Multi-omics cancer image classification method and system combining graph automatic encoder and dynamic evidence fusion

Abstract

The invention relates to a multi-omics cancer image classification method and system combining a graph automatic encoder and dynamic evidence fusion, and belongs to the field of bioinformatics and medical diagnosis. The method comprises the following steps: capturing a complex relationship between samples through a graph automatic encoder, compressing high-dimensional omics data into low-dimensional potential representation, and meanwhile, ensuring that original structure information is reserved through reconstruction loss; a dynamic evidence fusion technology is adopted, the quality of each group of study data is evaluated based on subjective logic, and higher weight is given to high-quality data by evaluating category probability distribution and uncertainty of samples. According to the method, the relationship between different omics data can be better captured, so that the classification precision of the model is improved. The dynamic evidence fusion technology further integrates the uncertainty in different omics data, and improves the robustness of the model on heterogeneous data.

Description

Technical Field

[0001] The present invention relates to a multi-omics cancer image classification method and system combining a graph autoencoder and dynamic evidence fusion, and belongs to the field of bioinformatics and medical diagnosis. Background Art

[0002] The accuracy of cancer classification is the key to achieving more precise and personalized cancer diagnosis and treatment. Multi-omics data can provide more comprehensive and multi-level cancer information. Therefore, cancer image classification methods that integrate multi-omics data have become one of the effective means to deal with the complexity of cancer. At the same time, this method also shows great potential in improving the accuracy of cancer image classification. With the increasing maturity of high-throughput sequencing technology, multi-omics data has become easier to obtain, and researchers can comprehensively study and analyze cancer from multiple molecular levels such as genomics, transcriptomics, and proteomics. However, with the rapid increase in the amount of multi-omics data, how to effectively integrate multi-omics data to improve the accuracy of cancer image classification has become a difficult problem that needs to be solved urgently in this field.

[0003] The high dimensionality and heterogeneity of multi-omics data are its two main characteristics. High dimensionality means that each sample in the omics data contains a large number of features. This often leads to increased computational complexity and may cause overfitting. Heterogeneity refers to the difference in quality of different omics data. Due to the uncertainty in the generation of multi-omics data, it is difficult to completely avoid noise, damage, or data out of the distribution interval. Therefore, there are significant differences in the quality of omics. Existing methods usually regard the quality of omics as equally important, ignoring this difference, which may lead to misleading classification results, which is obviously inconsistent with the actual situation. Therefore, addressing these two issues is the key to achieving effective multi-omics data integration. Summary of the invention

[0004] The present invention provides a multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion, which is used to solve the problem of low cancer image classification accuracy caused by the heterogeneity of multi-omics data and the problem of increased computational complexity caused by the high dimensionality of multi-omics data.

[0005] The technical solution of the present invention is: a multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion, the method comprising:

[0006] Step 1: Use graph autoencoders to handle the high dimensionality of multi-omics data;

[0007] Step 2: Use dynamic evidence fusion to deal with the heterogeneity of multi-omics data.

[0008] Furthermore, in the Step 1, the complex relationship between samples is captured by a graph autoencoder, and the high-dimensional omics data is compressed into a low-dimensional potential representation, while the reconstruction loss is used to ensure that the original structural information is retained.

[0009] Furthermore, in Step 2, dynamic evidence fusion technology is used to evaluate the quality of each omics data, and high-quality data is weighted by evaluating the category probability distribution and uncertainty of the samples.

[0010] Furthermore, the Step 1 includes:

[0011] Step 1.1. Build a graph structure for each omics data to generate an initial fully connected graph. The nodes in the fully connected graph represent the samples in the omics data, and the edges represent the similarity between samples. The edges are constructed by calculating the cosine similarity between samples, and then the adjacency matrix is ​​constructed.

[0012] The elements in the adjacency matrix, i.e. the cosine similarity A between samples ij The calculation formula is shown in (1):

[0013]

[0014] Among them, s(x i ,x j ) refers to the cosine similarity function, x i 、x j represents the sample feature vector, ∈ is the cosine similarity threshold, which is used to filter edges to construct a sparse graph to improve computational efficiency and preserve local similarity; i and j represent two arbitrary sample numbers in the omics data, and n represents the total number of samples in the omics data;

[0015] Step 1.2: Use a graph autoencoder to represent high-dimensional omics data as a low-dimensional latent representation. The graph autoencoder consists of three graph convolutional layers (GCNs). The feature representation matrix calculation formula is shown in (2):

[0016]

[0017] Where A is the adjacency matrix of the graph (including self-loops); D is the degree matrix of the adjacency matrix; H (l) is the feature representation matrix of the lth layer (initially the input feature matrix); W (l) is the trainable weight matrix of the graph convolutional layer, σ is the ReLU activation function; H (l+1) is the feature representation matrix of the l+1th layer. After multi-layer graph convolution operations, the output H (l)It is a low-dimensional potential representation of each sample. This low-dimensional representation not only compresses the original high-dimensional data, but also retains the graph structure information between samples.

[0018] Step 1.3, use the decoder to reconstruct the original graph structure data from the low-dimensional latent representation to ensure that the latent representation effectively retains the original structure information; the decoder reconstructs the adjacency matrix in a similar way to the inner product. The calculation formula for reconstructing the adjacency matrix is ​​shown in (3):

[0019]

[0020] where Z is the low-dimensional latent representation matrix obtained from the encoder, The adjacency matrix between samples output from the decoder is reconstructed and compared with the original adjacency matrix A to ensure that sufficient information about the relationship between samples is retained in the potential representation;

[0021] The reconstruction error and the regularized loss (KL divergence) of the latent representation are used to optimize the model. The overall loss function calculation formula is shown in (4):

[0022]

[0023] in, represents the overall loss function, i.e., the parameters used to optimize the entire model; Represents the reconstruction loss, that is, comparing the original matrix A with the reconstructed matrix The difference between ; β represents the weight coefficient, which is used to balance the relative influence of the reconstruction error and the KL divergence term; Represents KL divergence loss, which makes the distribution of potential representation Z close to the given prior distribution and realizes the regularization of latent space;

[0024] The reconstruction error is used to measure the quality of the reconstructed image, and the regularization loss is used to constrain the latent space so that the encoder output has a good distribution;

[0025] Step 1.4, generate the category probability distribution, that is, generate the category probability distribution of the sample through the classifier and use it as the input of the variational Dirichlet distribution; input the low-dimensional potential representation of each sample into the fully connected layer, and generate the category probability distribution of the sample through the Softmax activation function. The calculation formula of the category probability distribution P(y=c|x) of the sample is shown in (5):

[0026]

[0027] Among them, C represents the total number of categories, Z is the low-dimensional representation of the sample, W is the weight of the classifier, and W c is the weight vector of the cth class, is its transpose (row vector form), and the inner product with the input feature vector Z obtains the unnormalized score for the category; k in the summation symbol is used to traverse the index of all categories, from 1 to C (the total number of categories); x represents the high-dimensional feature space representation of the original input sample, y represents the category label of the sample, and y=c means that the sample belongs to category c.

[0028] Furthermore, the Step 2 includes:

[0029] Step 2.1, using the category probability distribution of the previous step as input, train and estimate the variational Dirichlet distribution polynomial concentration parameter of the sample to reflect the distribution uncertainty of the sample in each category. Specifically, the category probability distribution is used as input, and the variational Dirichlet distribution is used to fuse the relevant evidence to output the concentration parameter of the sample in each category α=[α1,...,α k ], in order to deal with the inaccuracy of the prior distribution, the variational technique (KL divergence) is introduced to train and optimize the concentration parameter α; the calculation formula of the variational Dirichlet distribution is shown in (6):

[0030]

[0031] Where B(α) is a normalization factor to ensure that the Dirichlet distribution integral is 1; K is the total number of categories or the number of vector dimensions; is the i-th component of a set of probability distribution parameters on the simplex, representing the probability value of category i, α i It is a hyperparameter (concentration parameter) that controls the shape of the distribution and determines the degree of prior preference for the probability component of a certain category; S K is the K-dimensional simplex to which μ belongs, i.e., the definition of μ;

[0032] Since the Dirichlet distribution only describes the prior distribution of class probabilities, directly calculating the posterior distribution usually leads to more inaccurate classification results; therefore, the variational technique is introduced to adjust the concentration parameter by optimizing the variational lower bound so that the target approximates the posterior distribution; the following is a detailed introduction to the application of the variational technique on the Dirichlet distribution:

[0033] (1) Dirichlet distribution parameters:

[0034] Assume that a sample set is given The omics number is m, the sample label For omics m, the following formulas (7) and (8) represent the calculation process of the Dirichlet distribution parameters, which are specifically based on the category probability distribution μ m And the label distribution y is calculated as:

[0035]

[0036] y~p(y|μ m)=Mult(y|μ m )(8)

[0037] Among them, μ m Representing omics data through a graph autoencoder (GAE) Output class probability distribution, θ m are the model parameters used to generate the parameters of the Dirichlet distribution; the model is derived from the input data x m Through the parameter θ m Generates the concentration parameter α of the Dirichlet distribution m , then from α m Sampling the class probability vector μ m ;Mult(y|μ m ) means a multinomial distribution of labels;

[0038] (2) Variational lower bound:

[0039] It can be expressed by the following formula (9):

[0040]

[0041] in, is the parameter θ m The probability encoder under is similar to the one introduced earlier. D KL represents the KL divergence, Represents the marginal likelihood The lower bound of can be expressed by formula (10) as follows:

[0042]

[0043] Step 2.2, based on the optimized variational Dirichlet distribution parameters, calculate the uncertainty u of the sample under each omics and the belief quality of the sample category under the omics The combination of the two can evaluate the quality of each omics and achieve multi-omics integration under the guidance of simplified DST combination rules;

[0044] (1) Evaluate the quality of omics. For each omics m, the calculation relationship is shown in (11):

[0045]

[0046] Among them, u m represents the classification uncertainty at the m-th omics level; represents the belief quality or credibility assignment of the m-th omics level to category k; K is the total number of categories, which is the number of defined classification dimensions here.

[0047] (2) Calculate the belief quality and uncertainty. The calculation formula is shown in (12):

[0048]

[0049] in, It is called Dirichlet intensity; is the basic parameter or amount of evidence related to category k under the m-th omics, Depend on The converted parameters Usually one of the parameters of the Dirichlet distribution; S in the formula m It reflects that m is the dimension (or data source) index;

[0050] ③Calculate the quality of omics:

[0051] All omics can be represented as a set Each omics representation is shown in (13):

[0052]

[0053] According to the simplified DST combination rules, the integration is done as follows:

[0054]

[0055] M is the final comprehensive quality function integrating all M^m using the DST synthesis rule, representing the overall belief distribution after the fusion of multi-omics evidence; M V is the quality function at the v-th omics level, which contains the belief quality assignments for each category and uncertainty.

[0056] The present invention also provides a system for multi-omics cancer image classification combining graph autoencoder and dynamic evidence fusion, the system comprising: a module for executing the above-mentioned multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion.

[0057] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion is implemented.

[0058] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion.

[0059] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the above-mentioned multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion.

[0060] The beneficial effects of the present invention are:

[0061] 1. This paper effectively solves the high dimensionality and heterogeneity problems of multi-omics data by combining graph autoencoders with dynamic evidence fusion technology; graph autoencoders can compress high-dimensional data into low-dimensional potential representations while retaining the structural information of the data, greatly reducing the computational complexity;

[0062] 2. By utilizing the low-dimensional feature extraction capability of graph autoencoders, the present invention can better capture the relationship between different omics data, thereby improving the classification accuracy of the model; dynamic evidence fusion technology further integrates the uncertainty in different omics data and improves the robustness of the model on heterogeneous data;

[0063] 3. By modeling category uncertainty using variational Dirichlet distribution, the present invention can not only provide classification results, but also provide the distribution of samples in different categories, helping researchers to better understand the decision-making process of the model and improve the interpretability of the results;

[0064] 4. The present invention provides new technical support for assisting in early cancer screening, efficacy evaluation, prognosis analysis, and studying the molecular mechanisms of cancer; the present invention can be applied to a variety of biomedical fields such as cancer image classification, disease prediction, and drug target discovery, especially in the research of processing multi-omics data, and has high application value; the present invention has powerful feature extraction and data fusion capabilities, which enables it to adapt to different biological problems and has broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a model diagram in the present invention;

[0066] Figure 2 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0067] Example 1: Figure 1-Figure 2 As shown, a multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion, the method comprising:

[0068] Step 1: Use graph autoencoders to handle the high dimensionality of multi-omics data;

[0069] Furthermore, in the Step 1, the complex relationship between samples is captured by a graph autoencoder, and the high-dimensional omics data is compressed into a low-dimensional potential representation, while the reconstruction loss is used to ensure that the original structural information is retained.

[0070] Furthermore, the Step 1 includes:

[0071] Step 1.1. Build a graph structure for each omics data to generate an initial fully connected graph. The nodes in the fully connected graph represent the samples in the omics data, and the edges represent the similarity between samples. The edges are constructed by calculating the cosine similarity between samples, and then the adjacency matrix is ​​constructed.

[0072] The elements in the adjacency matrix, i.e. the cosine similarity A between samples ij The calculation formula is shown in (1):

[0073]

[0074] Among them, s(x i ,x j ) refers to the cosine similarity function, x i 、x j represents the sample feature vector, ∈ is the cosine similarity threshold, which is used to filter edges to construct a sparse graph to improve computational efficiency and preserve local similarity; i and j represent two arbitrary sample numbers in the omics data, and n represents the total number of samples in the omics data;

[0075] Step 1.2: Use a graph autoencoder to represent high-dimensional omics data as a low-dimensional latent representation. The graph autoencoder consists of three graph convolutional layers (GCNs). The feature representation matrix calculation formula is shown in (2):

[0076]

[0077] Where A is the adjacency matrix of the graph (including self-loops); D is the degree matrix of the adjacency matrix; H (l) is the feature representation matrix of the lth layer (initially the input feature matrix); W (l) is the trainable weight matrix of the graph convolutional layer, σ is the ReLU activation function; H (l+1) is the feature representation matrix of the l+1th layer. After multi-layer graph convolution operations, the output H (l) It is a low-dimensional potential representation of each sample. This low-dimensional representation not only compresses the original high-dimensional data, but also retains the graph structure information between samples.

[0078] Step 1.3, use the decoder to reconstruct the original graph structure data from the low-dimensional latent representation to ensure that the latent representation effectively retains the original structure information; the decoder reconstructs the adjacency matrix in a similar way to the inner product. The calculation formula for reconstructing the adjacency matrix is ​​shown in (3):

[0079]

[0080] where Z is the low-dimensional latent representation matrix obtained from the encoder, The adjacency matrix between samples output from the decoder is reconstructed and compared with the original adjacency matrix A to ensure that sufficient information about the relationship between samples is retained in the potential representation;

[0081] The reconstruction error and the regularized loss (KL divergence) of the latent representation are used to optimize the model. The overall loss function calculation formula is shown in (4):

[0082]

[0083] in, represents the overall loss function, i.e., the parameters used to optimize the entire model; Represents the reconstruction loss, that is, comparing the original matrix A with the reconstructed matrix The difference between ; β represents the weight coefficient, which is used to balance the relative influence of the reconstruction error and the KL divergence term; Represents KL divergence loss, which makes the distribution of potential representation Z close to the given prior distribution and realizes the regularization of latent space;

[0084] The reconstruction error is used to measure the quality of the reconstructed image, and the regularization loss is used to constrain the latent space so that the encoder output has a good distribution;

[0085] Step 1.4, generate the category probability distribution, that is, generate the category probability distribution of the sample through the classifier and use it as the input of the variational Dirichlet distribution; input the low-dimensional potential representation of each sample into the fully connected layer, and generate the category probability distribution of the sample through the Softmax activation function. The calculation formula of the category probability distribution P(y=c|x) of the sample is shown in (5):

[0086]

[0087] Among them, C represents the total number of categories, Z is the low-dimensional representation of the sample, W is the weight of the classifier, and W c is the weight vector of the cth class, is its transpose (row vector form), and the inner product with the input feature vector Z obtains the unnormalized score for the category; k in the summation symbol is used to traverse the index of all categories, from 1 to C (the total number of categories); x represents the high-dimensional feature space representation of the original input sample, y represents the category label of the sample, and y=c means that the sample belongs to category c.

[0088] Step 2: Use dynamic evidence fusion to deal with the heterogeneity of multi-omics data.

[0089] Furthermore, in Step 2, dynamic evidence fusion technology is used to evaluate the quality of each omics data, and high-quality data is weighted by evaluating the category probability distribution and uncertainty of the samples.

[0090] Furthermore, the Step 2 includes:

[0091] Step 2.1, using the category probability distribution of the previous step as input, train and estimate the variational Dirichlet distribution polynomial concentration parameter of the sample to reflect the distribution uncertainty of the sample in each category. Specifically, the category probability distribution is used as input, and the variational Dirichlet distribution is used to fuse the relevant evidence to output the concentration parameter of the sample in each category α=[α1,...,α k ], in order to deal with the inaccuracy of the prior distribution, the variational technique is introduced to train and optimize the concentration parameter α; the calculation formula of the variational Dirichlet distribution is shown in (6):

[0092]

[0093] Where B(α) is a normalization factor to ensure that the Dirichlet distribution integral is 1; K is the total number of categories or the number of vector dimensions; is the i-th component of a set of probability distribution parameters on the simplex, representing the probability value of category i, α i It is a hyperparameter (concentration parameter) that controls the shape of the distribution and determines the degree of prior preference for the probability component of a certain category; S K is the K-dimensional simplex to which μ belongs, i.e., the definition of μ;

[0094] Since the Dirichlet distribution only describes the prior distribution of class probabilities, directly calculating the posterior distribution usually leads to more inaccurate classification results; therefore, the variational technique is introduced to adjust the concentration parameter by optimizing the variational lower bound so that the target approximates the posterior distribution; the following is a detailed introduction to the application of the variational technique on the Dirichlet distribution:

[0095] (1) Dirichlet distribution parameters:

[0096] Assume that a sample set is given The omics number is m, the sample label For omics m, the following formulas (7) and (8) represent the calculation process of the Dirichlet distribution parameters, which are specifically based on the category probability distribution μ m And the label distribution y is calculated as:

[0097]

[0098] Among them, μ m Representing omics data through a graph autoencoder (GAE) Output class probability distribution, θ m are the model parameters used to generate the parameters of the Dirichlet distribution; the model is derived from the input data x m Through the parameter θ m Generates the concentration parameter α of the Dirichlet distribution m , then from α m Sampling the class probability vector μ m ;Mult(y|μ m ) means a multinomial distribution of labels;

[0099] (2) Variational lower bound:

[0100] It can be expressed by the following formula (9):

[0101]

[0102] in, is the parameter θ m The probability encoder under is similar to the one introduced earlier. D KL represents the KL divergence, Represents the marginal likelihood The lower bound of can be expressed by formula (10) as follows:

[0103]

[0104] Step 2.2, based on the optimized variational Dirichlet distribution parameters, calculate the uncertainty u of the sample under each omics and the belief quality of the sample category under the omics The combination of the two can evaluate the quality of each omics and achieve multi-omics integration under the guidance of simplified DST combination rules;

[0105] (1) Evaluate the quality of omics. For each omics m, the calculation relationship is shown in (11):

[0106]

[0107] Among them, u m represents the classification uncertainty at the m-th omics level; represents the belief quality or credibility assignment of the m-th omics level to category k; K is the total number of categories, which is the number of defined classification dimensions here.

[0108] (2) Calculate the belief quality and uncertainty. The calculation formula is shown in (12):

[0109]

[0110] in, It is called Dirichlet intensity; is the basic parameter or amount of evidence related to category k under the m-th omics, Depend on The converted parameters Usually one of the parameters of the Dirichlet distribution; S in the formula m It reflects that m is the dimension (or data source) index;

[0111] (3) Calculate the quality of omics:

[0112] All omics can be represented as a set Each omics representation is shown in (13):

[0113]

[0114] According to the simplified DST combination rules, the integration is done as follows:

[0115]

[0116] M is the final comprehensive quality function integrating all M^m using the DST synthesis rule, representing the overall belief distribution after the fusion of multi-omics evidence; M V is the quality function at the v-th omics level, which contains the belief quality assignments for each category and uncertainty.

[0117] The present invention also provides a system for multi-omics cancer image classification combining graph autoencoder and dynamic evidence fusion, the system comprising:

[0118] A processing module for handling the high dimensionality of multi-omics data using graph autoencoders;

[0119] Fusion module for handling heterogeneity of multi-omics data using dynamic evidence fusion.

[0120] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion is implemented.

[0121] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion.

[0122] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the above-mentioned multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion.

[0123] In order to verify the effect of the present invention, experiments were conducted on the BRCA and ROSMAP datasets using existing methods and the method of the present invention. Experiments on the BRCA and ROSMAP datasets show that compared with the best existing methods, the method proposed in the present invention improves ACC, F1_weighted and F1_macro by 4.34%, 4.97% and 7.62% on the BRCA dataset, respectively; on the ROSMAP dataset, it improves by 1.60%, 0.97% and 3.78%, respectively. The experimental results show that this method can effectively deal with the quality differences in multi-omics data and significantly improve the classification accuracy.

[0124] Table 1 shows the experimental results of the BRCA dataset

[0125]

[0126]

[0127] Table 2 Experimental results of the ROSMAP dataset

[0128]

[0129] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A multi-omics cancer image classification method combining graph autoencoders and dynamic evidence fusion, characterized by: The method comprises: Step 1: Use graph autoencoders to handle the high dimensionality of multi-omics data; Step 2: Use dynamic evidence fusion to deal with the heterogeneity of multi-omics data.

2. A multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion according to claim 1, characterized in that: In Step 1, the complex relationship between samples is captured by a graph autoencoder, and the high-dimensional omics data is compressed into a low-dimensional potential representation, while the reconstruction loss is used to ensure that the original structural information is preserved.

3. The multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion according to claim 1, characterized in that: In Step 2, dynamic evidence fusion technology is used to evaluate the quality of each group of data, and high-quality data is weighted by evaluating the category probability distribution and uncertainty of the samples.

4. The multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion according to claim 1, characterized in that: The Step 1 includes: Step 1.

1. Build a graph structure for each omics data to generate an initial fully connected graph. The nodes in the fully connected graph represent the samples in the omics data, and the edges represent the similarity between samples. The edges are constructed by calculating the cosine similarity between samples, and then the adjacency matrix is ​​constructed. The elements in the adjacency matrix, i.e. the cosine similarity A between samples ij The calculation formula is shown in (1): Among them, s(x i ,x j ) refers to the cosine similarity function, x i 、x j represents the sample feature vector, ∈ is the cosine similarity threshold, i and j represent two arbitrary sample numbers in the omics data, and n represents the total number of samples in the omics data; Step 1.2: Use graph autoencoders to represent high-dimensional omics data as low-dimensional latent representations. The graph autoencoder consists of three graph convolutional layers, and the feature representation matrix calculation formula is shown in (2): Among them, A is the adjacency matrix of the graph, D is the degree matrix of the adjacency matrix, and H (l) is the feature representation matrix of the lth layer, W (l) is the trainable weight matrix of the graph convolutional layer, σ is the ReLU activation function, and H (l+1) is the feature representation matrix of the l+1th layer; Step 1.3, use the decoder to reconstruct the original graph structure data from the low-dimensional potential representation; the decoder reconstructs the adjacency matrix in a similar way to the inner product. The calculation formula for reconstructing the adjacency matrix is ​​shown in (3): where Z is the low-dimensional latent representation matrix obtained from the encoder, is the reconstructed adjacency matrix between samples output from the decoder; The reconstruction error and the regularized loss of the latent representation are used to optimize the model. The overall loss function calculation formula is shown in (4): in, represents the overall loss function, represents the reconstruction loss, β represents the weight coefficient, represents KL divergence loss; Step 1.4, input the low-dimensional potential representation of each sample into the fully connected layer, and generate the class probability distribution of the sample through the Softmax activation function. The calculation formula of the class probability distribution of the sample P(y=c|x) is shown in (5): Among them, C is the number of categories, Z is the low-dimensional representation of the sample, W is the weight of the classifier, and W c is the weight vector of the cth category, k is the index used to traverse all categories, x represents the representation of the high-dimensional feature space of the original input sample, y represents the category label of the sample, and y=c means that the sample belongs to category c.

5. The multi-omics cancer image classification method combining graph autoencoder and dynamic evidence fusion according to claim 1, characterized in that: The Step 2 includes: Step 2.1, take the category probability distribution as input, use the variational Dirichlet distribution to fuse the relevant evidence and output the concentration parameter α of the sample in each category = [α1,...,α k ], the variational technique is introduced to train and optimize the concentration parameter α; the calculation formula of the variational Dirichlet distribution is shown in (6): Where B(α) is the normalization factor, K is the total number of categories or the number of vector dimensions, is the probability value of category i, α i is a hyperparameter that controls the shape of the distribution, S K is the K-dimensional simplex to which μ belongs; Step 2.2, based on the optimized variational Dirichlet distribution parameters, calculate the uncertainty u of the sample under each omics and the belief quality of the sample category under the omics (1) Evaluate the quality of omics. For each omics m, the calculation relationship is shown in (7): Among them, u m represents the classification uncertainty at the m-th omics level; represents the belief quality or credibility assignment of the m-th omics level to category k; K is the total number of categories; (2) Calculate the belief quality and uncertainty. The calculation formula is shown in (8): in, It is called Dirichlet intensity; is the basic parameter or amount of evidence related to category k under the m-th omics, is one of the parameters of the Dirichlet distribution, (3) Calculate the quality of omics: All omics can be represented as a set Each omics representation is shown in (9): According to the simplified DST combination rules, the integration is done as follows: Where M is the final comprehensive quality function integrating all M^m using the DST synthesis rule, M V is the quality function at the V-omics level, which contains the belief quality assignments for each category and uncertainty.

6. A system for multi-omics cancer image classification combining graph autoencoders and dynamic evidence fusion, characterized in that The system comprises: a module for executing a multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion as described in any one of claims 1 to 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, a multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion is implemented as described in any one of claims 1 to 3.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements a multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements a multi-omics cancer image classification method combining a graph autoencoder and dynamic evidence fusion as described in any one of claims 1 to 5.

Images