brain regions based on the set brain atlas, the set brain atlas can be loaded, so that the set brain atlas is in the same space (such as MNI space) as the fMRI data. Then the binary mask corresponding to each set brain region in the set brain atlas can be extracted, according to which each brain region in the brain map data is determined, and according to which the time series data of each brain region in the brain map data can be calculated based on the time series of each voxel in the brain region.

[0044] In order to analyze the correlation of a certain brain region to other brain regions, a calculation strategy is designed to calculate the weighted average value of the time series of each voxel in the brain region based on the functional connection as the time series data of the brain region.

[0045] The time series data of the brain region is calculated by the following formula:

[0046]

[0047] Wherein, M(t) represents the average signal value of brain region M at time node t, n is the total amount of voxels in brain region M, x i (t) represents the signal value of voxel i in brain region M at time node t, T is the total amount of time nodes, r i is the weight corresponding to voxel i in brain region M, the weight r i satisfies:

[0048] r i =max{r ic (k),r′ ic (k)},k∈[0,K],K∈Z + ,K<T, (2)

[0049]

[0050] wherein max{r ic (k),r′ ic (k)} represents the maximum value of r ic (k) and r′ ic (k), k is the number of lag nodes, K is the maximum number of lag nodes (for example, 3, then k needs to calculate four lag cases from 0 to 3), and K is a positive integer, K < T, r ic (k) represents the correlation coefficient of the voxel i in the brain region M and the target brain region, is the average value of the signal of the voxel i in the brain region M at each time node, x c (t+k) is the signal value of the reference time sequence of the target brain region at time node (t+k), is the average value of the signal of the reference time sequence of the target brain region at each time node, r′ ic (k) also represents the correlation coefficient of the voxel i in the brain region M and the target brain region, x i (t+k) is the signal value of the voxel i in the brain region M at time node (t+k), x c (t) is the signal value of the reference time sequence of the target brain region at time node (t).

[0051] The brain map data of the target object obtained in this way can reflect the connection of the brain region at the voxel level to some extent. Such time sequence data is beneficial to mining more detailed brain region connection information when analyzing the nervous system diseases reflected by the brain region connection strength, so as to improve the accuracy in evaluating such nervous system diseases.

[0052] The above is the introduction of generating the brain map data of the target object by using a specific strategy. The following will introduce the nervous system disease evaluation method based on the brain network map of the present embodiment.

[0053] Firstly, step S10 can be run.

[0054] Step S10: Obtain the brain map data of the target object, wherein the brain map data comprises time sequence data of M brain regions in the brain of the target object and a time-varying connection matrix.

[0055] In the present embodiment, the brain map data of the target object can be obtained, which is generated by a specific brain map generation strategy and comprises time sequence data of M brain regions in the brain of the target object and a time-varying connection matrix. The time sequence data of each brain region is calculated based on the weighted average value of the time sequence data of each voxel in the brain region according to the functional connection, as introduced in the foregoing.

[0056] After obtaining the brain mapping data of the target object, step S20 can be performed.

[0057] Step S20: Based on the brain mapping data of the target object, a sliding window analysis is performed to determine a brain region feature sequence.

[0058] In this embodiment, the length of the sliding window is set to 30 seconds, and the step is 10 seconds, and a total of K window data is determined. Taking 10 minutes of time series data as an example, scanning once every 2 seconds (i.e., 2 seconds for a signal value), the time series data has 300 nodes, the sliding window is set to 30 seconds (15 nodes), and the step is 10 seconds (5 nodes), so that 58 window data can be obtained.

[0059] Then, for each window data:

[0060] The correlation coefficient of each specified brain region and other brain regions in the kth window data is calculated to obtain a correlation coefficient matrix corresponding to the kth window data, wherein the number of specified brain regions S is not more than the total number of brain regions M in the brain mapping data.

[0061] It should be noted that the specified brain regions corresponding to the specified nervous system disease can be different, and the selection of the specified brain regions can utilize prior knowledge (i.e., existing research) to select, or can add some brain regions with relatively high correlation with the specified nervous system disease after principal component analysis (or elastic network analysis), of course, if the calculation efficiency is not considered, the whole brain region analysis can also be directly performed, that is, the number of specified brain regions S is equal to the total number of brain regions M, which is not limited here.

[0062] The specific calculation method adopts the following scheme: for the specified brain region i in the kth window data, i∈[1,S], the correlation coefficient of the specified brain region i and other brain regions j (other brain regions do not contain the specified brain region i at this time, and the number is M-1) is calculated:

[0063]

[0064] wherein, indicates the ΔT node lag correlation coefficient of the specified brain region i and other brain regions j in the kth window data, ΔT∈[0,3], which is determined based on the specified nervous system disease, T k is the number of time nodes of the kth window data, is the signal value of the specified brain region i at time node t in the window data, is the signal average value of the specified brain region i at each time node in the window data, is the signal value of the other brain region j at time node (t+ΔT) in the window data, is the signal average value of the other brain region j at each time node in the window data.

[0065] It should be noted that the value of the time lag node number ΔT can be one value or multiple values, but when multiple values are taken, the multiple values need to be processed, and the form of the correlation coefficient matrix fused needs to be adjusted accordingly. In this embodiment, one time lag is taken as an example, so the form of the correlation coefficient matrix given is an Sx(M-1) matrix, but in other implementation manners, the correlation coefficients of multiple time lags can be considered at the same time, so as to obtain the correlation coefficient matrix in the corresponding form.

[0066] For example, when the nervous system disease is specified as schizophrenia, there is usually abnormality in the prefrontal-thalamus-cerebellum loop, which is manifested as impaired cognitive function, and the default network is out of synchronization, which is manifested as affecting self-referential thinking. Therefore, when the nervous system disease is specified as schizophrenia, it is mainly dominated by short time lag, and the connection state when ΔT is 1 is mainly considered. In other embodiments, the connection state when ΔT is 3 can also be used for auxiliary analysis. When the nervous system disease is specified as depression, the connection between the limbic system and the prefrontal lobe is usually weakened, which is manifested as ineffective emotional regulation; the reward network function is low, which is manifested as lack of pleasure, and the default network is overactive, which is manifested as enhanced self-referential thinking, but the information transmission efficiency is reduced. Therefore, when the nervous system disease is specified as depression, it is mainly dominated by medium time lag, and the connection state when ΔT is 2 is mainly considered to reflect the global network efficiency. Similarly, the connection state when ΔT is 3 can also be used for auxiliary analysis to analyze the long-range information transmission efficiency across brain regions. When the nervous system disease is specified as Alzheimer's disease, the default network is usually disintegrated, which is manifested as delayed information transmission from the posterior cingulate gyrus to the hippocampus; the global network efficiency is reduced, which is manifested as reduced information transmission speed, and there is memory integration disorder, etc. Therefore, when the nervous system disease is specified as Alzheimer's disease, it is mainly dominated by long time lag, and the connection state when ΔT is 3 is mainly considered to reflect the long-range information transmission efficiency. Therefore, this embodiment takes the connection state analysis when ΔT takes a single value as an example, which should not be regarded as a limitation of the present application.

[0067] Accordingly, the correlation coefficient matrix of the S specified brain regions and the (M-1) other brain regions in the kth window data can be combined into an Sx(M-1) matrix as the correlation coefficient matrix corresponding to the kth window data.

[0068] After obtaining the correlation coefficient matrix corresponding to each window data, the correlation coefficient matrices corresponding to the K window data can be combined into a brain region feature sequence according to the window order, that is, a KxSx(M-1) brain region feature sequence.

[0069] After obtaining the brain region feature sequence, step S30 can be performed.

[0070] Step S30: inputting the brain region feature sequence into the nervous system disease evaluation model to determine the evaluation index of the target object in the specified nervous system disease, wherein the specified nervous system disease includes one of schizophrenia, depression, and Alzheimer's disease.

[0071] In this embodiment, before introducing this step, the nervous system disease evaluation model of this embodiment is introduced. Please refer to Figure 2 The nervous system disease evaluation model includes an input layer, a feature extraction layer, a full connection layer, and an output layer. The input layer is connected with the feature extraction layer, the feature extraction layer adopts S parallel bidirectional LSTM networks, each bidirectional LSTM network is connected with the full connection layer, and the full connection layer is connected with the output layer.

[0072] For example, the input layer is used to receive the input brain region feature sequence and split the brain region feature sequence into K correlation coefficient matrices, and process them according to the sequence K: split each correlation coefficient matrix into S (M-1) column vectors according to the row, transpose them into row vectors, and then input them into the S bidirectional LSTM networks of the feature extraction layer.

[0073] Each bidirectional LSTM network of the feature extraction layer can perform feature extraction based on the input data (for example, the first bidirectional LSTM network receives the split data of the input layer), realize parallel processing of the data of each brain region, and input the extracted features to the full connection layer.

[0074] Here, the full connection layer is used to perform full connection processing on the input features, complete feature fusion and nonlinear mapping, realize cross-brain region interaction modeling, and input the fused features to the output layer.

[0075] The output layer mainly determines the evaluation index of the target object in the specified nervous system disease based on the fused features and outputs it. In this embodiment, the output layer adopts a sigmoid function to determine the binary classification probability (for example, the probability of being classified as healthy and the probability of being classified as suffering from the specified nervous system disease), as the evaluation index of the target object in the specified nervous system disease.

[0076] The model building here is relatively simple and can be built with PyTorch. The logic code for building the model is not shown here. During the training of the model, due to the small sample size and sample imbalance problem existing in the application scenario (i.e., the patient data sample is small, while the healthy data sample is large), it needs to be considered, so when building the model, a custom loss function needs to be defined to deal with this sample imbalance problem.

[0077] The loss function in this embodiment is designed as:

[0078]

[0079] where L is the loss function, n is the number of samples, y l is the sample label, taking 0 or 1, 0 represents healthy, 1 represents suffering from a specified nervous system disease, is the predicted label, is the sample balance weight, is the focal adjustment factor, is the accuracy of the predicted probability.

[0080] where the sample balance weight satisfies:

[0081]

[0082] n=n h +n p , (9)

[0083] where α y＝1 represents the weight of the disease label, n h is the number of healthy labels, n is the total number of sample labels, i.e. the number of samples, α y＝0 represents the weight of the healthy label, n p is the number of disease labels.

[0084] And the focal adjustment factor satisfies:

[0085]

[0086] where γ base is the basic adjustment factor, which is a constant (for example, take 2), β is the adjustment coefficient, which is a constant (for example, take 3), is the validation set accuracy of the current model on this class label.

[0087] In the process of building the model, it is necessary to customize this loss function, and the pseudo code is as follows:

[0088] classDynamicFocalLoss(nn.Module):

[0089] def__init__(self,gamma_base=2,beta=3):

[0090] super().__init__()

[0091] self.gamma_base=gamma_base

[0092] self.beta=beta

[0093] self.val_acc = {'0': 0.9, '1': 0.5} # initial validation set accuracy

[0094] def forward(self, y_pred, y_true):

[0095] # compute class weights

[0096] n_healthy = (y_true == 0).sum().item()

[0097] n_patient = y_true.size(0) - n_healthy

[0098] alpha_1 = n_healthy / y_true.size(0)

[0099] alpha_0 = n_patient / y_true.size(0)

[0100] # adaptive gamma

[0101] gamma_1 = self.gamma_base + self.beta * (1 - self.val_acc['1'])

[0102] gamma_0 = self.gamma_base + self.beta * (1 - self.val_acc['0'])

[0103] # compute p_t

[0104] p_t = torch.where(y_true == 1, y_pred, 1 - y_pred)

[0105] # compute loss weights

[0106] alpha_t = torch.where(y_true == 1, alpha_1, alpha_0)

[0107] gamma_t = torch.where(y_true == 1, gamma_1, gamma_0)

[0108] modulator = (1 - p_t) ** gamma_t

[0109] # base cross-entropy

[0110] bce = F.binary_cross_entropy(y_pred, y_true, reduction='none')

[0111] # total loss

[0112] loss = (alpha_t * modulator * bce).mean()

[0113] return loss

[0114] To optimize the problems of small sample and sample imbalance, in the design of the loss function, the traditional cross-entropy is used as the basis, and a double-weight adjustment mechanism is introduced. The class balance weight is used to dynamically adjust according to the real-time batch sample distribution, which can alleviate the problems of small sample and health / patient sample imbalance. At the same time, an adaptive focus factor is used to adjust the attention to difficult samples (such as early Alzheimer's disease) based on the validation set accuracy. Therefore, when the classification of a certain class is difficult, the adaptive focus factor can be increased by the classification accuracy on the validation set, so as to pay more attention to the difficult samples. On the contrary, if the sample classification is easy, the adaptive focus factor can be reduced by the classification accuracy on the validation set, so as to reduce the attention paid to the easy samples, avoiding over-adjustment. In addition, the accuracy of the prediction probability is also added to the base adjustment, which is also conducive to dealing with the imbalance of the sample. The weight of the correct prediction is reduced, and the weight of the incorrect prediction is increased, which can effectively deal with the sample imbalance problem in the evaluation of the specified nervous system disease, so as to improve the classification accuracy and enhance the accuracy of the evaluation of the specified nervous system disease.

[0115] Please refer to Figure 3 , Figure 3 FIG. 1 is a schematic diagram of a nervous system disease evaluation system 20 based on a brain network diagram.

[0116] The nervous system disease evaluation system 20 based on the brain network diagram can include a data acquisition module 21 configured to acquire brain diagram data of a target object, wherein the brain diagram data includes time series data and a time-varying connection matrix of M brain regions in the brain of the target object.

[0117] The feature processing module 22 is configured to perform sliding window analysis based on the brain diagram data of the target object to determine a brain region feature sequence.

[0118] The disease evaluation module 23 is configured to input the brain region feature sequence into a nervous system disease evaluation model to determine an evaluation index of the target object for a specified nervous system disease, wherein the specified nervous system disease includes one of schizophrenia, depression, and Alzheimer's disease.

[0119] In summary, the embodiment of the present application provides a neural system disease evaluation method and system based on brain network graph. Based on brain graph data (including time series data and time-varying connection matrix of M brain regions in the brain of a target object), sliding window analysis is performed (using a sliding window strategy with a 30-second window length and a 10-second step), the dynamic evolution mode of functional connection can be captured in the time dimension by combining the ΔT time lag correlation coefficient, multi-scale dynamic feature extraction is realized, and brain region feature sequences are obtained. The ΔT time lag can be adjusted according to the type of the neural system disease to be evaluated (for example, Alzheimer's disease needs to combine time lag analysis of brain region connection state, which usually has the characteristics of information delay transmission, so as to reflect on the time lag of the connection state), so as to adapt to the evaluation of different specified neural system diseases. In the design of the neural system disease evaluation model, bidirectional LSTM is used as the core, and a fully connected layer is introduced to model the cross-brain region interaction. Based on the extraction of local dynamic characteristics of each brain region, the feature mapping of the fully connected layer can be used to capture the collaborative abnormalities between brain regions. Compared with the traditional static functional connection analysis and shallow dynamic network modeling scheme, when evaluating the specified neural system diseases (schizophrenia, depression, and Alzheimer's disease), the present scheme captures short-time dynamic characteristics through a sliding window, and uses bidirectional LSTM to model long-term time dependence, thereby overcoming the limitations of static analysis and shallow dynamic analysis, having higher sensitivity, and thus being beneficial to improving the accuracy of neural system disease evaluation.

[0120] In this document, the term "only" is used to distinguish one entity or operation from another entity or operation, but does not necessarily require or imply any such actual relationship or order between the entities or operations.

[0121] The above only describes the embodiments of the present application and does not limit the protection scope of the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for evaluating nervous system diseases based on brain network diagram, characterized in that: include: Acquiring brain map data of a target subject, wherein the brain map data includes time series data and a time-varying connectivity matrix of M brain regions in the target subject's brain; Based on the brain map data of the target subject, sliding window analysis is performed to determine the characteristic sequence of brain regions; The brain region feature sequence is input into the neurological disease evaluation model to determine the evaluation index of the target object for the specified neurological disease, where the specified neurological disease includes one of schizophrenia, depression, and Alzheimer's disease.

2. The method for evaluating nervous system diseases based on brain network diagram according to claim 1, characterized in that: The brain map data of the target subject is generated based on a specific brain map generation strategy, and the time series data of each brain region is calculated by the weighted average of the time series data of each voxel in the brain region based on functional connectivity.

3. The method for evaluating nervous system diseases based on brain network diagram according to claim 1, characterized in that: Based on the brain map data of the target subject, a sliding window analysis is performed to determine the brain region feature sequence, including: The length of the sliding window is set to 30 seconds, the step length is 10 seconds, and a total of K window data are determined; For each window data: calculate the correlation coefficient between each designated brain region in the k-th window data and other brain regions, and obtain the correlation coefficient matrix corresponding to the k-th window data, where the number S of designated brain regions does not exceed the total number M of all brain regions in the brain map data; The correlation coefficient matrices corresponding to the K window data are combined into a brain region feature sequence according to the window order.

4. The method for evaluating nervous system diseases based on brain network diagram according to claim 3, characterized in that: Calculate the correlation coefficients between each designated brain region in the k-th window data and other brain regions, and obtain the correlation coefficient matrix corresponding to the k-th window data, including: For the specified brain region i in the k-th window data, i∈[1,S], calculate the correlation coefficient between the specified brain region i and other brain regions j: in, represents the delta T node lag correlation coefficient between the specified brain region i and other brain regions j in the k-th window data, delta T∈[0,3], determined based on the specified neurological disease, T k is the number of time nodes of the k-th window data, is the signal value of the specified brain area i at time node t in the window data, is the average signal value of the specified brain region i at each time node in the window data, is the signal value of other brain regions j at the time node (t+ΔT) in the window data, is the average signal value of other brain regions j at each time node in the window data; The correlation coefficients of the S specified brain regions and (M-1) other brain regions in the k-th window data are combined into an S×(M-1) matrix as the correlation coefficient matrix corresponding to the k-th window data.

5. The method for evaluating nervous system diseases based on brain network diagram according to claim 3, characterized in that: The neurological disease evaluation model includes an input layer, a feature extraction layer, a fully connected layer, and an output layer. The input layer is connected to the feature extraction layer. The feature extraction layer uses S parallel bidirectional LSTM networks. Each bidirectional LSTM network is connected to the fully connected layer. The fully connected layer is connected to the output layer. The input layer is used to receive the input brain region feature sequence and split the brain region feature sequence into K correlation coefficient matrices, and process them according to the sequence K: each correlation coefficient matrix is ​​split into S (M-1) column vectors according to the row, and after transposition, they are respectively input into the S bidirectional LSTM networks of the feature extraction layer; Each bidirectional LSTM network in the feature extraction layer is used to extract features based on the input data, realize parallel processing of data from each brain region, and input the extracted features into the fully connected layer; The fully connected layer is used to perform fully connected processing on the input features, complete feature fusion and nonlinear mapping, realize cross-brain region interaction modeling, and input the fused features into the output layer; The output layer is used to determine and output the evaluation index of the target object for a specified nervous system disease based on the fused features.

6. The method for evaluating nervous system diseases based on brain network diagram according to claim 5, characterized in that: The loss function is designed as: Among them, L is the loss function, n is the number of samples, y l is the sample label, which takes 0 or 1, 0 means healthy, 1 means suffering from a specified neurological disease, To predict the label, is the sample balancing weight, is the focus adjustment factor, is the accuracy of the predicted probability.

7. The method for evaluating nervous system diseases based on brain network diagram according to claim 6, characterized in that: Sample balance weight satisfy: n=n h +n p , Among them, α y＝1 represents the weight of the diseased label, n h is the number of healthy labels, n is the total number of sample labels, i.e. the number of samples, α y＝0 represents the weight of the health label, n p is the number of diseased labels.

8. The method for evaluating nervous system diseases based on brain network diagram according to claim 6, characterized in that: Focus adjustment factor satisfy: Among them, γ base is the basic adjustment factor, is a constant, β is the adjustment coefficient, is a constant, is the validation set accuracy of the current model on this category label.

9. The method for evaluating nervous system diseases based on brain network diagram according to claim 5, characterized in that: The output layer uses the sigmoid function.

10. A neurological disease evaluation system based on brain network diagram, characterized in that: include: A data acquisition module is used to acquire brain map data of a target subject, wherein the brain map data includes time series data and a time-varying connectivity matrix of M brain regions in the target subject's brain; The feature processing module is used to perform sliding window analysis based on the brain map data of the target object to determine the brain region feature sequence; The disease evaluation module is used to input the brain region feature sequence into the neurological disease evaluation model to determine the evaluation index of the target object for a specified neurological disease, where the specified neurological disease includes one of schizophrenia, depression, and Alzheimer's disease.

Images