A method for establishing a classification model for brain imaging data based on partial label learning using neighbor propagation

Abstract

The application belongs to the technical field of brain image analysis, and discloses a brain image data classification model establishment method based on a near neighbor propagation bias label learning, which comprises the following steps: S1, acquiring a training data set and corresponding original labels, solving correlation coefficients between samples and degrees of centrality of the samples, and generating a typical sample set and a category center set according to the correlation coefficients and the degrees of centrality; S2, determining candidate labels of training samples, and constructing a candidate label set; S3, performing label propagation by using a K+N near neighbor sample graph through the training data set, and establishing a classification model. The application can improve the accuracy and precision of the classification model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of brain image analysis and the field of partial label learning, and particularly relates to a new brain image data classification model establishment method based on a near neighbor propagation partial label learning. BACKGROUND

[0002] Traditional mental illness classification is faced with problems such as high overlap of clinical symptoms of multiple diseases, and imperfect diagnosis method based on symptoms, and thus needs to be improved.

[0003] At present, the classification of mental illness is mainly based on the subjective description of patients on their symptoms and the observation of the behavior of patients by clinicians or informants. However, many mental illnesses have similar clinical symptoms, and mental illness patients often have clinical features of multiple diseases at the same time, and the main clinical symptoms of a specific individual can also change significantly over time, which greatly interferes with the classification of mental illness. Therefore, the clinical classification results obtained by physician assessment or patient self-assessment often have the disadvantage of low accuracy due to the lack of biological evidence.

[0004] The use of brain image data can improve the classification accuracy of mental illness. However, the theoretical premise of the classification algorithm in the prior art is that the data and its corresponding label are reliable. The method often learns information from the training data and establishes a model, and then completes the classification of the class of new samples through the trained model. However, many data sets have the problem of incomplete and inaccurate label information. Artificial annotation work requires high professional accomplishment of the annotator, but in reality, such annotation requirements are often difficult to meet, so part of the data has the problem of label ambiguity. For example, in the field of mental illness, the symptoms of schizophrenia and bipolar disorder overlap, and doctors have difficulty in giving reliable diagnosis of the two diseases, and the disease label corresponding to the brain image data of the mental illness patient has the problem of low accuracy, so the traditional classification method is not suitable for brain image classification work in the field of mental illness. The classification based on partial label learning is a method of extracting effective information from incomplete and inaccurate label information, and is more suitable for brain image data classification.

[0005] Therefore, it is necessary to improve the classification model establishment method in the prior art to obtain a brain image classification model with high accuracy and improve the classification accuracy of the classification model. SUMMARY

[0006] The present application overcomes the deficiencies in the prior art and solves the technical problem of providing a brain image data classification model establishment method based on near neighbor propagation partial label learning.

[0007] In order to solve the above technical problems, the technical scheme adopted by the present application is as follows: a brain image classification model establishment method based on near neighbor propagation partial label learning, comprising the following steps:

[0008] S1, acquire a training data set and corresponding original labels; calculate a similarity matrix and a degree centrality between samples in the same category of the training data set, determine typical samples of each category according to the degree centrality, then take the mean of all typical samples in the same category group as the initial category center of the category; construct a typical sample set according to the typical samples under each category, and construct a category center set according to the initial category center of each category;

[0009] S2, determine candidate labels of samples in the training data set, and construct a candidate label set;

[0010] S3, establish a classification model by using K+N neighbor sample graph through the training data set, which specifically includes:

[0011] S301, take the typical sample and its K neighbor node samples as the samples of the current layer, and determine the candidate labels of the neighbor node samples;

[0012] S302, construct a K+N neighbor sample graph of the samples of the current layer according to the typical sample set, the category center and the sample features of the training samples; the K+N neighbor sample graph includes K neighbor sample nodes and N neighbor typical sample nodes;

[0013] S303, infer the labels of the neighbor node samples of the typical sample, determine the classification output labels of the neighbor node samples, then calculate the degree centrality of each sample in the same category group, increase the number of typical samples of each category according to the degree centrality, and update the typical sample set, the category center set and the candidate label set;

[0014] S304, determine whether all sample labels are determined, if not, return to step S301; if yes, output the final classification output labels of all samples, and complete the establishment of the classification model.

[0015] The brain image data includes brain function magnetic resonance imaging, brain structure magnetic resonance data, diffusion tensor imaging data, electroencephalogram data and magnetoencephalogram data.

[0016] The features of the brain image data include brain function connection features, brain gray matter density and volume features, brain structure connection features, electroencephalogram signal features and features after multi-modal fusion of brain image data.

[0017] In step S1, the matrix element of the similarity matrix is a correlation coefficient, and the correlation coefficient is a Pearson correlation coefficient, a Gaussian kernel solving correlation coefficient or a regression solving correlation coefficient.

[0018] In step S303, the method for determining the classification output labels of the neighbor samples is:

[0019] Cyclically calculating a temporary confidence matrix and a label confidence matrix F (t) , and the calculation formula is:

[0020]

[0021]

[0022]

[0023] Wherein, t represents the number of propagation; F (0) represents the initial label confidence matrix, and the average of the number of candidate labels in the candidate label set of the sample is taken as the initial label confidence of the current sample; α is the weight, W K represents the K nearest neighbor weight matrix, W N represents the N nearest neighbor weight matrix; F (t-1) represents the label confidence matrix obtained in the t-1th propagation is a temporary confidence matrix element in the temporary confidence matrix element in the label confidence matrix F (t) ; represents the temporary confidence that the classification output label of the i th sample in the candidate label set is the class l, y l represents the label of the class l, S i represents the candidate label set, n represents the number of training samples in the layer, and q represents the number of classes.

[0024] When the convergence condition is reached, the cycle is stopped, and the label corresponding to the maximum confidence value in the label confidence matrix is taken as the label of the nearest neighbor sample.

[0025] In the step S1, the degree centrality is the total sum of sample similarity;

[0026] The method for determining the typical sample of each class is that: the degree centrality values of each sample are sorted, and A% of the samples with high degree centrality in the same class group are selected as the typical samples.

[0027] In the step S303, the specific method for increasing the number of typical samples of each class according to the degree centrality is:

[0028] The degree centrality values of each sample in the same class group are sorted, and B% of the non-typical samples with high degree centrality in the same class group are selected as new typical samples and added to the typical sample set, wherein B is greater than A.

[0029] In the step S2, the specific method for determining the candidate label of the training set sample is:

[0030] Identify the original label of a typical sample as a classification output label;

[0031] Calculate the similarity of the remaining samples to each class center, and select the label corresponding to the class center with the highest similarity as a candidate label, and add it to the candidate label set of the current sample together with the original label.

[0032] In step S302, the construction method of the K+N neighbor sample graph structure of the current layer sample is:

[0033] Based on the typical sample set, class center and sample feature of the training sample, a full connection original graph of the training sample is constructed;

[0034] The full connection original graph is sparsified twice to obtain a K neighbor sample graph and an N neighbor typical sample graph;

[0035] The K neighbor sample graph and the N neighbor typical sample graph are superimposed to obtain a K+N neighbor sample graph of the current layer sample.

[0036] The application provides a brain image data classification model establishment method based on near neighbor propagation partial label learning, which is different from existing classification algorithms, avoids excessive dependence of current mental illness diagnosis classification algorithms on the diagnosis results of clinical doctors, combines sample label propagation with graph structure, establishes a model, and has the following effects compared with the prior art:

[0037] 1. The application measures the relationship between samples based on a K+N neighbor sample graph, effectively utilizes the information of more reliable samples with partial label learning algorithm, and improves the accuracy of sample classification.

[0038] 2. There are few studies on partial label classification auxiliary mental illness diagnosis, and the partial label algorithm defaults the initial label as any label in the candidate label set, which conforms to the current situation that the high overlap of mental illness causes the disease to be difficult to accurately diagnose.

[0039] 3. The K+N neighbor sample graph generated by the application has nodes representing samples and weight edges reflecting the similarity between a pair of samples. The K+N neighbor sample graph structure conforms to the characteristics of mental illness data and is suitable for mining effective information of mental illness data.

[0040] 4. Compared with the traditional partial label algorithm, the application has higher classification accuracy, and therefore has more advantages. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 It is a flowchart of a brain image classification model establishment method based on near neighbor propagation partial label learning provided by the embodiment of the application.

[0042] Figure 2 is a K+N neighbor sample graph generation process schematic diagram based on the present application of a near neighbor propagation-based partial label classification algorithm.

[0043] Figure 3 is a accuracy performance evaluation result graph of the present application of a near neighbor propagation-based partial label classification algorithm under five data sets and six partial labels.

[0044] Figure 4 is a precision performance evaluation result graph of the present application of a near neighbor propagation-based partial label classification algorithm under five data sets and six partial labels. DETAILED DESCRIPTION

[0045] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] As shown in Figure 1 The present application provides a brain image classification model establishment method based on near neighbor propagation partial label learning, and the most core is a near neighbor propagation partial label classification algorithm. When the model of the present application is established, the training set samples are used to train the classification model, the over-reliance of the current mental illness diagnosis classification algorithm on the diagnosis results of the clinical doctors is broken, the training set sample labels are re-determined, and a K+N neighbor sample graph model with higher accuracy is obtained through the accuracy of the training samples. Using the model, the labels of new samples can be more accurately predicted and judged.

[0047] S1, obtain a training data set and corresponding original labels, solve the correlation coefficient between samples and the degree centrality of the samples, and generate an initial typical sample set and an initial category center set accordingly. The typical sample refers to a sample with reliable label and obvious features, therefore, in this embodiment, part of the samples with high degree centrality in each original label category are selected as typical samples. In the same category sample of the training data set, the correlation coefficient between samples is calculated as the similarity, the degree centrality is solved based on the similarity matrix, the samples are sorted according to the degree centrality, and the samples ranked in the front A% are selected as typical samples. Then, the mean of all typical samples in the same category group is taken as the initial category center of the category. The initial typical sample set is composed of the typical samples of the original categories, and the initial category center set is composed of the category centers.

[0048] Specifically, in this embodiment, the brain imaging data includes functional magnetic resonance imaging (FMRI) data, structural magnetic resonance imaging (MRI) data, diffusion tensor imaging (DTI) data, electroencephalography (EEG) data, and magnetic brain data. The data label is the category of the mental illness corresponding to the data.

[0049] Specifically, in this embodiment, the characteristics of the brain imaging data include: brain functional connectivity characteristics, brain gray matter density and volume, brain structural connectivity characteristics, EEG signal characteristics, and characteristics after multimodal fusion of brain images.

[0050] Specifically, in this embodiment, within samples of the same category in the training set, the correlation coefficient between samples is calculated as similarity, and a similarity matrix is ​​constructed to calculate degree centrality. The sum of the sample similarities is used as the degree centrality for sorting. The samples in the top A% of the degree centrality sort within the category group are selected as representative samples, and the mean of all representative samples within the category group is used as the category center. The initial representative sample set is formed based on the representative samples of the original category, and the initial category center set is formed based on the category centers.

[0051] In step S1, the correlation coefficient is the Pearson correlation coefficient, the correlation coefficient is solved by Gaussian kernel or regression.

[0052] S2. Determine the candidate labels of the samples in the training set and construct the candidate label set S i In step S2, the specific method for determining the candidate labels of the remaining samples in the training set is:

[0053] Typical samples are samples with reliable labels and distinct features. Therefore, the original labels of typical samples are used as the classification output labels. The candidate label set for typical samples does not introduce candidate label noise, maintaining the unique certainty of the labels in the candidate label set. In addition to the typical samples, candidate labels are determined for the remaining instance samples to ultimately generate a candidate label set. The similarity between a sample and each category center indicates the likelihood that the sample belongs to that label. The similarity between the current sample and each category center is calculated. In addition to the original label, the label corresponding to a category center with a high similarity is selected as a candidate label and added to the candidate label set for the current sample along with the original label. Finally, candidate labels are determined for all training samples to generate a candidate label set.

[0054] In this embodiment, based on the category center and the similarity relationship between the instance sample and the category center solved in step S1, the candidate label of each sample is determined, and the candidate label set S is constructed. i The higher the similarity between an instance and the category center, the more likely that category will be the classification output label for that instance. Based on the similarity relationship, the label with the highest similarity, excluding the initial label, is selected as the candidate label to construct the candidate label set for the current instance, ultimately forming the candidate label set for all training instances.

[0055] S3. Iteratively complete the construction of the K+N nearest neighbor sample graph model of the training sample and the label inference process of the training sample, where the K value represents the number of nearest neighbor sample nodes in the K+N nearest neighbor sample graph structure, and the N value represents the number of nearest typical sample nodes. The K value and the N value are set based on the experimental results of multiple data sets and the characteristics of brain imaging data. Specifically, in the embodiment of the present invention, the K value range is 8-30, and the N value range is 3-8.

[0056] The iterative construction of the K+N nearest neighbor sample graph model for the training samples involves determining the samples in the current layer and their candidate labels, establishing a K+N nearest neighbor sample graph for the samples in the current layer, performing label propagation on the samples, disambiguating the candidate label set, and determining their classification output labels. The K+N nearest neighbor sample graph model training is completed in a layered and iterative manner. The specific operation method is as follows:

[0057] S301. Determine the samples in this layer and their candidate labels. The samples in this layer are all samples involved in this iteration and are the set of samples with determined labels and samples to be propagated in this layer. The samples to be propagated in this layer are samples to be disambiguated from the candidate label set in this iteration. In this embodiment, neighboring node samples of samples with determined labels are selected as samples to be propagated. After determining the samples in this layer, candidate labels for these samples are extracted from the candidate label set and used for label inference for these samples.

[0058] S302: Construct a K+N nearest neighbor sample graph for samples in this layer. In this embodiment, when constructing the K+N nearest neighbor sample graph, for each sample, its connection with the K nearest neighbor sample nodes is retained, while its connection with the nodes corresponding to the N nearest typical samples is also retained. These two connections are superimposed to form the K+N nearest neighbor sample graph.

[0059] like Figure 2 As shown, in this embodiment, the method for constructing the K+N nearest neighbor sample graph structure of the samples in this layer in S302 is:

[0060] First, based on the typical sample set, category centers, and sample features of the training samples, a fully connected original graph of the training samples is constructed. The fully connected original graph is then thinned twice to obtain a K-nearest neighbor sample graph and an N-nearest neighbor typical sample graph. The K-nearest neighbor sample graph is an undirected asymmetric graph generated by thinning the original graph by retaining the weighted edges between each node and its K nearest neighbors. The N-nearest neighbor typical sample graph is an undirected asymmetric graph generated by thinning the original graph by retaining the weighted edges between each node and N typical sample nodes. The K-nearest neighbor sample graph and the N-nearest neighbor typical sample graph are superimposed to obtain the K+N nearest neighbor sample graph of the sample.

[0061] The application considers not only the information amount of K neighbor nodes, but also the information amount of N neighbor typical samples when generating a graph structure, so that the sample can absorb more effective information as much as possible during the label iteration process. The application defines this graph result as a K+N neighbor sample graph structure. In the K+N neighbor sample graph, the node represents the sample, and the weighted edge reflects the similarity between a pair of nodes. The K neighbor weighted matrix is used The similarity relationship between each node and the neighbor node is recorded, and n represents the number of samples in the current layer. The N neighbor weighted matrix is used The similarity relationship between each node and the neighbor typical sample node is recorded. Based on the partial label learning algorithm and the K+N neighbor sample graph structure, the application re-measures the relationship between the subjects, re-labels the subjects with low label reliability, and can improve the accuracy of sample label matching. The selection of K and N parameters of the NNPL classification model is based on data and prior knowledge settings. In the example experiment, K+N neighbor sample graph classification model is established using K=10 and N=5 values for test set label inference.

[0062] S303, label inference is performed on the samples in the current layer, the candidate labels are disambiguated, the classification output labels of the samples are determined, and the typical sample set, the class center set and the candidate label set are updated. In this step, the final label of the sample in the current layer is determined according to the current K+N neighbor sample graph. In addition, after determining the final label of the sample in the current layer, the typical sample set and the class center set need to be updated. In this embodiment, the top B% of non-typical samples in the degree centrality ranking in the same class group are selected as new typical samples, the typical sample set is updated, and the typical samples belonging to the same class are averaged to update the class center set. According to the updated class center set, the candidate labels of the samples in the candidate label set are updated.

[0063] In this embodiment, the specific operation steps of the iteration propagation in S303 are as follows:

[0064] After determining the samples in the current layer, label propagation is performed on the samples in the current layer, the candidate label set of the samples in the current layer is disambiguated, and the classification output labels of the samples in the current layer are determined. In this embodiment, the top B% of non-typical samples in the degree centrality ranking in the same class group are selected as new typical samples, the typical sample set is updated, and the typical samples belonging to the same class are averaged to update the class center set. According to the updated class label set, the candidate labels of the samples in the candidate label set are updated.

[0065] In this embodiment, the specific operation steps of the label propagation in S303 to determine the classification output labels of the samples in the current layer are as follows:

[0066] After generating the K+N neighbor sample graph of the samples in the current layer, iteration label propagation is performed on all samples for label inference. Through the label propagation, the K neighbor sample graph G Kand N nearest neighbor typical sample graph G N The two undirected asymmetric graphs disambiguate the candidate label set, and finally complete the label inference process, determine the classification output labels of the samples in the candidate label set, and complete the disambiguation of the candidate label set.

[0067] Determine the sample set of this layer and the K nearest neighbor sample graph G K and N nearest neighbor typical sample graph G N After that, label propagation is started to complete label inference of all samples in the label set. Label inference is achieved by determining the label confidence matrix F = [f i，c ] n×q F is a matrix composed of the confidence of each category of the sample classification output label. i，c Indicates the confidence that the classification output label of the i-th sample is category C, n represents the number of samples in this layer, and q represents the number of categories. The initial label confidence matrix F (0) =[f (0) i，c ] n×q , obtained by solving formula (1), where y c Represents the label of category C.

[0068]

[0069] The label confidence matrix is ​​iterated according to formula (2). t represents the tth propagation, represents the temporary confidence matrix obtained by the t-th propagation, W K represents the K nearest neighbor weight matrix, W N Represents the N nearest neighbor weight matrix. α∈(0,1) represents the influence weight of the result of the t-1th iteration on the tth iteration, and (1-α) reflects the influence weight of the initial label confidence matrix on the tth iteration.

[0070]

[0071] Then middle After processing by formula (3), it is ensured that the sum of the elements of the confidence vector corresponding to each sample is still 1. Finally, the confidence matrix F is obtained (t) , used for the next iteration.

[0072]

[0073] Among them, t represents the number of transmissions, F (0) represents the initial label confidence matrix, α is the weight, W K represents the K nearest neighbor weight matrix, W N represents the N nearest neighbor weight matrix; F (t-1)denotes the label confidence matrix obtained at the t-1th propagation; denotes an element in the temporary confidence matrix denotes an element in the temporary confidence matrix denotes an element in the label confidence matrix F (t) denotes an element in the label confidence matrix F denotes the confidence that the classification output label of the i-th sample in the candidate label set is class 1, y l denotes the label of class l, S i denotes the candidate label set, n denotes the number of training samples at the current layer, and q denotes the number of classes.

[0074] When the convergence condition is reached, the cycle is stopped, and the label corresponding to the maximum confidence value in the confidence matrix is taken as the label of the sample. After obtaining the final confidence matrix, the label with the largest weight in the candidate label set is selected as the final label of the current iteration of the current sample, and the disambiguation of the candidate label set is completed. In this embodiment, the convergence condition is that the difference between the label confidences in two iterations is less than a set threshold.

[0075] S304, it is judged whether the iteration of all training samples is completed. If the iteration is not completed, the iteration of S301 to S303 is continued. If the iteration is completed, the final classification output label of all samples is output, and the establishment of the classification model of the training set is completed.

[0076] Specifically, in this embodiment, after the classification model is established, the label of a new sample can be predicted. The specific method is: for a new sample, a K+N neighbor sample graph model is constructed based on the training samples and the new sample, and the label of the new sample is predicted.

[0077] In this embodiment, after the K+N neighbor sample graph model is obtained by using the training set data, the label of a new sample can be determined by the K+N neighbor sample graph model, and the label prediction and judgment of the new sample in the test set are completed. The specific method is: the data of the new sample is obtained, the candidate label is determined, the candidate label set is generated, and the generation method is according to the method of step S2. The new sample is added to the K+N neighbor sample graph generated by the training set to generate a new K+N neighbor sample graph, and the generation method is according to the method of step S302. Finally, the label of the new sample is propagated according to the newly obtained K+N neighbor sample graph model, and the inference prediction of the label of the new sample is completed.

[0078] The algorithm performance evaluation stage is based on UCI and other data sets for experiments, and six partial label learning methods are compared comprehensively, and the results of ten-fold experiments are shown in box plots, and the performances of the nearest neighbor propagation-based partial label learning method (NNPL) of the application and other algorithms, such as PL-KNN (Partial Learning K-Nearest Neighbor classification), IPAL (an Instance-based PArtial Label learning), CLPL (Convex Loss for Partial Labels), PL-LEAF (Partial Label LEArning via Feature-aware disambiguation), M3PL (MaxiMum Margin Partial Label learning) and PALOC (PAtial Label learning via One-vs-one deComposition) in accuracy and precision are summarized. Through classification performance evaluation, it is proved that the nearest neighbor propagation-based partial label learning method (NNPL) of the application generally has better classification effect compared with other methods.

[0079] Figure 3 The accuracy evaluation results of various algorithms for different data sets. Figure 3 The three subgraphs (a), (b) and (c) in FIG. 1 show the accuracy evaluation results of various algorithms on three UCI data sets, and it can be seen that the accuracy of the NNPL algorithm of the application is generally at a high level in the ten-fold evaluation results. Figure 3 The two subgraphs (d) and (e) in FIG. 1 show the accuracy evaluation results of different algorithms on two real data sets. It can be seen that the NNPL algorithm of the application performs better on the MSRCv2 data set, and performs better than the other two partial label learning algorithms PL-KNN and IPAL improved by the label propagation idea on the Lost data set.

[0080] Figure 4 The precision evaluation results of various algorithms for different data sets. As shown in Figure 4 As shown in (a) and (e), the precision evaluation results of the NNPL algorithm of the application are slightly better than those of other algorithms on the Glass data set and the Lost data set. As shown in Figure 4 (b) and (d) show that the precision of the NNPL algorithm of the application is significantly better than that of other algorithms on the Seed data set and the MSRCv2 data set. As shown inFigure 4 (c) The results of the accuracy evaluation of the NNPL algorithm of the present application are similar to those of most algorithms in the Waveform dataset. Although the NNPL of the present application does not have the best performance in every dataset, the accuracy of the NNPL algorithm of the present application is generally at a high value.

[0081] The present application provides a brain image classification model establishment method based on near neighbor propagation partial label learning. Unlike existing classification algorithms, the present application avoids excessive dependence on the diagnosis results of clinicians in current mental illness diagnosis classification algorithms, combines graph structure to re-propagate sample labels, and establishes a model. The characteristics and innovations of the present application are mainly as follows:

[0082] 1) The present application re-measures the relationship between subjects based on the partial label learning algorithm and the K+N neighbor sample graph structure, re-labels subjects with low label credibility, and improves the accuracy of sample label classification.

[0083] 2) There are few studies on partial label classification assisting mental illness diagnosis. In the weakly supervised learning algorithm, the initial default label is only one label, but the partial label algorithm assumes that the initial label is any label in the candidate label set. This candidate label set assumption is very consistent with the current situation that mental illness is difficult to accurately diagnose due to the high overlap of disease symptoms.

[0084] 3) The present application draws on the semi-supervised learning research ideas of graphs, and the K+N neighbor sample graph structure generated by the present application has nodes representing samples and weighted edges reflecting the similarity between a pair of nodes. The label information is propagated from the labeled samples to the unlabeled samples by combining the structure information of the constructed graph. Whether this graph structure or manifold assumption or similarity assumption is very consistent with the characteristics of mental illness data, and is very suitable for mining effective information of such data.

[0085] 4) Compared with traditional partial label algorithms, the classification accuracy of the present application is better, and the present application is more suitable for application conditions where the training data labels are not accurate, so the present application has more advantages.

[0086] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for establishing a brain image classification model based on neighbor propagation partial label learning, characterized in that: The following steps are involved: S1. Obtain the training data set and the corresponding original labels; Calculate the similarity matrix and degree centrality between samples in the same category of the training data set, and determine the typical samples of each category based on the degree centrality. Then, take the mean of all typical samples in the same category group as the initial category center of the category. Construct a typical sample set based on the typical samples of each category, and construct a category center set based on the initial category center of each category. S2. Determine the candidate labels for the samples in the training dataset and construct a candidate label set; S3. Using the training data set, a classification model is established using the K+N nearest neighbor sample graph, which specifically includes: S301: Take the typical sample and its K nearest neighbor node samples as samples of this layer, and determine the candidate labels of its nearest neighbor node samples; S302: Construct a K+N nearest neighbor sample graph of samples in this layer based on the typical sample set, category center, and sample features of the training samples; the K+N nearest neighbor sample graph includes K nearest neighbor sample nodes and N nearest typical sample nodes; S303: Perform label inference on neighboring node samples using typical samples to determine the classification output labels of the neighboring node samples. Then, calculate the degree centrality of each sample in the same category group, increase the number of typical samples in each category based on the degree centrality, and update the typical sample set, category center set, and candidate label set. S304: Determine whether all sample labels are determined. If not, return to step S301; if so, output the final classification output labels of all samples to complete the establishment of the classification model.

2. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 1, characterized in that: The brain imaging data includes: functional magnetic resonance imaging data, structural magnetic resonance imaging data, diffusion tensor imaging data, electroencephalogram data, and magnetic brain data; The characteristics of the brain imaging data include: brain functional connectivity characteristics, brain gray matter density and volume characteristics, brain structural connectivity characteristics, EEG signal characteristics, and characteristics after multimodal fusion of brain images.

3. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 1, characterized in that: In step S1, the matrix element of the similarity matrix is ​​a correlation coefficient, and the correlation coefficient is a Pearson correlation coefficient, a Gaussian kernel correlation coefficient, or a regression correlation coefficient.

4. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 1, characterized in that: In step S303, the method for determining the classification output label of the neighboring sample is: Loop calculation of temporary confidence matrix And label confidence matrix F (t) , the calculation formula is: Where t represents the number of transmissions; F (0) Represents the initial label confidence matrix, and takes the mean of the number of candidate labels in the candidate label set of the sample as the initial label confidence of the current sample; α is the weight, W K represents the K nearest neighbor weight matrix, W N represents the N nearest neighbor weight matrix; F (t-1) Represents the label confidence matrix obtained during the t-1th propagation; is the temporary confidence matrix The elements in Represents the label confidence matrix F (t) Elements in Indicates the temporary confidence that the classification output label of the i-th sample in the candidate label set is category l, y l represents the label of category l, S i represents the candidate label set, n represents the number of training samples in this layer, and q represents the number of categories; When the convergence condition is reached, the loop stops and the label corresponding to the maximum confidence value in the label confidence matrix is ​​used as the label of the neighboring sample.

5. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 1, characterized in that: In step S1, degree centrality is the sum of sample similarities; The method for determining the typical samples of each category is: sort the degree centrality values ​​of each sample, and select the samples with the highest degree centrality in the same category group as the typical samples.

6. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 5, characterized in that: In step S303, the specific method of increasing the number of typical samples of each category according to degree centrality is: Sort the degree centrality values ​​of each sample in the same category group, select the top B% of atypical samples in the degree centrality ranking within the same category group as new typical samples, and add them to the typical sample set, where B is greater than A.

7. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 1, characterized in that: In step S2, the specific method for determining the candidate labels of the training set samples is: The original labels of typical samples are identified as classification output labels; Calculate the similarity between the remaining samples and each category center. In addition to the original label, select the label corresponding to the category center with the highest similarity as the candidate label, and add it together with the original label to the candidate label set of the current sample; finally determine the candidate labels of all training samples and generate a candidate label set.

8. The method for establishing a brain image classification model based on neighbor propagation partial label learning according to claim 1, characterized in that: In step S302, the method for constructing the K+N nearest neighbor sample graph structure of the samples in the current layer is: Based on the typical sample set, category center and sample features of the training samples, a fully connected original graph of the training samples is constructed; The fully connected original graph is thinned twice to obtain a K-nearest neighbor sample graph and an N-nearest neighbor typical sample graph; The K nearest neighbor sample graph and the N nearest neighbor typical sample graph are superimposed to obtain the K+N nearest neighbor sample graph of the samples in this layer.

Images