Mental disease prediction method based on reciprocal multi-network noise label learning model

Abstract

The invention discloses a mental disease prediction method based on a reciprocal multi-network noise label learning model, and belongs to the technical field of medical image processing. In order to solve the problems that inherent subjectivity exists in the mental disease classification process at present, and a wrong diagnosis label is generated due to the fact that pathology and neural mechanisms behind mental disorders generally do not completely meet clinical diagnosis standards, the method comprises the steps that neuroimaging data are utilized; and the diagnosis accuracy is improved under the condition that potential clinical diagnosis deviation or error exists. By maximizing the consistency of deep neural networks in identifying and utilizing subjects with accurate diagnostic tags and suspected diagnostic tags, the present method effectively utilizes multiple deep neural networks to learn from data of suspected diagnostic tags. According to the method, on the basis of the original diagnostic tag, the re-determined biotype shows more obvious inter-group differences, and more significant functional connections exist.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a method for predicting mental illness based on a reciprocal multi-network noise label learning model. Background Art

[0002] Currently, the diagnosis of mental disorders mainly relies on clinical interviews based on diagnostic criteria. However, due to the subjective nature of these assessments, ensuring the accuracy of the diagnosis can be challenging. In addition, the overlap of symptoms between different mental disorders may lead to inaccurate symptom-based diagnostic methods, increasing the risk of incorrect diagnostic labels. Although many studies have used neuroimaging measures to build classifiers to distinguish different mental disorders and explored biomarkers by treating diagnostic labels as category labels, the inaccuracy of diagnostic labels may affect the performance of classifiers and the reliability of biomarkers. In addition to the inherent subjectivity of current diagnostic methods, the pathological and neural mechanisms behind mental disorders often do not fully conform to clinical diagnostic criteria, such as the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) or the International Classification of Diseases (ICD-11).

[0003] Therefore, neuroimaging-based classifiers trained using these diagnostic labels may encounter noise or inconsistencies as the models are forced to map neurobiological patterns onto labels that may not fully reflect disease mechanisms. This discrepancy between clinical classifications and biological reality makes disease classification and biomarker discovery particularly challenging, highlighting the urgent need to develop methods that focus on biological information to reduce the negative impact of incorrect diagnostic labels on model training and biomarker discovery, thereby better aligning classification with the underlying biological processes of mental disorders. Summary of the invention

[0004] In view of the inherent subjectivity in the current classification of mental illnesses, and the fact that the pathological and neural mechanisms behind mental disorders usually do not fully meet the clinical diagnostic criteria, leading to the generation of incorrect diagnostic labels, the present invention provides a reciprocal multi-network noise label learning model for the classification of mental illnesses. The model can classify and predict subjects using brain imaging data, and can further analyze the prediction results obtained based on the training model to assist doctors or professionals in diagnosing and treating mental illnesses. This method is expected to provide new ideas for neuroscience research and medical diagnosis of brain diseases.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The method for predicting mental illness based on a reciprocal multi-network noise label learning model includes the following steps:

[0007] Step 1, construct a training dataset for the reciprocal multi-network noisy label learning model;

[0008] Step 1.1, first preprocess the brain image data of the subject to obtain the standard brain image of the subject, use the brain image analysis method to estimate the brain image measurement from the brain image data, and extract the brain image features;

[0009] Step 1.2: Construct a matrix of brain image features for all subjects

[0010] Where N represents the number of subjects, D represents the number of brain imaging features of each subject; x i represents the characteristics of the i-th subject, is the i-th row of the X matrix;

[0011] Step 1.3: Construct the diagnostic label matrix corresponding to the subject

[0012] Among them, C represents the number of disease categories of the subjects, y i represents the diagnostic label of the ith subject, is the i-th row of the Y matrix;

[0013] Step 1.4, set two empty subject subscript sets S Clean and S Noisy ;

[0014] Among them, S Clean Used to store the subscript of the subject with the correct diagnostic label, S Noisy Used to store the subscripts of subjects whose diagnostic labels are questionable;

[0015] Step 2, training a reciprocal multi-network noisy label learning model;

[0016] Step 2.1, initialize two neural networks θ1 and θ2, and two mean teacher neural networks corresponding to the two neural networks θ1 and θ2 and

[0017] Step 2.2: Input the brain image feature matrix X and the corresponding diagnostic label matrix Y of the subjects in the training data set into the two neural networks θ1 and θ2 of the reciprocal multi-network noise label learning model; the neural networks θ1 and θ2 output the prediction results f(x i ,θ1) and f(x i ,θ2), i∈(1,2,…,N);

[0018] Calculate the classification loss term for each subject

[0019]

[0020] Calculate the consistency loss of the two neural networks' predictions for each subject

[0021]

[0022] The classification loss And the consistency loss term The weighted sum of ∫ constitutes the joint loss for each subject

[0023]

[0024] Among them, λ1 is a weight parameter used to balance the classification loss term And the consistency loss term

[0025] Step 2.3: The joint loss term for each subject in the training dataset is The values ​​of are sorted in ascending order; the subjects whose joint loss items are lower than the set threshold are identified as subjects with accurate diagnostic labels, and their subscripts i are recorded in the subscript set S of subjects with accurate diagnostic labels. Clean In the above example, the subjects whose joint loss items are higher than the set threshold are identified as subjects with questionable diagnostic labels, and their subscript i is recorded in the subscript set S of subjects with questionable diagnostic labels. Noisy middle;

[0026] Step 2.4: Subscript the set S of subjects with accurate diagnostic labels Clean The brain image features of the corresponding subjects are input into two mean teacher neural networks and The prediction results are obtained and

[0027] Calculate the soft distribution loss term for each subject with accurate diagnostic label

[0028]

[0029] The soft distribution loss term With joint loss Sum the total loss for each subject with an accurate diagnosis label. Then sum the total loss for all subjects with accurate diagnosis labels to calculate the overall clean loss term |S Clean | indicates S CleanThe number of accurate subject subscripts of diagnostic labels stored in ;

[0030]

[0031] Step 2.5: Subscript the subjects with questionable diagnostic labels into the set S Noisy The corresponding brain image features of the subjects are input into two mean teacher neural networks and The prediction results are obtained and

[0032] Calculate the soft distribution loss term for each subject whose diagnostic label is questionable

[0033]

[0034] Soft distribution loss term And the consistency loss term Sum the total loss for each subject with a questionable diagnostic label. Then sum the total losses of all subjects with questionable diagnostic labels to calculate the overall noise loss term |S Noisy | indicates S Noisy The number of subscripts of subjects with questionable diagnostic labels stored in;

[0035]

[0036] Step 2.6, calculate the sum of losses for all subjects

[0037]

[0038] Among them, λ2 is a weight parameter used to balance the loss of subjects with accurate diagnostic labels Loss of subjects with questionable diagnostic labels

[0039] Step 2.7, use the gradient descent method to update the parameters of the two neural networks θ1 and θ2, and use the exponential moving average method to update their corresponding mean teacher neural network and Parameters; if the total loss of the model Converge and get a trained model; if the total loss of the model If it does not converge, go to step 2.2 and repeat until it converges;

[0040] Step 3: Preprocess and extract brain image features for subjects with unknown diagnostic labels; then input them into the trained reciprocal multi-network noise label learning model for analysis and output the predicted results of their mental illness categories.

[0041] Furthermore, the preprocessing in step 1 specifically includes inter-layer correction, head motion correction, spatial normalization and spatial smoothing.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] 1) This invention proposes a new method to address the inherent subjectivity of current diagnostic methods and the incorrect diagnostic labels caused by the fact that the pathological and neural mechanisms behind mental disorders usually do not fully meet the clinical diagnostic criteria, which leads to low classification accuracy.

[0044] 2) The present invention not only uses the brain image features and diagnostic labels of subjects with accurate diagnostic labels to explore the mapping pattern between brain image features and diagnostic labels, but also uses the brain image features of subjects with questionable diagnostic labels to assist in this exploration process.

[0045] 3) The present invention can reduce the negative impact of subjects with questionable diagnostic labels on model training, thereby obtaining a model that can more accurately identify the type of disease in the subject, assisting doctors or professionals in diagnosing and treating mental illness. This will provide new research ideas for fields such as neuroscience research and medical diagnosis of brain diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of a method for predicting mental illness based on a reciprocal multi-network noise label learning model of the present invention;

[0047] Figure 2 Schematic diagram of the T-value results of the two-sample t-test between different categories of functional connectivity data;

[0048] Figure 3 is the average connection strength of each individual in the functional connectivity data. DETAILED DESCRIPTION

[0049] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0050] Example 1

[0051] The method for predicting mental illness based on a reciprocal multi-network noise label learning model includes the following steps:

[0052] Step 1, construct a training dataset for the reciprocal multi-network noisy label learning model;

[0053] Step 1.1, brain imaging data from patients with different mental illnesses and normal subjects were collected from the site, and preprocessing operations including inter-layer correction, head motion correction, spatial normalization and spatial smoothing were performed on these data to obtain standard brain images of the subjects, and brain imaging analysis methods were used to estimate brain imaging measures from the brain imaging data, and feature selection or feature extraction methods were used to extract features from the brain imaging measures;

[0054] Step 1.2: Construct a matrix of brain image features for all subjects

[0055] Where N represents the number of subjects, D represents the number of brain imaging features of each subject; x i represents the characteristics of the i-th subject, is the i-th row of the X matrix;

[0056] Step 1.3: Construct the diagnostic label matrix corresponding to the subject

[0057] Among them, C represents the number of disease categories of the subjects, y i represents the diagnostic label of the ith subject, is the i-th row of the Y matrix;

[0058] Step 1.4, set two empty subject subscript sets S Clean and S Noisy ;

[0059] Among them, S Clean Used to store the subscript of the subject with the correct diagnostic label, S Noisy Used to store the subscripts of subjects whose diagnostic labels are questionable;

[0060] Step 2, training a reciprocal multi-network noisy label learning model;

[0061] Step 2.1, initialize two neural networks θ1 and θ2, and two mean teacher neural networks corresponding to the two neural networks θ1 and θ2 and

[0062] Step 2.2: Input the brain image feature matrix X and the corresponding diagnostic label matrix Y of the subjects in the training data set into the two neural networks θ1 and θ2 of the reciprocal multi-network noise label learning model; the neural networks θ1 and θ2 output the prediction results f(x i ,θ1) and f(x i ,θ2), i∈(1,2,…,N);

[0063] Calculate the classification loss term for each subject

[0064]

[0065] Calculate the consistency loss of the two neural networks' predictions for each subject

[0066]

[0067] The classification loss And the consistency loss term The weighted sum of ∫ constitutes the joint loss for each subject

[0068]

[0069] Among them, λ1 is a weight parameter used to balance the classification loss term And the consistency loss term

[0070] Step 2.3: The joint loss term for each subject in the training dataset is The values ​​of are sorted in ascending order; the subjects whose joint loss items are lower than the set threshold are identified as subjects with accurate diagnostic labels, and their subscripts i are recorded in the subscript set S of subjects with accurate diagnostic labels. Clean In the above example, the subjects whose joint loss items are higher than the set threshold are identified as subjects with questionable diagnostic labels, and their subscript i is recorded in the subscript set S of subjects with questionable diagnostic labels. Noisy middle;

[0071] Step 2.4: Subscript the set S of subjects with accurate diagnostic labels Clean The brain image features of the corresponding subjects are input into two mean teacher neural networks and The prediction results are obtained and

[0072] Calculate the soft distribution loss term for each subject with accurate diagnostic label

[0073]

[0074] The soft distribution loss term With joint loss Sum the total loss for each subject with an accurate diagnosis label. Then sum the total loss for all subjects with accurate diagnosis labels to calculate the overall clean loss term |S Clean | indicates S Clean The number of accurate subject subscripts of diagnostic labels stored in ;

[0075]

[0076] Step 2.5: Subscript the subjects with questionable diagnostic labels into the set S Noisy The corresponding brain image features of the subjects are input into two mean teacher neural networks and The prediction results are obtained and

[0077] Calculate the soft distribution loss term for each subject whose diagnostic label is questionable

[0078]

[0079] Soft distribution loss term And the consistency loss term Sum the total loss for each subject with a questionable diagnostic label. Then sum the total losses of all subjects with questionable diagnostic labels to calculate the overall noise loss term |S Noisy | indicates S Noisy The number of subscripts of subjects with questionable diagnostic labels stored in;

[0080]

[0081] Step 2.6, calculate the sum of losses for all subjects

[0082]

[0083] Among them, λ2 is a weight parameter used to balance the loss of subjects with accurate diagnostic labels Loss of subjects with questionable diagnostic labels

[0084] Step 2.7, use the gradient descent method to update the parameters of the two neural networks θ1 and θ2, and use the exponential moving average method to update their corresponding mean teacher neural network and Parameters; if the total loss of the model Converge and get a trained model; if the total loss of the model If it does not converge, go to step 2.2 and repeat until it converges;

[0085] Step 3: Preprocess and extract brain image features for subjects with unknown diagnostic labels; then input them into the trained reciprocal multi-network noise label learning model for analysis and output the predicted results of their mental illness categories.

[0086] The reciprocal multi-network noise label learning model for predicting mental illness of the present invention was experimented on the brain imaging data of bipolar disorder subjects and schizophrenia subjects in the Bipolar-Schizophrenia Network on Intermediate Phenotypes dataset. The method was evaluated by comparing the inter-group differences of the original diagnostic label group and the inter-group differences of the two groups after the original diagnostic group was re-predicted using the model trained by the method, as well as the inter-group differences between the original diagnostic group and healthy people and the inter-group differences between the re-predicted group and healthy people, and the number of significant functional connections of the above comparison groups. Figure 2 As shown, (af) are the T-value results of the two-sample t-test between different categories of functional connectivity data, which are used to represent the group differences between different categories. In (af), 53 functional networks are divided into seven functional domains, including subcortical (SC), auditory (AU), sensorimotor (SM), visual (VI), cognitive control (CC), default mode (DM), and cerebellum (CB) domains. (a) is the group difference between bipolar disorder subjects (BP) and schizophrenia subjects (SZ). (b) is the group difference between healthy controls (HC) and BP. (c) is the group difference between HC and SZ. (d) is the group difference between the two groups (Biotype 1 and Biotype 2) after the original diagnostic group is re-predicted using the model trained by this method. (e) is the group difference between Biotype1 and HC. (f) is the group difference between Biotype 2 and HC. This suggests that the differences between the re-diagnosed groups Biotype 1 and Biotype 2, and between HC and each Biotype, were more pronounced than those between BP and SZ, HC and BP, and HC and SZ. Figure 3 As shown, Figure 3(ac) are the mean connectivity strengths of individuals in each functional connectivity data set with p-values ​​< 0.005 (without correction for multiple comparisons) in BP and SZ using two-sample t-tests. (df) are the mean connectivity strengths of individuals in each functional connectivity data set with p-values ​​< 0.005 (with Bonferroni correction) in Biotype 1 and Biotype 2 using two-sample t-tests. Figure 3 The number of functional connections showing significant differences between the re-diagnosis groups Biotype 1 and Biotype 2 was much greater than that between BP and SZ.

[0087] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for predicting mental illness based on a reciprocal multi-network noise label learning model, characterized in that: Step 1, construct a training dataset for the reciprocal multi-network noisy label learning model; Step 1.1, first preprocess the brain image data of the subject to obtain the standard brain image of the subject, use the brain image analysis method to estimate the brain image measurement from the brain image data, and extract the brain image features; Step 1.2: Construct a matrix of brain image features for all subjects Where N represents the number of subjects, D represents the number of brain imaging features of each subject; x i represents the characteristics of the i-th subject, is the i-th row of the X matrix; Step 1.3: Construct the diagnostic label matrix corresponding to the subject Among them, C represents the number of disease categories of the subjects, y i represents the diagnostic label of the ith subject, is the i-th row of the Y matrix; Step 1.4, set two empty subject subscript sets S Clean and S Noisy ; Among them, S Clean Used to store the subscript of the subject with the correct diagnostic label, S Noisy Used to store the subscripts of subjects whose diagnostic labels are questionable; Step 2, training a reciprocal multi-network noisy label learning model; Step 2.1, initialize two neural networks θ1 and θ2, and two mean teacher neural networks corresponding to the two neural networks θ1 and θ2 and Step 2.2: Input the brain image feature matrix X and the corresponding diagnostic label matrix Y of the subjects in the training data set into the two neural networks θ1 and θ2 of the reciprocal multi-network noise label learning model; the neural networks θ1 and θ2 output the prediction results f(x i ,θ1) and f(x i ,θ2), i∈(1,2,…,N); Calculate the classification loss term for each subject i∈(1,2,…,N); Calculate the consistency loss of the two neural networks' predictions for each subject i∈(1,2,…,N); The classification loss And the consistency loss term The weighted sum of ∫ constitutes the joint loss for each subject i∈(1,2,…,N); Among them, λ1 is a weight parameter used to balance the classification loss term And the consistency loss term Step 2.3: The joint loss term for each subject in the training dataset is The values ​​of are sorted in ascending order; the subjects whose joint loss items are lower than the set threshold are identified as subjects with accurate diagnostic labels, and their subscripts i are recorded in the subscript set S of subjects with accurate diagnostic labels. Clean In the above example, the subjects whose joint loss items are higher than the set threshold are identified as subjects with questionable diagnostic labels, and their subscript i is recorded in the subscript set S of subjects with questionable diagnostic labels. Noisy middle; Step 2.4: Subscript the set S of subjects with accurate diagnostic labels Clean The brain image features of the corresponding subjects are input into two mean teacher neural networks and The prediction results are obtained and i∈S Clean ; Calculate the soft distribution loss term for each subject with an accurate diagnostic label i∈S Clean ; The soft distribution loss term With joint loss Sum the total loss for each subject with an accurate diagnosis label. Then sum the total loss for all subjects with accurate diagnosis labels to calculate the overall clean loss term |S Clean | indicates S Clean The number of accurate subject subscripts of diagnostic labels stored in ; Step 2.5: Subscript the subjects with questionable diagnostic labels into the set S Noisy The corresponding brain image features of the subjects are input into two mean teacher neural networks and The prediction results are obtained and i∈S Noisy ; Calculate the soft distribution loss term for each subject whose diagnostic label is questionable i∈S Noisy ; Soft distribution loss term And the consistency loss term Sum the total loss for each subject with a questionable diagnostic label. Then sum the total losses of all subjects with questionable diagnostic labels to calculate the overall noise loss term |S Noisy | indicates S Noisy The number of subscripts of subjects with questionable diagnostic labels stored in; Step 2.6, calculate the sum of losses for all subjects Among them, λ2 is a weight parameter used to balance the loss of subjects with accurate diagnostic labels Loss of subjects with questionable diagnostic labels Step 2.7, use the gradient descent method to update the parameters of the two neural networks θ1 and θ2, and use the exponential moving average method to update their corresponding mean teacher neural network and Parameters; if the total loss of the model Converge and get a trained model; if the total loss of the model If it does not converge, go to step 2.2 and repeat until it converges; Step 3: Preprocess and extract brain image features for subjects with unknown diagnostic labels; then input them into the trained reciprocal multi-network noise label learning model for analysis and output the predicted results of their mental illness categories.

2. The method for predicting mental illness based on a reciprocal multi-network noise label learning model according to claim 1, characterized in that: The preprocessing in step 1 specifically includes inter-layer correction, head motion correction, spatial normalization and spatial smoothing.

Images