Intelligent medical diagnosis method and device based on semi-supervised deep non-parallel support vector machine

Abstract

The application discloses an intelligent medical diagnosis method and device based on a semi-supervised deep non-parallel support vector machine, and the method comprises the following steps: 1) constructing and preprocessing a medical data set; 2) augmenting supervised samples and unsupervised samples; 3) calculating a reconstruction loss and a consistency loss; 4) screening high-quality pseudo labels by using a confidence degree; 5) constructing a classification task loss function, wherein a similar ''one-to-remainder'' is adopted to process a multi-classification learning problem; 6) constructing a model overall loss function; 7) constructing a medical diagnosis grading evaluation decision function; and 8) predicting a grading evaluation result of a new medical sample.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer-aided medical diagnosis, and in particular to intelligent medical diagnosis data mining tasks for different patients' genetic information and clinical characteristics. Specifically, the present application proposes an intelligent medical diagnosis method and device based on a semi-supervised deep nonparallel support vector machine (SDNSVM) for the hierarchical diagnosis of brain glioma. BACKGROUND

[0002] With the development of intelligent medical diagnosis technology, the accuracy and efficiency of medical diagnosis have been significantly improved. Intelligent medical diagnosis is mainly applied to early identification of diseases, prognosis prediction, treatment plan formulation, reduction of misdiagnosis and missed diagnosis rates, real-time monitoring and early warning of patients, and other aspects. By combining advanced machine learning technology, intelligent medical diagnosis systems can analyze a large amount of clinical data and image information, provide auxiliary decision support, and improve the accuracy and efficiency of diagnosis.

[0003] Among numerous medical applications, brain glioma hierarchical diagnosis has particularly important research value and broad application prospects. Brain glioma refers to a tumor originating from brain glial cells and is one of the most common primary intracranial tumors. Although its pathogenesis is not fully understood, the following two key factors have a significant impact on technological progress: long-term exposure to high-dose ionizing radiation is a known inducing factor, and genetic mutations of certain rare syndromes-related high-penetrance genes are also important risk factors. The annual incidence of brain glioma in China is 5-8 / 100,000, and the 5-year mortality rate is second only to pancreatic cancer and lung cancer among all tumors. According to histological and imaging criteria, brain glioma is mainly divided into two categories: low-grade glioma (LGG) and glioblastoma multiforme (GBM). Low-grade glioma grows slowly and has low malignancy, while glioblastoma multiforme grows rapidly and has high malignancy and poor prognosis. In addition to traditional histological and imaging characteristics, clinical and genetic mutation factors (such as patients' age, gender, and specific genetic mutation status) also play an important role in accurate grading and treatment strategy formulation.

[0004] However, in the field of machine learning, the brain glioma grading diagnosis method faces two main challenges: first, the cost of obtaining samples with diagnostic information is high, and it is difficult to obtain enough data, which causes the model to be unable to learn enough information from a small amount of data, thereby affecting the classification effect. Second, the classes of the glioma dataset are imbalanced, that is, the low-grade glioma (LGG) data is much more than the glioblastoma multiforme (GBM) data, which causes the model to ignore the minority class during the learning process, resulting in false judgments. Therefore, brain glioma grading diagnosis is an imbalanced learning problem.

[0005] In this context, relevant scholars have proposed several solutions to address the above challenges: [1] Yoon J, Zhang Y, Jordon J, et al. Vime: Extending the success of self-and semi-supervised learning to tabular domain [J]. Advances in Neural Information Processing Systems, 2020, 33: 11033-11043. [2] Pant H, Sharma M, Soman S. Twin neural networks for the classification of large unbalanced datasets [J]. Neurocomputing, 2019, 343: 34-49. [3] Qi Z, Tian Y, Shi Y. Laplacian twin support vector machine for semi-supervised classification [J]. Neural networks, 2012, 35: 46-53. [4] Ren Y, Zhu H, Tian Y, et al. A Laplacian svm based semi-supervised classification using multi-local linear model [J]. IEEE Transactions on Electrical and Electronic Engineering, 2021, 16(3): 455-463. Literature [1] pre-trains table data through self-supervised learning, then uses semi-supervised learning for classification tasks, and uses mask vectors to damage data to expand table data. Literature [2] extends TWSVM to the neural network framework and proposes a double neural network to address the class imbalance problem. Two neural networks are trained to learn linear classifiers in feature space, similar to the hyperplane model of TWSVM, to handle imbalanced datasets. Literature [3] introduces a manifold regularization term combined with SVM to solve the semi-supervised problem. Literature [4] integrates latent boundary information into the kernel function through an interpretable model, improving theoretical interpretability. Although these methods have alleviated the semi-supervised learning and SVM imbalance problem to some extent, there are still the following challenges:

[0006] 1. Literature [1] requires pre-training before semi-supervised training, resulting in a significant increase in training time.

[0007] 2. Although the document [2] solves the problem of class imbalance, the multi-classification problem requires a large number of parameters to be solved, increasing the amount of calculation and time.

[0008] 3. The documents [3] and [4] do not consider the imbalance between classes, resulting in bias of the model to the minority class.

[0009] To this end, the present application proposes an intelligent medical diagnosis method based on semi-supervised deep nonparallel support vector machine to more effectively cope with these challenges and improve the accuracy and efficiency of brain glioma grading diagnosis. SUMMARY

[0010] In order to overcome the brain glioma grading diagnosis problem in intelligent medical treatment, the present application proposes an intelligent medical diagnosis method and device based on semi-supervised deep nonparallel support vector machine (Semi-supervised Deep Nonparallel Support Vector Machine, SDNSVM). The method is realized through a neural network framework, combines the loss function of nonparallel support vector machine, fully utilizes a small amount of labeled data and a large amount of unlabeled data, solves the problems of insufficient data and class imbalance, and improves the accuracy and efficiency of brain glioma grading diagnosis.

[0011] Specifically, the technical innovations of the present application include the following aspects:

[0012] 1) The present application introduces a self-supervised numerical feature augmentation mechanism. A feature generator is constructed by a mask vector and data features to generate new features to augment the original table data, which is used to enhance the mutual relationship learning between features and improve the generalization ability of the model.

[0013] 2) The present application proposes a pseudo-label semi-supervised learning strategy. By using the prediction results of the model on unlabeled data as pseudo-labels, the training set is expanded and the model performance is improved. In order to ensure the high quality of the pseudo-labels, the present application introduces a confidence screening mechanism, thereby enhancing the reliability of unsupervised sample classification.

[0014] 3) In order to solve the problem of class imbalance, the present application increases new samples by randomly sampling from minority class samples, so that the proportion of each class of samples reaches balance. The model avoids the preference for the majority class in the learning process, improves the recognition ability of the model to the minority class, and ensures the accuracy and fairness of classification.

[0015] 4) The present application constructs a nonparallel intra-class and inter-class loss mechanism. In order to handle complex multi-class learning scenarios,

[0016] By combining the local information of intra-class and inter-class, and using parameter sharing mechanism to find the optimal hyperplane,

[0017] The learning ability for multi-class information is enhanced, and the learning efficiency of the model is provided.

[0018] For better understanding and description, the present application formulates unified symbol representation rules, wherein scalar is represented by non-bold lowercase or uppercase letters, such as x, X. Vector is represented by bold lowercase letters, such as x. Matrix is represented by bold uppercase letters, such as X. Table 1 is a symbol description table related to the present application.

[0019]

[0020]

[0021] Table 1 SDNSVM symbol description table

[0022] The present application is an intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine (SDNSVM), and the total process is as shown in the figure. Figure 1 The specific steps of the intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine are described as follows.

[0023] Step 1: Construct and preprocess the medical data set.

[0024] This step realizes the construction and preprocessing of the medical data set, including data screening and preprocessing, data normalization and data labeling, and unbalanced data balancing. The specific description is as follows:

[0025] (1.1) Data screening and preprocessing. First, the clinical medical data is screened and analyzed, and 20 most frequent mutation gene sites and 2 clinical features: gender and diagnosis age are extracted. In order to ensure the integrity and quality of the data, the original data is preprocessed, including filtering instances with missing key information, converting the confirmed age from string format to continuous value, and excluding irrelevant columns (such as medical record number, preliminary diagnosis and project).

[0026] (1.2) Data normalization and data labeling. In order to adapt to the input requirements of the model, the data is normalized. Normalization ensures that the numerical ranges of different features are consistent, thereby improving the effect of model training. Specifically, after normalization, the data is labeled according to the doctor's diagnosis result to obtain supervised samples where x i ∈R d is a sample feature with dimension d, y i is a sample label, and n l is the size of the supervised sample; for data without labels, it is an unsupervised sample where n u is the size of the unsupervised sample.

[0027] (1.3) Unbalanced data balancing. Check the preprocessed brain glioma dataset to ensure the accuracy and consistency of the data. To solve the problem of class imbalance in the dataset, randomly sample the minority class samples to increase the number of minority class samples, so that the data reaches a balanced state. This prevents the model from developing a preference for the majority class during the learning process, thereby improving the recognition ability of the minority class.

[0028] First, calculate the number of samples in each class, denoted as n k , and determine the target sample size. Assume that the class with the most samples is n max , the goal is to increase the number of samples in all classes to n max . Then calculate the number of samples that need to be added, for each minority class k, the number of samples that need to be added is:

[0029]

[0030] Finally, randomly select samples from the minority class k, and add these samples to the dataset. Through this process, the number of samples in all classes will tend to be the same, thereby balancing the dataset.

[0031] Step two: augment supervised samples and unsupervised samples.

[0032] This step realizes the augmentation function of the dataset samples, by constructing a mask generator and a feature generator, integrating the augmented output, thereby improving the training effect and generalization ability of the model. Inspired by self-supervised learning, a series of enhancement transformations are performed on the original training data x i , u i , to generate new training samples, learn the potential relationship between data samples, and learn useful feature representations, achieving the diversity of data samples.

[0033] The specific description is as follows:

[0034] (2.1) Construct a mask generator. The generator outputs a binary mask with dimension D, denoted as m = [m1, …, m D ] T ∈ {0, 1} D , where m D is randomly sampled from a Bernoulli distribution with probability p m , that is:

[0035]

[0036] (2.2) Construct a feature generator (Pretext generator). The generator takes the input mask m and calculates the corresponding subtraction operation (1-m), and multiplies (1-m) ⊙ s by element-by-element multiplication with the original sample feature. The original feature sample s (which can be a supervised sample or an unsupervised sample) is randomly shuffled column by column to obtain Multiply element-wise with the mask m, denoted as The final operation can be expressed as:

[0037]

[0038] Among them, g m is the feature generator, is the enhancing features, ⊙ is the element-wise multiplication, s is the sample, is a randomly shuffled feature matrix. Formula (3) is interpreted as retaining the original samples of the part with a value of 1 in mask m plus the part with a value of 0 in mask m.

[0039] (2.3) Integrate the augmented output. According to formula (3), take the supervised sample x, unsupervised samples u1, u2 and mask m for sample s, and get the corresponding supervised sample augmented features and unsupervised sample augmentation features

[0040] Step 3: Calculate the reconstruction loss and consistency loss.

[0041] This step implements the function of calculating reconstruction and consistency loss, including building a feature representation network model, calculating the logits corresponding to the features, calculating the reconstructed feature representation, calculating the reconstruction loss of the sample, and calculating the mask reconstruction loss. The reconstruction loss measures the difference between the model's reconstructed features and the original features, while the consistency loss ensures that the output of the neural network remains consistent under different conditions. The network model structure diagram of the present invention is shown in the figure below. Figure 2 The specific description is as follows:

[0042] (3.1) Constructing a feature representation network model F. In order to fully utilize the powerful feature extraction capability of the neural network, the present invention uses a network model to extract the feature representation of the sample in the nonlinear space, replacing the nonlinear transformation implemented by the traditional kernel function. Specifically, the network F consists of an encoder f enc and decoder f dec The encoder f enc is responsible for mapping the input data into a compact feature space, while the decoder f dec It is used to reconstruct the input data, and the specific form is as follows:

[0043]

[0044] where, is the function composition symbol. F is composed of an encoder f enc and a decoder f dec as a whole network. The encoder-decoder structure not only can effectively capture the complex nonlinear characteristics of samples, but also can improve the accuracy and robustness of feature expression.

[0045] (3.2) Calculate the corresponding logits of the feature. The augmented feature is concatenated into the input of the encoder f enc to obtain the corresponding logits, denoted as

[0046]

[0047] where, W e , W e and b e are network weights and biases, is the vector concatenation symbol. By concatenating the augmented feature, it is input into the encoder and the corresponding logits are output.

[0048] (3.3) Calculate the reconstructed feature representation. The augmented feature is concatenated into the input of the encoder-decoder model to obtain the supervised sample reconstructed feature , the unsupervised reconstructed feature and the mask reconstructed feature

[0049]

[0050] where, W d and b d are network weights and biases. By concatenating the augmented feature and inputting it into the encoder-decoder model, the corresponding reconstructed feature is output.

[0051] (3.4) Calculate the reconstruction loss of the sample. The reconstruction loss can evaluate the difference between the reconstructed sample and the original sample, and this calculation process is shown in Figure 3 .

[0052] Calculate the mean square error loss of the supervised sample:

[0053]

[0054] where, n l is the size of the supervised sample, the original sample feature x and the reconstructed sample feature When minimizing this loss, the difference between x and can be reduced.

[0055] Compute the reconstruction loss of unsupervised samples:

[0056]

[0057] where n u is the number of unsupervised samples, and u1is the original unsupervised sample feature and is the reconstructed unsupervised sample feature. The difference between u1and is reduced to ensure the consistency of the reconstructed feature and the original feature.

[0058] (3.5) Compute the mask reconstruction loss. The cross-entropy loss is computed between the original mask m and the reconstructed mask :

[0059]

[0060] where D is the size of the mask m. The difference in probability distribution between the original mask m and the reconstructed mask is evaluated by computing the cross-entropy loss.

[0061] (3.6) Compute the consistency loss of unsupervised samples. The consistency loss of logits, and is computed:

[0062]

[0063] where n u is the number of unsupervised samples. The prediction results of the same sample after enhancement, and can still remain consistent, so as to evaluate the ability of the model to maintain stability in predicting data changes.

[0064] Step four: Use confidence to filter high-quality pseudo labels.

[0065] This step realizes the effective use of unsupervised samples by generating and filtering high-quality pseudo labels. It includes generating pseudo label vectors for unsupervised samples and filtering out high-quality pseudo labels to enhance the reliability of unsupervised sample prediction. By using a confidence threshold to filter pseudo labels, it can ensure that the pseudo labels used by the model during training have high accuracy and credibility. The specific description is as follows:

[0066] (4.1) Pseudo label vector for unsupervised samples. According to formula (5), is the logits obtained by only using the encoder f enc for unsupervised samples. For There are:

[0067]

[0068] where, is the output for each class, k is the class. is the vector of each class output.

[0069] The formula (11) is obtained Take the absolute value, get Through the activation function softmin, the real number mapped to 0-1 is obtained, denoted as And the normalization ensures the sum to be 1, that is:

[0070]

[0071] where, σ is the symbol of the activation function softmin. The sum of all probabilities is 1.

[0072] (4.2) Screening high-quality pseudo labels. A scalar hyperparameter confidence threshold τ∈[0, 1] is introduced, and the predicted results above the threshold are retained as the class pseudo label. On this basis, increasing the confidence threshold can screen pseudo label samples with higher credibility. First take its q k The maximum probability is compared with the confidence threshold τ, and the value greater than the threshold is retained.

[0073]

[0074] where, is the indicator function, max(q k ) is the maximum class probability of unsupervised samples, and when the maximum class probability is greater than the confidence threshold τ , it is 1, otherwise 0.

[0075] According to formula (13), if it is greater than the confidence threshold τ, the pseudo label is taken out and denoted as

[0076]

[0077] where, the pseudo label is the class with the maximum probability in q k .

[0078] Step five: build the classification task loss function.

[0079] The step realizes loss function design of the classification task, including constructing intra-class non-parallel loss function and constructing inter-class non-parallel loss function. By combining local information of intra-class and inter-class and adopting a parameter sharing mechanism, the application finds an optimal hyperplane f k (x) for each class, so as to improve classification accuracy and robustness of the model, and the form is as follows:

[0080]

[0081] Wherein, f k (x) is the proximal hyperplane of the class k. w k and b k are the weight and bias of the corresponding hyperplane, f enc is an encoder, and x is the feature of the sample. And the application can be used for multi-classification, wherein a similar "one-to-many" idea is adopted to process the multi-classification learning problem, that is, for K-class classification, K binary classifiers are constructed, and the kth class is the positive class in the kth classifier, and the remaining classes are all trained as negative classes. The intra-class and inter-class loss calculation process of the application is as shown in Figure 4 .

[0082] (5.1) Constructing intra-class non-parallel loss function. For the kth class of all sample points s, construct the intra-class loss function. The output of the kth class is required to be minimum. Taking a sample point i as an example, according to the true label y and the pseudo label , the output of the kth class is first screened out and denoted as:

[0083]

[0084] Then the minimum distance of the kth class is calculated, and the loss function is constructed as follows:

[0085]

[0086] The 2-norm of the kth class is calculated, and the distance from the kth class to the origin is calculated.

[0087] In addition, the minimum output of the class other than the kth class is required to be at least α away from the hyperplane to which the kth class belongs. Taking a sample point i as an example, the loss function is constructed as follows:

[0088]

[0089] Wherein, α∈R is the margin. Formula (18) calculates the minimum output of the class other than the kth class, and formula (19) requires that the class other than the kth class and the hyperplane to which the kth class belongs have a margin distance.

[0090] (5.2) Constructing inter-class non-parallel loss function. For the kth class and the class other than the kth class Constructing inter-class loss function. The minimum output of the class except the kth class is required The distance to the kth class is at least a, and the loss function is constructed as follows for a sample point i:

[0091]

[0092] where a e R is the margin. Equation (20) requires that the kth class and the classes other than the kth class have a margin distance.

[0093] Step six: constructing the overall loss function of the model.

[0094] This step integrates an overall loss function, including promoting intra-class sample compactness, inter-class sample discrimination, model reconstruction accuracy, and prediction consistency. Integrating these target loss functions into a unified loss function can guide the model to continuously optimize during the training process, thereby improving the overall generalization ability of the model. The specific description is as follows:

[0095] (6.1) Fusion of intra-class and inter-class loss functions. In this sub-step, the intra-class loss function and the inter-class loss function are fused to optimize the intra-class compactness and inter-class discrimination at the same time. Specifically, by combining equation (17), equation (19) and equation (20):

[0096]

[0097] where w ik , is a regularization term, and λ1 e R, λ2 e R, λ3 e R are weighting parameters. Analyzing equation (21), the first term of the objective function is a regularization term, in order to realize the structural risk minimization, that is, to balance the model complexity and the model accuracy, and to avoid overfitting; the second term makes the distance from the kth class sample point to the hyperplane of the class closer; the third term requires that the sample points of other classes are at least a margin distance from the kth class sample points, that is, the slack variable is the measurement error when the constraint condition is not met, aiming to maximize the separation of the two classes of sample points; the fourth term requires that the sample points of other classes are at least a margin distance from the hyperplane to which the kth class sample points belong, that is, other classes are as far away from the hyperplane to which the kth class sample points belong as possible.

[0098] The supervised loss function is:

[0099]

[0100] which represents the loss for all supervised samples.

[0101] The unsupervised loss is:

[0102]

[0103] Analysis is made on formula (23), is a supervised sample maximum class probability, when the maximum class probability is greater than the confidence threshold τ, then ik is 0, indicating that high-quality pseudo-labels are screened out for training.

[0104] (6.2) Construct a unified model overall loss function. In this sub-step, the intra-class and inter-class losses are combined with the model reconstruction loss and the prediction consistency loss to construct a unified model overall optimization loss function. According to the above losses (22), (23) and the reconstruction losses (7), (8), (9) and the consistency loss (10) of the step, the overall loss function is integrated.

[0105] Loss=λ x ·(L xr +L x )+λ u ·(L ur +L uc +L u )+λ m ·L m (24)

[0106] Wherein, λ x ∈R, λ u ∈R, λ m ∈R is a weight parameter.

[0107] Step seven: medical diagnosis grading evaluation stage.

[0108] This step aims to realize the grading diagnosis of brain glioma data. For a new data sample, that is, brain glioma data x, the class of the new sample is judged by calculating the distance of x to each non-parallel near-end hyperplane. Each non-near-end hyperplane represents a class (low-grade glioma LGG, glioblastoma GBM). The class of the nearest hyperplane is the class of the new sample. The decision function is as follows:

[0109]

[0110] Wherein, |·| represents the absolute value, k=1, …, K represents the kth class.

[0111] The present application aims at the problem of class imbalance in brain glioma grading estimation under a small amount of data, can be used for identifying tumors, predicting their malignancy and recurrence risk, and proposes an SDNSVM method, which performs more superiorly when the data set has characteristics such as a small amount of data and class imbalance learning.

[0112] The present application has the advantages that:​

[0113] 1) The present application introduces a self-supervised numerical feature augmentation mechanism. A feature generator is constructed by a mask vector and data features to generate new features to augment the original table data, which is used to enhance the mutual relationship learning between features and improve the generalization ability of the model.

[0114] 2) The present application proposes a pseudo-label semi-supervised learning strategy. By using the prediction results of the model on unlabeled data as pseudo-labels, the training set is expanded and the model performance is improved. In order to ensure the high quality of the pseudo-labels, the present application introduces a confidence screening mechanism, thereby enhancing the reliability of unsupervised sample classification.

[0115] 3) In order to solve the problem of class imbalance, the present application increases new samples by randomly sampling from minority class samples, so that the proportion of each class sample is balanced. The model avoids the preference for majority classes in the learning process, improves the recognition ability of the model for minority classes, and ensures the accuracy and fairness of classification.

[0116] 4) The present application constructs a non-parallel intra-class and inter-class loss mechanism. In order to deal with complex multi-class learning scenarios, the local information of intra-class and inter-class is combined, and the parameter sharing mechanism is used to find the optimal hyperplane, which enhances the learning ability of multi-class information and provides the learning efficiency of the model. BRIEF DESCRIPTION OF DRAWINGS

[0117] Figure 1 The present application is a method for the total flowchart.

[0118] Figure 2 The present application is a network model structure diagram.

[0119] Figure 3 The present application is a calculation of reconstruction and consistency loss flowchart.

[0120] Figure 4 The present application is a calculation of intra-class and inter-class loss flowchart. DETAILED DESCRIPTION

[0121] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings to explain the technical solutions of the present application. The embodiments described herein are only used to explain the present application and do not limit the present application.

[0122] Example 1

[0123] This embodiment relates to an intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine, and the specific steps are as follows:

[0124] Step 1: Construct and preprocess the medical data set.

[0125] This step realizes the construction and preprocessing of medical data set, including data screening and preprocessing, data normalization and data labeling, and unbalanced data balancing. The specific description is as follows:

[0126] (1.1) Data screening and preprocessing. First, the clinical medical data is screened and analyzed, and 20 most frequent mutation genes and 2 clinical characteristics: gender and diagnosis age are extracted. In order to ensure the integrity and quality of the data, the original data is preprocessed, including filtering instances with missing key information, converting the confirmed age from string format to continuous value, and excluding irrelevant columns (such as medical record number, preliminary diagnosis and project).

[0127] (1.2) Data normalization and data labeling. In order to adapt to the input requirements of the model, the data is normalized. Normalization ensures that the numerical range of different features is consistent, thereby improving the effect of model training. Specifically, after normalization, the data is labeled according to the doctor's diagnosis result, and supervised samples where x i ∈R d is the sample feature of dimension d, y i is the corresponding label of the sample, and n l is the size of the supervised sample; for the data without label, it is an unsupervised sample where n u is the size of the unsupervised sample.

[0128] (1.3) Unbalanced data balancing. The preprocessed glioma data set is checked to ensure the accuracy and consistency of the data. In view of the problem of class imbalance in the data set, the number of minority class samples is increased by random sampling, so that the data reaches a balanced state. Thus, it prevents the model from producing preference for the majority class during the learning process, thereby improving the recognition ability of the minority class.

[0129] First, calculate the number of samples in each class, denoted as n k , and then determine the target sample number. Assuming that the class with the most samples is n max , the target is to increase the number of samples in all classes to n max . Then calculate the number of samples that need to be increased. For each minority class k, the number of samples that need to be increased is:

[0130]

[0131] Finally, randomly select samples from the minority class k, and add these samples to the data set. Through this process, the number of samples in all classes will tend to be the same, thereby balancing the data set.

[0132] Step two: augmenting supervised samples and unsupervised samples.

[0133] This step realizes the augmentation function of the data set samples. By constructing a mask generator and a feature generator, integrating the augmented output, the training effect and the generalization ability of the model are improved. Inspired by self-supervised learning, a series of enhancement transformations are performed on the original training data x i , u i to generate new training samples, learn the potential relationship between data samples, and learn useful feature representations to achieve the diversity of data samples.

[0134] The specific description is as follows:

[0135] (2.1) Construct a mask generator. The generator outputs a binary mask of dimension D, denoted as m = [m1, …, m D ] T ∈{0, 1} D , where m D is randomly sampled in a Bernoulli distribution with probability p m , that is:

[0136]

[0137] (2.2) Construct a feature generator (Pretext generator). The generator calculates the corresponding subtraction operation (1-m) from the input mask m, and obtains (1-m) o s by element-wise multiplication of (1-m) and the original sample feature. Randomly shuffle the original feature sample s (which can be a supervised sample or an unsupervised sample) column by column to obtain Multiply by the mask m element by element, denoted as The final operation can be represented as:

[0138]

[0139] where g m is the feature generator, is the augmented feature (enhancing features), o is the element-wise multiplication, s is the sample, is the randomly shuffled feature matrix. Equation (3) explains that the part with the value of 1 in the mask m is retained and the original sample with the value of 0 in the mask m is added.

[0140] (2.3) Integrate augmented output. According to equation (3), take the supervised sample x, the unsupervised samples u1, u2 and the mask m for the sample s, respectively, to obtain the corresponding supervised sample augmented feature and the unsupervised sample augmented feature

[0141] Step three: calculate reconstruction loss and consistency loss.

[0142] This step realizes the function of calculating reconstruction and consistency loss, including constructing a feature representation network model, calculating the corresponding logits of the features, calculating the reconstructed feature representation, calculating the reconstruction loss of the sample, and calculating the mask reconstruction loss. The reconstruction loss measures the difference between the reconstructed features and the original features, while the consistency loss ensures that the output of the neural network remains consistent under different conditions. The network model structure diagram of the present application is shown in Figure 2 , and the specific description is as follows:

[0143] (3.1) Construct a feature representation network model F. In order to fully utilize the powerful feature extraction capability of neural networks, the present application adopts a network model to extract the feature representation of the sample in a nonlinear space, instead of the traditional nonlinear transformation realized by kernel functions. Specifically, the network F is composed of an encoder f enc and a decoder f dec , both of which adopt a 3-layer fully connected layer structure. The encoder f enc is responsible for mapping the input data to a compact feature space, while the decoder f dec is used to reconstruct the input data, which is specifically as follows:

[0144]

[0145] wherein, is the function composition symbol. F is composed of the encoder f enc and the decoder f dec as a whole network. The encoder-decoder structure not only effectively captures the complex nonlinear features of the sample, but also improves the accuracy and robustness of feature representation.

[0146] (3.2) Calculate the corresponding logits of the features. Concatenate the augmented feature into the encoder f enc to obtain the corresponding logits, denoted as

[0147]

[0148] wherein, W e , wherein W e and b e are network weights and biases, is the vector concatenation symbol. By concatenating the augmented feature, it is input into the encoder and outputs the corresponding logits.

[0149] (3.3) Calculate the reconstructed feature representation. Concatenate the augmented feature Concatenate the input encoder-decoder model to obtain the supervised sample reconstruction feature Unsupervised reconstruction feature And mask reconstruction feature

[0150]

[0151] where W d and b d are network weights and biases. By concatenating the augmented features and inputting the encoder-decoder model, the corresponding reconstruction feature is output.

[0152] (3.4) Calculate the reconstruction loss of the sample. The reconstruction loss can evaluate the difference between the reconstructed sample and the original sample. The calculation process is shown in Figure 3 .

[0153] Calculate the supervised sample mean square error loss:

[0154]

[0155] where n l is the size of the supervised sample, the original sample feature x and the reconstructed sample feature When minimizing this loss, the difference between x and can be reduced.

[0156] Calculate the reconstruction loss of the unsupervised sample:

[0157]

[0158] where n u is the size of the unsupervised sample, the original unsupervised sample feature u1 and the reconstructed unsupervised sample feature To minimize this loss, the difference between u1 and is reduced, thereby ensuring the consistency of the reconstructed feature and the original feature.

[0159] (3.5) Calculate the mask reconstruction loss. Calculate the cross-entropy loss between the original mask m and the reconstructed mask .

[0160]

[0161] where D is the size of the mask m. By calculating the cross-entropy loss, the difference between the probability distribution of the original mask m and the reconstructed mask is evaluated.

[0162] (3.6) Calculate the consistency loss of the unsupervised sample. Calculate the logits of the unsupervised sample, and consistency loss:

[0163]

[0164] where n u is the unsupervised sample size. The prediction results of the enhanced samples, and can still remain consistent, thereby evaluating the model's ability to maintain stability in predicting data changes.

[0165] Step four: screening high-quality pseudo labels using confidence.

[0166] This step achieves effective use of unsupervised samples by generating and screening high-quality pseudo labels. It includes generating pseudo label vectors for unsupervised samples and screening high-quality pseudo labels to enhance the reliability of unsupervised sample predictions. By using a confidence threshold to screen pseudo labels, we can ensure that the pseudo labels used by the model during training have high accuracy and credibility. The specific description is as follows:

[0167] (4.1) Pseudo label vector for unsupervised samples. According to equation (5), is the logits obtained by the encoder f enc for unsupervised samples. For we have:

[0168]

[0169] where is the output of each class, and k is the class. represents the vector of each class output.

[0170] Taking the absolute value of obtained from equation (11), we get Through the activation function softmin, we get a real number mapped between 0 and 1, denoted as and normalized to ensure that the sum is 1, i.e.:

[0171]

[0172] where σ is the symbol of the activation function softmin. The sum of all probabilities is 1.

[0173] (4.2) Screening high-quality pseudo labels. Introduce a scalar hyperparameter confidence threshold τ ∈ [0, 1], and the prediction result above this threshold retains the class pseudo label. On this basis, increasing the confidence threshold can screen pseudo label samples with higher credibility. First, take its qk The maximum probability in q is compared with a confidence threshold τ, and if greater than the threshold, it is retained.

[0174]

[0175] wherein, is an indicator function, and max(q k ) is the maximum class probability of the unsupervised sample, and when the maximum class probability is greater than the confidence threshold τ, it is 0 otherwise.

[0176] According to formula (13), if greater than the confidence threshold τ, the pseudo label is taken out and marked as

[0177]

[0178] wherein, the pseudo label is the class with the maximum probability in q k .

[0179] Step five: constructing a classification task loss function.

[0180] This step realizes the loss function design of the classification task, including constructing an intra-class loss function and constructing an inter-class loss function. By combining the local information of the intra-class and the inter-class and adopting a parameter sharing mechanism, the present application finds an optimal hyperplane f k (x) for each class, thereby improving the classification accuracy and the robustness of the model, and the form is as follows:

[0181]

[0182] wherein, f k (x) is the proximal hyperplane of the class k. w k and b k are the weight and the bias of the corresponding hyperplane, and f enc is an encoder, and x is the feature of the sample. And the present application can be used for multi-classification, wherein a similar “one-versus-rest” idea is adopted to process the multi-classification learning problem, that is, for K-class classification, K binary classifiers are constructed, and in the kth classifier, the kth class is the positive class, and the rest of the classes are trained as the negative class. The intra-class and inter-class loss calculation process of the present application is shown in Figure 4 .

[0183] (5.1) Constructing an intra-class loss function. For the kth class of all sample points s, an intra-class loss function is constructed. The output of the kth class is required to be minimum. Taking a sample point i as an example, according to the real label y and the pseudo label , the output of the kth class is first screened out and marked as:

[0184]

[0185] Recalculate the minimum distance of the kth class, and the loss function is constructed as follows:

[0186]

[0187] Take the 2-norm of the kth class, and calculate the distance from the kth class to the origin.

[0188] In addition, the minimum output of the class other than the kth class is required The distance between the kth class and the hyperplane to which it belongs is at least α. Taking a sample point i as an example, the loss function is constructed as follows:

[0189]

[0190] where α ∈ R is the margin. Equation (18) calculates the minimum output of the class other than the kth class, and equation (19) requires that the class other than the kth class and the hyperplane to which the kth class belongs have a margin distance.

[0191] (5.2) Construct the non-parallel inter-class loss function. For the kth class of all sample points s and the class other than the kth class Construct the inter-class loss function. The minimum output of the class other than the kth class is required The distance between the kth class and the hyperplane to which it belongs is at least α. Taking a sample point i as an example, the loss function is constructed as follows:

[0192]

[0193] where α ∈ R is the margin. Equation (20) requires that the class other than the kth class and the kth class have a margin distance.

[0194] Step six: construct the overall loss function of the model.

[0195] This step integrates an overall loss function, including promoting intra-class sample compactness, inter-class sample discrimination, model reconstruction accuracy, and prediction consistency. Combining these target loss functions into a unified loss function can guide the model to continuously optimize during the training process, thereby improving the overall generalization ability of the model. The specific description is as follows:

[0196] (6.1) Fusion of intra-class and inter-class loss functions. In this sub-step, the intra-class loss function and the inter-class loss function are fused to optimize the intra-class compactness and inter-class discrimination at the same time. Specifically, by combining equations (17), (19) and (20):

[0197]

[0198] where w ik , For the regularization term, λ1∈R, λ2∈R, λ3∈R are the trade-off parameters. Analyzing equation (21), the first term of the objective function is the regularization term, which is to minimize the target function in order to achieve the structural risk minimization, that is, to balance the model complexity and model accuracy, and to avoid overfitting; the second term is to make the distance between the k-th class sample point and the hyperplane of this class closer; the third term requires that the distance between other class sample points and the k-th class sample point is at least one margin distance, that is, the slack variable is the measurement error when the constraint condition is not met, which aims to maximize the separation of the two classes of sample points; the fourth term requires that the distance between other class sample points and the hyperplane of the k-th class sample points is at least one margin distance, that is, other classes are as far away from the hyperplane of the k-th class sample points as possible.

[0199] The supervised loss function is:

[0200]

[0201] which represents the loss for all supervised samples.

[0202] The unsupervised loss is:

[0203]

[0204] Analyzing equation (23), is an indicator function, max(q ik ) is the maximum class probability of the unsupervised sample, and when the maximum class probability is greater than the confidence threshold τ, then otherwise 0, which means that high-quality pseudo-labels are filtered out for training.

[0205] (6.2) Construct a unified model overall loss function. In this sub-step, the intra-class and inter-class losses are combined with the model reconstruction loss and the prediction consistency loss to construct a unified model overall optimization loss function. According to the above losses (22), (23) and the reconstruction losses (7), (8), (9) and the consistency loss (10) of step, the overall loss function is integrated.

[0206] Loss=λ x ·(L xr +L x )+λ u ·(L ur +L uc +L u )+λ m ·L m (24)

[0207] wherein, λ x ∈R, λ u ∈R, λ m ∈R are the trade-off parameters.

[0208] Step seven: medical diagnosis hierarchical evaluation stage.

[0209] This step aims to achieve the hierarchical diagnosis of brain glioma data. For a new data sample, i.e. brain glioma data x, the class of the new sample is determined by calculating the distance of x to each non-parallel proximal hyperplane. Each non-proximal hyperplane represents a class (low-grade glioma LGG, glioblastoma GBM). The class of the nearest hyperplane is the class of the new sample. The decision function is as follows:

[0210]

[0211] Where |. | represents the absolute value, k = 1, …, K represents the kth class.

[0212] In order to verify the robustness of the method of the present application under very few labeled data, the experiment takes 5%, 10%, and 20% of the labeled data of the training set of the brain glioma medical data set respectively. In order to ensure the fairness of the experimental results, the present application carries out experiments based on different random seeds, taking 0, 1, 2, 3, and 4 seeds. In the experiment, Deep Neural Networks (DNN), SVM and Lap-SVM are selected as comparison methods, among which DNN and SVM are supervised methods, and Lap-SVM and SDNSVM are semi-supervised methods. Accuracy (ACC), Recall, Precision, F1 Score and G_mean are used as method evaluation indexes, and the specific definitions are as follows:

[0213] Accuracy (ACC) is the proportion of correctly predicted classes in all predictions.

[0214]

[0215] Where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative.

[0216] Recall reflects the ability of the model to identify positive class samples. It refers to the proportion of all actual kth class samples that are correctly predicted as the kth class.

[0217]

[0218] Precision reflects the proportion of actual positive class samples in the samples predicted by the model as positive class. It refers to the proportion of all samples predicted as the kth class that are actually the kth class.

[0219]

[0220] The F1 Score combines precision and recall, providing a harmonic mean of the two.

[0221]

[0222] The G_mean reflects the overall performance of the model in handling imbalanced data. By considering the recognition ability of both positive and negative classes, G_mean provides a more comprehensive evaluation of imbalanced classification.

[0223]

[0224] The medical diagnosis output of the present application is for two different brain glioma levels: low-grade glioma (LGG) and glioblastoma multiforme (GBM). To verify the robustness of the SDNSVM method in the case of a small amount of labeled data, three groups of scenarios with different amounts of labeled data are designed, specifically 5%, 10%, and 20% of labeled data in the training set data. Tables 2, 3, and 4 record the grading diagnostic performance of the SDNSVM method of the present application compared with DNN, SVM, and Lap-SVM methods in five evaluation indexes under the condition of only 5%, 10%, and 20% labeled data.

[0225]

[0226] Table 2 Comparison of brain glioma grading diagnostic performance of each method under 5% labeled data

[0227] Table 2 shows that under the condition of 5% labeled data, the SDNSVM method proposed in the present application performs well, is almost not affected by the small amount of labeled data, and performs better than other algorithms. Under the supervision of DNN and SVM, the ACC of SDNSVM is 3.96% higher than that of supervised methods; compared with Lap-SVM, the performance of SDNSVM is improved by 14.07%. In terms of Recall, Precision, and F1 Score, the SDNSVM method is optimal. In addition, in terms of G_mean, SDNSVM far surpasses Lap-SVM, indicating that SDNSVM performs well in handling class imbalance problems.

[0228]

[0229] Table 3 Comparison of brain glioma grading diagnostic performance of each method under 10% labeled data

[0230] The results in Table 3 show that, under the condition of 10% labeled data ratio, the Lap-SVM has certain instability and insufficient accuracy in classifying samples. In contrast, the performance of the SDNSVM is significantly better than that of the Lap-SVM, and the SDNSVM shows better classification effect in comparison with the supervised method. The SDNSVM achieves the optimal results in F1 Score and G_mean indexes, indicating that the SDNSVM has good performance in dealing with the class imbalance problem.

[0231]

[0232] Table 4 Comparison of performance of each method in the diagnosis of brain glioma grading under the condition of 20% labeled data ratio

[0233] The results in Table 4 show that, under the condition of 20% labeled data ratio, the SDNSVM still performs well, and each index is significantly better than that of the Lap-SVM; compared with the supervised method such as DNN and SVM, the SDNSVM still shows certain advantages under the same labeled data ratio, verifying the effectiveness of the semi-supervised method.

[0234] From the results in Tables 2, 3 and 4, it can be seen that, under the condition of decreasing amount of labeled data and class imbalance, the SDNSVM can still maintain high classification performance. In general, the SDNSVM method proposed in the present application exhibits good robustness in the scene of diagnosis of brain glioma grading under class imbalance.

[0235] Embodiment 2

[0236] The present embodiment relates to an intelligent medical diagnosis device based on a semi-supervised deep non-parallel support vector machine, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the intelligent medical diagnosis method based on the semi-supervised deep non-parallel support vector machine of embodiment 1.

[0237] The above is the preferred embodiment of the present application, and the protection scope of the present application is not limited thereto. Any modification or replacement within the technical scope disclosed by the present application without creative labor is covered by the protection scope of the present application. The protection scope of the present application is subject to the protection scope defined by the claims.

Claims

1. An intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine, comprising the following steps: Step one: constructing and pre-processing medical dataset; including data screening and preprocessing, data normalization and data labeling, and unbalanced data balancing; Step two: Augment supervised samples and unsupervised samples, generate masks and features through the construction of mask generator and feature generator, integrate augmented output, so as to improve the training effect and generalization ability of the model; based on self-supervised learning, a series of enhancement transformations are performed on the original sample features , to generate new training samples and learn the potential relationship between samples; Step three: calculating reconstruction loss and consistency loss, including constructing a feature representation network model, calculating the corresponding logits of the features, calculating the reconstructed feature representation, calculating the reconstruction loss of the sample, and calculating the mask reconstruction loss; the reconstruction loss measures the difference between the reconstructed features and the original features, while the consistency loss ensures that the output of the neural network remains consistent under different conditions; Step four: using confidence to filter high-quality pseudo-labels, including generating pseudo-label vectors for unsupervised samples and filtering high-quality pseudo-labels; Step five: constructing the classification task loss function, including constructing the non-parallel intra-class loss function and constructing the non-parallel inter-class loss function; by combining the local information of intra-class and inter-class and adopting the parameter sharing mechanism, an optimal hyperplane is found for each class ; Step six: constructing a model overall loss function, including promoting intra-class sample compactness, inter-class sample discriminability, model reconstruction accuracy and prediction consistency; these target loss functions are fused into a unified loss function; Step seven: medical diagnosis hierarchical evaluation, realizing the hierarchical diagnosis of brain glioma data; for a new sample, i.e. brain glioma sample features , the category of the new sample is judged by calculating the distance to each non-parallel proximal hyperplane; each non-proximal hyperplane represents a category, wherein the categories include low-grade glioma LGG and glioblastoma GBM; The class of the nearest hyperplane is the class of the new sample.

2. The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machines as claimed in claim 1, wherein, Step one specifically includes: (1.1) Data screening and preprocessing; first, screening and analyzing clinical medical samples, extracting the 20 most frequent mutation genes and 2 clinical features: gender and diagnosis age; to ensure data integrity and quality, the original sample is preprocessed, including filtering instances with missing key information, converting diagnosis age from string format to continuous value, and excluding irrelevant columns, including medical record number, preliminary diagnosis and project; (1.2) Data normalization and data labeling; normalize the data samples; after normalization, label the samples according to the doctor's diagnosis result to obtain supervised samples wherein is a sample feature of dimension , is a sample corresponding label, is a supervised sample size; for samples without labels, it is an unsupervised sample wherein is an unsupervised sample size; (1.3) Unbalanced data balancing; check the preprocessed brain glioma dataset, and for the problem of unbalanced classes in the dataset, increase the number of minority class samples by randomly sampling the minority class samples to make the data balanced; specifically: first, calculate the number of samples of each class, denoted as , and then determine the target sample number; the sample number of the class with the most samples is , so the target is to increase the sample number of all classes to ; then calculate the number of samples to be increased, for each minority class , the number of samples to be increased is: (1) Finally, a number of samples are randomly selected from the minority class of samples and added to the dataset. ​ 3.The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine of claim 1, wherein, Step two specifically includes: (2.1) Construct a mask generator; the generator outputs a binary mask of dimension is denoted as where is randomly sampled in a Bernoulli distribution with probability , i.e. (2) (2.2) Construct a feature generator; this generator will input a mask and compute a corresponding subtraction operation , according to and sample features element-wise multiplication ; sample features are randomly shuffled column by column to get and mask element-wise multiplication, denoted as ; the final operation can be represented as: (3) wherein, is a feature generator, is an enhancing feature, is an element-wise multiplication, is a sample feature, is a randomly shuffled sample feature; equation (3) is interpreted as preserving the mask the part of the values in the mask the original sample; values of the part of the values in the mask (2.3) integrating augmented outputs; according to equation (3), augmenting features of the sample respectively supervised sample features , unsupervised sample features , and masks , obtaining corresponding supervised sample augmented features and unsupervised sample augmented features , . 4.The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine of claim 3, wherein, Step three specifically includes: (3.1) Constructing a feature representation network model ; Using network model To extract the feature representation of the sample in the nonlinear space, specifically, the network By encoder and decoder Composition, all adopt 3-layer fully connected layer structure; encoder Responsible for mapping the input data into a compact feature space, while the decoder It is used to reconstruct the input data, and the specific form is as follows: (4) wherein is a function composition symbol; composed by an encoder and a decoder composing a whole network; (3.2) compute feature respective logits; augment feature , , concatenate input encoder , to get respective logits, denoted as , , ; (5) wherein, and are network weights and biases, is a vector concatenation symbol; by concatenating the augmented features, it inputs the encoder and outputs the corresponding logits; (3.3) computing a reconstructed feature representation; augmenting the feature , , concatenating the input encoder-decoder model to obtain supervised sample reconstructed features unsupervised sample reconstructed features and mask reconstructed features : (6) wherein, and are network weights and biases; by concatenating the augmented features and inputting an encoder-decoder model, output corresponding reconstructed features; (3.4) Calculate the reconstruction loss of the sample; the reconstruction loss can evaluate the difference between the reconstructed features of the autoencoder and the original sample features; Calculate the supervised sample mean square error loss: (7) wherein, is the supervised sample size, the original sample feature and the sample reconstructed feature ; when minimizing the loss, the difference between and can be reduced; Calculate the reconstruction loss of the unsupervised sample: (8) wherein, is the unsupervised sample size, the original unsupervised sample feature and the reconstructed unsupervised sample feature ; to minimize the loss, reduce the difference between and , thereby ensuring the consistency of the reconstructed feature and the original sample feature; (3.5) Calculate the mask reconstruction loss; for the original mask and reconstruction mask , calculate the cross entropy loss: (9) wherein, is a mask of size ; the difference in probability distribution between the original mask and the reconstructed mask is evaluated by computing the cross-entropy loss; (3.6) computing a consistency loss for the unsupervised samples; computing unsupervised sample logits, and consistency loss: (10) wherein, is the unsupervised sample size; the prediction result after enhancement in the case of the same sample, and can still maintain consistency, so as to evaluate the ability of the model to maintain stability on the prediction level to the data change. 5.The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine of claim 4, wherein, Step four specifically includes: (4.1) generating pseudo label vector for unsupervised samples; described by formula (5), For unsupervised samples only through the encoder Get logits; for There are: (11) wherein, outputting, for each category, a category; denotes a vector outputting each category The resulting formula (11) is Taking the absolute value, we get ; The activation function softmin gives a mapping to real numbers between 0 and 1, noted as and the normalization guarantees the sum to be 1, i.e. (12) wherein, is the symbolic representation of the activation function softmin; the sum of all probabilities is 1; (4.2) Filter high-quality pseudo labels; introduce a scalar hyperparameter confidence threshold , and keep the class pseudo label if the prediction is above this threshold the maximum class probability in the batch and the confidence threshold , and keep the class pseudo label if the prediction is above this threshold (13) wherein, is an indicator function, is the maximum class probability of the unsupervised sample, when the maximum class probability is greater than a confidence threshold then 0 otherwise. According to the formula (13), if greater than the confidence threshold , the pseudo label is taken out as : (14) wherein the pseudo label is a pseudo label of the class with the highest probability in the middle. 6.The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine of claim 5, wherein, Step five specifically includes: optimal hyperplane is of the form: (15) wherein, is a class of near-end hyperplanes; and are weights and bias of the corresponding hyperplanes, is an encoder, is a sample feature; for class classification, construct binary classifiers, in the th classifier, the th class is the positive class, and the rest are trained as negative classes; the intra-class and inter-class loss calculation process is as follows: (5.1) Construct a non-parallel intra-class loss function; for all sample points of the class, construct an intra-class loss function; require the output of the class to be the minimum; take a sample point as an example, according to the real label and the pseudo label , first filter out the output of the class, denoted as: (16) Recalculating the minimum distance of the class, the loss function is constructed as follows: (17) Take the 2-norm of the class to get the distance from the origin. Take the 2-norm of the class to get the distance from the origin. Take the 2-norm of the class to get the distance from the origin. In addition, the minimum output of a class other than the class is required The distance from the hyperplane to which the class belongs is at least For a sample point , the loss function is constructed as follows: (18) (19) wherein is the margin; formula (18) gives the minimum output for classes other than the first class; formula (19) requires the margin distance for classes other than the first and second classes; and formula (20) gives the margin distance for the second class. (5.2) Construct a non-parallel inter-class loss function; for all sample points of the first class and classes other than the first class construct an inter-class loss function; require the minimum output of classes other than the first class to be at least distance from the first class, for a sample point , the loss function is constructed as follows: (20) wherein, is the margin, equation (20) requires that the class other than the class of classes have a margin of distance.

7. The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machines as claimed in claim 6, wherein, Step six specifically includes: (6.1) Fusion of intra-class and inter-class loss functions; by combining equations (17), (19) and (20): (21) wherein , is a regularization term, , , is a trade-off parameter; For all supervised samples, the supervised loss function is: (22) The unsupervised loss is: (23) In formula (23), is an indicator function, is the maximum class probability of the unsupervised sample, when the maximum class probability is greater than the confidence threshold then 0 otherwise. (6.2) Construct a unified model overall loss function; according to the above losses (22), (23) and reconstruction losses (7), (8), (9) and consistency losses (10), the overall loss function is integrated; (24) wherein , , is a trade-off parameter.

8. The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machines as claimed in claim 7, wherein, In step seven, the hierarchical diagnosis of brain glioma data is realized, and the decision function is as follows: (25) wherein denotes the absolute value, denotes the first category. 9.The intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine of claim 8, wherein: The data of step one is from the low-grade glioma (LGG) and glioblastoma (GBM) brain glioma sub-project of the Cancer Genome Atlas (TCGA) project; The confidence threshold of step four is 0.95; the trade-off parameter of step six , , the value of the trade-off parameter is taken to be 1.

10. An intelligent medical diagnostic apparatus based on semi-supervised deep non-parallel support vector machines, characterized by: A memory and one or more processors, the memory has stored executable code, and the one or more processors execute the executable code to implement the intelligent medical diagnosis method based on semi-supervised deep non-parallel support vector machine of any one of claims 1-9.

Images