Pseudo-tag variational self-encoding network-based mental disease biological typing method

Abstract

The invention provides a mental disease biological typing method based on a pseudo-tag variational self-encoding network, and aims to realize automatic biological typing of mental diseases through efficient classification and feature extraction. The method mainly comprises the following steps: classifying samples of mental patients and healthy control groups into N + 1 classes, and initializing pseudo tags; extracting sample features by adopting a coding network, comparing the sample features with the class mean features, and determining a sample pseudo tag; generating a high-level feature vector and performing signal reconstruction; and predicting a sample label through the classification sub-network. And then updating the pseudo labels and carrying out multi-round iteration until the sample classification is stable. And finally, outputting various types of typical biological signals by using the updated pseudo tag. According to the method, a loss function is constructed through five measurement indexes so as to optimize the model performance. The method is expected to promote biological research and application of mental diseases, and provides decision support for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to a biological classification method for mental diseases based on a pseudo-label variational autoencoder network, and belongs to the field of intelligent assisted classification diagnosis of mental diseases. Background Art

[0002] In the field of classification diagnosis of mental diseases, most of the existing methods are diagnosis methods based on symptomatology, that is, by observing and scoring the symptomatology of patients to determine the diseases or disease subtypes suffered by the patients. However, this symptom-based diagnosis method cannot reveal the variations in the brains of individuals and is not conducive to the formulation of subsequent treatment plans. Paying attention to the biological abnormalities in the brains of patients, rather than symptomatological abnormalities, is more helpful for the treatment of mental diseases in patients. In clinical practice, due to the high comorbidity rate of mental diseases and the fact that different mental diseases may present the same or similar symptoms, a classification diagnosis method for biological typing of mental diseases becomes particularly important. At the same time, biological data (such as electroencephalogram (EEG) data, magnetic resonance imaging (MRI) data, and various biological signal data derived therefrom, such as amplitude of low-frequency fluctuation (ALFF), regional homogeneity (Reho), gray matter thickness (GM), etc.) can more objectively reflect the variations in the brain. Therefore, the application of these biological data also provides a favorable guarantee for the realization of biological typing of mental diseases.

[0003] However, how to efficiently and correctly perform biological typing remains a challenging topic. The difficulty lies in that the relationship between the degree of brain variation and biological data is complex and not a simple linear mapping relationship. There is also a certain degree of uncertainty in brain variation. The development of a disease is a process from quantitative change to qualitative change, and the ability of individuals to tolerate this change varies. Therefore, there may be changes in the brain parenchyma but typical disease characteristics have not yet been reflected. At the same time, the human brain also has a considerable degree of adaptability and there is a compensatory mechanism. These uncertain factors all bring difficulties to biological typing. Therefore, the present invention uses a variational autoencoder network, a network based on a probability model, to simulate the above-mentioned various uncertainties, and assigns learnable pseudo-labels to different disease categories during the typing process to efficiently represent various types of biological characteristics adaptively. Finally, through pairwise comparison of the characteristics of the healthy group and various types of biological characteristics, a relatively effective biological typing of mental diseases is achieved. Summary of the Invention

[0004] The present invention provides a biological classification method for mental disorders based on a pseudo-label variational autoencoder network. The key lies in that, compared with the healthy control group, this method can use biological signal data to adaptively divide the samples in the mental disorder patient group into N biological types based on the variational autoencoder network. The variational autoencoder network involved in the present invention can fully consider the uncertainty of sample brain variations and use the probability mechanism in the variational autoencoder network to model this uncertainty. At the same time, this network requires that at the level of high-level features and the restoration of biological signals in the network output, the pairwise differences between each classification feature and the features of the healthy control group are maximized, so as to more efficiently distinguish each biological classification. At the same time, for the pseudo-label learning of each category of biological classification and the healthy control group category, the overall high-level features of the samples within the category are effectively aggregated, and the mean feature vector and variance feature vector corresponding to the category are formed, ultimately providing a guarantee for the formation of the typical biological signal of this category. Therefore, the method of the present invention provides a new perspective for the biological classification of mental disorders.

[0005] The technical solution of the present invention is as follows:

[0006] A biological classification method for mental disorders based on a pseudo-label variational autoencoder network, comprising the following steps:

[0007] Step 1: Assume that all samples in the mental disorder patient group can be divided into N categories according to biological classification, with corresponding labels from l1 to l N , and all samples in the healthy control group are divided into one category, with the corresponding label l N+1 , a total of N + 1 categories, and these categories respectively correspond to N + 1 sets, denoted as {S1,..., S N , S N+1}, where the sets S1 to S N correspond to the classification types l1 to l N of mental disorders, and the S N+1 set corresponds to the healthy control group category l N+1 . Randomly initialize the pseudo-labels of each category, which are composed of a class mean feature vector m li and a class variance feature vector v li . Each category can contain multiple samples, but at this time, each category is temporarily defined as an empty set and does not contain any samples;

[0008] Step 2: Input the biological signals x k of each sample in the mental disorder patient group and the healthy control group based on brain regions into the encoding sub-network module of this variational autoencoder network to obtain the sample mean feature vector m k and the sample variance feature vector v k corresponding to each sample;

[0009] Step 3: If the sample is a mental disorder patient, then the sample mean feature vector m of this samplek Compared with the class mean eigenvector m of various biological classification types of mental diseases li Take the pseudo-label of the class with the class mean eigenvector most similar to the sample mean eigenvector of the sample as the pseudo-label of the sample, and at the same time add the sample to the corresponding classification class li The corresponding set S i Among them, if the sample is a healthy sample, set the pseudo-label of the sample to the pseudo-label of the healthy control group class l N+1 and at the same time add the sample to the corresponding set S N+1 ;

[0010] Step 4: For each sample, generate a high-level eigenvector m k from the sample mean eigenvector m k and the sample variance eigenvector of the sample, and denote it as z k , where β is taken as a random number of a standard normal distribution; k

[0011] Step 5: Input the high-level eigenvector z k generated by each sample into the decoding sub-network module of the variational autoencoder network, and this module outputs a corresponding reconstructed biological signal y k ;

[0012] Step 6: Input the high-level eigenvector z k generated by each sample into a classification sub-network module, and this module outputs a classification label prediction result l k for the biological signal x k of the input sample, predicting whether each sample comes from the biological classification class or the healthy control group class to which it belongs;

[0013] Step 7: After all samples are input into the variational autoencoder network based on pseudo-labels and the samples in the sets S1 to S N+1 are added, update the pseudo-labels of each class, that is, the class mean eigenvector m li and the class variance eigenvector v li ;

[0014] Step 8: Perform one round of network training on the variational autoencoder network based on pseudo-labels, and return to Step 1 to perform iterative calculations of Steps 1 to 8. However, at this time, the pseudo-labels of each class in Step 1 are replaced with the updated pseudo-labels of the corresponding classes. After multiple rounds of iteration, when the samples included in each class are the same as those included in each class in the previous round of iteration, the iteration ends;

[0015] Step 9: Pass the class mean eigenvector mcorresponding to the pseudo-labels of the mental disease classification classes and the healthy control group classes through the decoding sub-network module of the variational autoencoder network to output the corresponding biological signal y li ​li as a typical biological signal of various types, so as to determine the final biological characteristics of each type in the biological classification of mental diseases.

[0016] In the above step 3, when taking the pseudo-label of the class with the most similar class mean feature vector and sample mean feature vector as the pseudo-label of the sample, the class mean feature vector m li and the sample mean feature vector m k The similarity measurement criterion is taken as the minimum criterion of the mean method or the minimum criterion of cosine similarity.

[0017] In the above step 7, when updating the pseudo-label, for the class mean feature vector m of each class li is the sample mean feature vector m of the samples contained in this class k The average value of, and the class variance feature vector v of each class li is the sample variance feature vector v of the samples contained in this class k The average value of.

[0018] In the above step 8, the network loss function is set, and this loss function is measured by the following five measurement values:

[0019] 1. In step 3, the relative entropy (KL divergence) measurement of each class (the samples contained in each set S i ), and the measurement value is denoted as K1;

[0020] 2. In step 5, the similarity measurement between the output biological signal y k and the input biological signal x of the corresponding sample in step 2 k , and the measurement value is denoted as K2;

[0021] 3. In step 6, comparing the classification label prediction result l of the input sample biological signal x k with the corresponding label l of this sample in step 3 k , calculating the similarity measurement, denoted as the measurement value K3; i 4. In step 7, for the updated class mean feature vectors m of each class

[0022] , calculate the sum of the Euclidean distance measurements between pairwise vectors, and the measurement value is denoted as K4; li 5. In step 8, for the updated class mean feature vectors m of each class

[0023] The corresponding typical biological signal y output by the decoding sub-network module of this variational auto-encoder network li , calculate the sum of the absolute values of the pairwise correlation values of the typical biological signals corresponding to all classes as the measurement, and the measurement value is denoted as K5; li

[0024] Finally, the loss function is K1 + K2 + K3 - K4 + K5, and this loss function is minimized during network training. Description of the Drawings

[0025] Figure 1 is the network model involved in a method for biological classification of mental disorders based on a pseudo-label variational autoencoder network according to the present invention;

[0026] Figure 2 is the classification effect obtained by the present invention in the biological classification of anxiety disorders. Detailed Embodiments

[0027] The following further elaborates on a method for biological classification of mental disorders based on a pseudo-label variational autoencoder network according to the present invention with reference to the accompanying drawings.

[0028] As Figure 1 shown, a method for biological classification of mental disorders based on a pseudo-label variational autoencoder network includes the following steps:

[0029] Step 1: Assume that all samples in the mental disorder patient group can be biologically classified into N categories, corresponding to labels l1 to l N , and all samples in the healthy control group are divided into one category, corresponding to label l N+1 , for a total of N + 1 categories. These categories respectively correspond to N + 1 sets, denoted as {S1,..., S N , S N+1}, where the sets S1 to S N correspond to the classification types l1 to l N of mental disorders, and the set S N+1 corresponds to the healthy control group category l N+1 . Randomly initialize the pseudo-labels for each category, which are composed of a class mean feature vector m li and a class variance feature vector v li . Each category can contain multiple samples, but at this time, each category is temporarily defined as an empty set and does not contain any samples;

[0030] Step 2: Input the biological signals x k of each sample in the mental disorder patient group and the healthy control group based on brain regions into the encoding sub-network module of this variational autoencoder network to obtain the sample mean feature vector m k and the sample variance feature vector v k corresponding to each sample;

[0031] Step 3: If the sample is a mental disorder patient, then compare the sample mean feature vector m k of this sample with the class mean feature vectors m liCompare, and take the pseudo-label of the class that is most similar to the sample mean feature vector of this sample among the class mean feature vectors as the pseudo-label of this sample. At the same time, add this sample to the corresponding classification class li to the corresponding set S i Among them, if the sample is a healthy sample, set the pseudo-label of this sample to the pseudo-label of the healthy control group class l N+1 and add this sample to the corresponding set S N+1 ;

[0032] Step 4: For each sample, generate a high-level feature vector m k from the sample mean feature vector m k of this sample and the sample variance feature vector, k denoted as z k , where β is taken as a random number from a standard normal distribution;

[0033] Step 5: Input the high-level feature vector z k generated by each sample into the decoding sub-network module of this variational auto-encoder network, and this module outputs a corresponding biological signal y k ;

[0034] Step 6: Input the high-level feature vector z k generated by each sample into a classification sub-network module, and this module outputs a classification label prediction result l k for the biological signal x k of the input sample, predicting whether each sample comes from the biological classification class or the healthy control group class it belongs to;

[0035] Step 7: After all samples are input into the variational auto-encoder network based on pseudo-labels and the samples in the sets S1 to S N+1 are added, update the pseudo-labels of each class, that is, the class mean feature vector m li and the class variance feature vector v li ;

[0036] Step 8: Perform one round of network training on this variational auto-encoder network based on pseudo-labels, and return to Step 1 to perform iterative calculations of Steps 1 to 8. However, at this time, the pseudo-labels of each class in Step 1 are replaced with the updated corresponding pseudo-labels of each class. After multiple rounds of iteration, when the samples included in each class are the same as those included in each class in the previous round of iteration, the iteration ends;

[0037] Step 9: Input the class mean feature vector m li corresponding to the pseudo-labels of the mental illness classification classes and the healthy control group classes into the decoding sub-network module of this variational auto-encoder network, and output the corresponding biological signal y li , as the typical biological signal of each class, thereby determining the final biological characteristics of each type in the biological classification of mental illness.

[0038] In step 3 above, when taking the pseudo-label of the class with the class mean feature vector most similar to the sample mean feature vector as the pseudo-label of the sample, the class mean feature vector m li and the sample mean feature vector m k The similarity measurement criterion is taken as the minimum criterion of the mean method or the minimum criterion of cosine similarity.

[0039] In step 7 above, when updating the pseudo-label, for the class mean feature vector m li of each class, it is the average value of the sample mean feature vector m k of the samples contained in the class. For the class variance feature vector v li of each class, it is the average value of the sample variance feature vector v k of the samples contained in the class.

[0040] In step 8 above, the network loss function is set, and the loss function is measured by the following five measurement values:

[0041] 1. In step 3, the relative entropy (KL divergence) measurement of each class (the samples contained in each set S i ), the measurement value is denoted as K1, and the specific measurement can be set as For K2, it is required to be minimized so that the sample mean feature vector and the sample variance mean vector of the samples contained in S i have a good clustering effect under the Gaussian probability distribution;

[0042] 2. In step 5, the similarity measurement between the output biological signal y k and the input biological signal x k of the corresponding sample in step 2, the measurement value is denoted as K2, and the specific measurement can be set as In the formula, K is the total number of input samples. For K2, it is required to be minimized so that the information content of the output biological signal and the input biological signal is consistent and information loss is minimized as much as possible;

[0043] 3. In step 6, the classification label prediction result l k of the input sample biological signal x k is compared with the corresponding label l i of this sample in step 3, and the similarity measurement is calculated, denoted as the measurement value K3. The specific measurement can be set as multi-class cross-entropy measurement. For K3, it is required to be minimized so that the classification label prediction result is consistent with the corresponding label of this sample;

[0044] 4. In step 7, for the updated class mean feature vectors m li of each class, calculate the sum of the Euclidean distance measurements between their vectors pairwise, denoted as the measurement value K4, and the specific measurement can be set as Maximize K4 to increase the differences among various types on the class mean feature vectors;

[0045] 5. In step 8, for the updated class mean feature vectors m of each class li The corresponding typical biological signal y output by the decoding sub-network module of the variational auto-encoder network li , calculate the sum of the absolute values of the pairwise correlation values of the typical biological signals corresponding to all classes as a metric, denoted as K5. The specific metric can be set as Minimize K5 to orthogonalize the typical biological signals of each class, minimize the overlap of biological features in different classes on the brain regions as much as possible, and improve the significance of each class in the brain region representation;

[0046] Finally, the loss function is K1 + K2 + K3 - K4 + K5, and this loss function is minimized during network training.

[0047] Implementation Example 1:

[0048] 1) Example experimental conditions

[0049] In this example experiment, functional magnetic resonance imaging data of an anxiety disorder patient group (including four subtypes, namely social anxiety disorder, 25 cases; generalized anxiety disorder, 48 cases; panic disorder, 55 cases; specific phobia, 53 cases, a total of 181 cases) and a healthy control group (108 cases) collected by Nanjing Brain Hospital were used and converted into corresponding amplitude low-frequency fluctuation (ALFF) data as biological signal data for biological typing of anxiety disorder, which was divided into 4 categories. At the same time, the healthy control group was divided into 1 category. More specifically, in the experiment, voxel-based ALFF data was subjected to template matching according to a brain template (AAL-116 template) to generate ALFF biological signal data based on brain regions. The ALFF biological signal data of 90 brain regions of the brain was selected and input into the pseudo-label variational auto-encoding network of the present invention to achieve biological typing of anxiety disorder. Here, in the encoding network, the dimension of the ALFF biological signal data of each sample input was 90×1, which passed through two fully connected layers in sequence, and an activation function (Relu) was connected after each connection layer. After passing through the first fully connected layer, features of 50×1 were output, and after passing through the second fully connected layer, features of 20×1 were output. Subsequently, the 20×1 features passed through two fully connected layers respectively. The first fully connected layer output a sample mean feature vector with a dimension of 20×1, and the second fully connected layer output a sample variance feature vector with a dimension of 20×1 as well. In the decoding network, two fully connected layers were also set, and an activation function (Relu) was connected after each connection layer. After passing through the first fully connected layer, features of 50×1 were output, and after passing through the second fully connected layer, features of 90×1 were output to construct reconstructed sample biological signal data and typical biological signal data of each biological typing category. For the relevant settings of the classification sub-network, reference can be made to the corresponding classification sub-network structure involved in "ADHD classification using auto-encoding neural network and binary hypothesis testing" (published in the journal Artificial intelligence in medicien). However, since in this example experiment, the dimension of the input data of this classification sub-network was 20×1, the input dimension of the first layer in the above classification sub-network was correspondingly modified to meet the dimension requirements of the input data of this experiment.

[0050] 2) Experimental content

[0051] For Example 1, biological typing of anxiety disorder samples was performed, and between-group analysis was carried out with the healthy group. As Figure 2As shown, the method involved in the present invention effectively classifies anxiety disorders into 4 types, and each type has a clear explanation. In type I, compared with the healthy control group, the abnormal brain regions are mainly concentrated in the orbital prefrontal region, which is mainly responsible for the function of reward feedback (amplification or inhibition), and is related to the amplification of anxiety factors. In type II, compared with the healthy control group, the abnormal brain regions are mainly concentrated in the occipital visual region. This is related to specific phobias in anxiety disorders, manifested as visual abnormal stimuli for fear objects (such as spiders, snakes, etc.). In type III, compared with the healthy control group, the abnormal brain regions are mainly concentrated in the superior frontal gyrus. This region is related to the ability to withstand psychological pressure. The abnormal response in this region is closely related to anxiety disorders. In type IV, compared with the healthy control group, the abnormal brain regions are mainly concentrated in the limbic network, and brain regions such as the hippocampus, parahippocampal gyrus, amygdala, and temporal pole are all related to emotion control, while the most direct manifestation of anxiety disorders is strong emotional changes. At the same time, there is very little overlap in the abnormal brain regions among these four types, and each type has an independent explanation. The above results show that when applying the method of the present invention to the biological classification of anxiety disorders, relatively reliable analysis results can be obtained. This method is expected to have a profound impact in the field of biological classification of mental diseases.

[0052] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the method involved in the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A biological classification method for mental disorders based on a pseudo-label variational autoencoder network, characterized in that, Including the following steps: Step 1: Assume that all samples in the group of mental illness patients can be biologically classified into N categories, with corresponding labels from l1 to l N , and all samples in the healthy control group are divided into one category, with the corresponding label l N+1 , for a total of N + 1 categories. These categories respectively correspond to N + 1 sets, denoted as {S1,..., S N , S N+1}, where the sets S1 to S N correspond to the classification types l1 to l N of mental illness, and the S N+1 set corresponds to the healthy control group class l N+1 . Randomly initialize the pseudo-labels for each category, which consist of a class mean feature vector m li and a class variance feature vector v li . Each category can contain multiple samples, but at this time, each category is temporarily defined as an empty set, containing no samples; Step 2: For each sample in the group of patients with mental illness and the healthy control group, based on the brain region bio-signal x k , input it into the encoder sub-network module of the variational auto-encoder network to obtain the sample mean feature vector m k and the sample variance feature vector v k ; Step 3: If the sample is a patient with mental illness, then the sample mean feature vector m of this sample k is compared with the class mean feature vectors mli of various biological classifications of mental illness. The pseudo-label of the class with the most similar class mean feature vector to the sample mean feature vector of this sample is taken as the pseudo-label of this sample, and at the same time, this sample is added to the corresponding classification class l i in the corresponding set S i If the sample is a healthy sample, then the pseudo-label of this sample is set to the pseudo-label of the healthy control group class l N+1 and at the same time, this sample is added to the corresponding set S N+1 ; Step 4: For each sample, generate a high-level feature vector m k from the sample mean feature vector m k and the sample variance feature vector, as m k +βv k , denoted as z , where β is taken as a random number from a standard normal distribution; Step 5: Input the high-level feature vector z generated by each sample into the decoding sub-network module of the variational auto-encoder network, and this module outputs a corresponding reconstructed biological signal y k , which is input into the decoding sub-network module of the variational auto-encoder network, and this module outputs a corresponding reconstructed biological signal y k ; Step 6: Input the high-level feature vector z generated by each sample into a classification sub-network module, which outputs a classification label prediction result l for the input sample biological signal x k , to predict whether each sample comes from the biological classification class or the healthy control group class to which it belongs k ; k ​ Step 7: Input all samples into the pseudo-label variational autoencoder network to implement the addition of samples from S1 to S N+1 After adding the samples of the set, update various pseudo-labels, that is, the class mean feature vector m li and the class variance feature vector v li ; Step 8: Conduct one round of network training on the pseudo-label variational autoencoder network, and return to Step 1 to perform iterative calculations for Steps 1 to 8. However, at this time, the pseudo-labels of each category in Step 1 are replaced with the updated pseudo-labels of the corresponding categories. After multiple rounds of iteration, when the samples included in each category are the same as those included in each category in the previous round of iteration, the iteration ends; Step 9: The corresponding class mean feature vectors m in the pseudo-labels of the mental illness classification types and the healthy control group are li output, through the decoding sub-network module of the variational auto-encoder network, the corresponding biological signal y li , as the typical biological signals of each type, thereby determining the final biological characteristics of each type in the biological classification of mental illnesses.

2. The method for biological classification of mental diseases based on the pseudo-label variational auto-encoder network according to claim 1, wherein in step 3, When taking the pseudo-label of the class with the class mean feature vector being most similar to the sample mean feature vector as the pseudo-label of the sample, the class mean feature vector m li and the sample mean feature vector m k The similarity measurement criterion is taken as the minimum criterion of the mean method or the minimum criterion of cosine similarity.

3. The biological typing method for mental disorders based on the pseudo-label variational autoencoder network according to claim 1, in the step 7, it is characterized in that When the pseudo-label is updated, the class mean feature vector m of each class li is the sample mean feature vector m of the samples contained in this class k The average value, and the class variance feature vector v of each class li is the sample variance feature vector v of the samples contained in this class k The average value.

4. A method for biological typing of mental disorders based on a pseudo-label variational autoencoder network according to claim 1, wherein in Step 8, it is characterized in that the network loss function is set, and the loss function is measured by the following five metrics: 4.

1. In step 3, the relative entropy (KL divergence) measure of each category (each sample contained in each set S i is denoted as K1 for the measurement value; 4.

2. In step 5, the output biological signal y k is the similarity measure with the input biological signal x of the corresponding sample in step 2 k , and the measured value is denoted as K2; 4.

3. In step 6, the input sample biological signal x k and the classification label prediction result l k of this sample in step 3 are compared with the corresponding label l i of this sample, and a similarity measure is calculated, denoted as the measure value K3; 4.

4. In step 7, for each updated class mean feature vector m li , calculate the total Euclidean distance metric between pairwise vectors, and denote the metric value as K4; 4.

5. In step 8, for each type of updated class mean feature vector m li the corresponding typical biological signal y output by the decoding sub-network module of the variational auto-encoder network li , calculate the sum of the absolute values of the pairwise correlation values of the typical biological signals corresponding to all classes as a metric, and denote the metric value as K5; Finally, the loss function is K1 + K2 + K3 - K4 + K5, and this loss function is minimized during network training. The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the method involved in the present invention, several improvements and deformations can be made, and these improvements and deformations should also be regarded as the protection scope of the present invention.

Images