Cervical cancer brachytherapy segmentation mode prediction method based on graph neural network

Abstract

The invention provides a cervical cancer brachytherapy segmentation mode prediction method based on a graph neural network. The method comprises the following steps: acquiring a cervical cancer data set; the cervical cancer data set is an electronic medical record containing clinical treatment information of each cervical cancer patient; preprocessing the cervical cancer data set; inputting the preprocessed cervical cancer data set into a GPT2 model for data generation so as to generate synthetic data; evaluating the quality of the synthetic data, and selecting the synthetic data with the highest similarity with the preprocessed cervical cancer data set; and selecting synthetic data with the highest similarity with the preprocessed cervical cancer data set as basic data, capturing a treatment feature relationship of the cervical cancer patient by using a dynamic feature aggregation graph neural network, and predicting a corresponding treatment scheme. According to the method, the synthetic data generated based on the GPT2 model is used, the problem of insufficient medical data is solved, and the treatment scheme of the patient is more accurately and effectively predicted by adopting the method of dynamically capturing the relation between the patient features through the graph neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of cervical cancer treatment segmentation pattern prediction, and in particular to a cervical cancer brachytherapy segmentation pattern prediction method based on graph neural network. Background Art

[0002] Cervical cancer is one of the four most common cancers in women worldwide and can be treated with surgery in its early stages. As the disease progresses, radiotherapy becomes the main treatment. The combination of external beam radiotherapy (EBRT) and intracavitary radiotherapy (i.e., brachytherapy) is now considered to be the most effective treatment for most patients with cervical cancer. Brachytherapy for cervical cancer refers to a radiotherapy method in which a radiation source is placed directly into or close to the tumor area for irradiation. This method plays an important role in the treatment of cervical cancer, especially in locally advanced or recurrent cases. Therefore, studying the relationship between patient clinical characteristics and radiotherapy regimens is crucial to developing personalized brachytherapy patterns.

[0003] However, challenges in cervical cancer treatment include the scarcity of chronic disease data, the lengthy treatment process, and the need for experienced planners to invest a lot of time in data accumulation. Models trained based solely on the plans of individual planners may inadvertently incorporate their subjective biases, making it very difficult to obtain a comprehensive collection of high-quality radiotherapy plans in clinical settings. Developing dose allocation plans for brachytherapy for cervical cancer requires multiple considerations, such as treatment range, duration, and dose selection. Although existing evidence-based clinical guidelines, such as the NCCN guidelines, provide basic treatment benchmarks, the diversity and complexity of patient conditions require personalized adjustments, which is a major challenge in clinical practice.

[0004] In recent years, the application of graph neural networks (GNNs) in healthcare has made remarkable progress, especially in predicting treatment plans based on clinical indicators. These advances are due to the ability of GNNs to interpret complex relationships in medical data, paving the way for personalized and precise medical interventions. A notable innovation in this field is the application of graph convolutional neural networks (Graph-CNNs) to gene expression data analysis to predict medical outcomes, such as breast cancer metastasis. This approach models molecular networks as graphs and assigns gene expression data to nodes. As a result, Graph-CNNs are able to identify gene expression patterns for specific health conditions. For example, the Graph Level Relevance Propagation (GLRP) technique has been adapted for use in Graph-CNNs, enabling them to generate interpretable, patient-specific molecular subnetworks. These subnetworks are essential for identifying drivers of tumor progression and helping to tailor precision medicine strategies. On an extensive breast cancer dataset, this approach played a key role in predicting distant metastasis and identifying patient-specific therapeutic vulnerabilities. In the field of bioinformatics, GNNs are widely used in node classification, link prediction, and graph generation, and are particularly good at biological network data analysis. Its application range covers a wide range of fields such as disease prediction, drug discovery, and biomedical imaging. The inherent advantages of GNNs, such as the ability to handle nonlinear patterns, interpretability, handling non-Euclidean geometry, modularity, and cross-modality, make them well suited for medical data analysis.

[0005] The treatment cost prediction technology solution recorded in the related technology (patent number: CN202110108182.4) integrates Bayesian network and regression analysis technology to improve the accuracy and robustness of treatment cost prediction when the amount of data is small, but it has defects in capturing the complex relationship between features; the treatment plan generation method technology solution recorded in another related technology (patent number: CN201811407145.8) uses a feedforward deep neural network technology with time connection to generate the patient's current treatment plan, and predicts the patient's future treatment plan through a sequence-to-sequence model, but requires a large amount of training data. The above technology has the following technical problems:

[0006] (1) The nonlinear relationship between multiple factors of patients is still a difficult problem in current clinical practice. When the relationship between data is very complex and difficult to express with simple rules, it is impossible to capture this complexity;

[0007] (2) It requires manual design and extraction of key features from data, which consumes a lot of time and manpower and is cumbersome;

[0008] (3) A large amount of data is required to train deep learning models. However, for some rare and chronic diseases, this technology cannot be used due to the scarcity of data.

[0009] Therefore, we urgently need a new method for predicting the segmentation pattern of cervical cancer brachytherapy to solve the technical problems not solved by the above technologies. Summary of the invention

[0010] The present invention aims to provide a method of using synthetic data generated by a generative pre-trained model GPT2 model to solve the problem of insufficient medical data, and adopt a method of dynamically capturing the relationship between patient characteristics using a graph neural network to more accurately and effectively predict the patient's treatment plan.

[0011] To this end, the purpose of the present invention is to propose a method for predicting segmentation patterns in cervical cancer brachytherapy based on graph neural networks.

[0012] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a method for predicting the segmentation pattern of cervical cancer brachytherapy based on graph neural network, the method comprising: step S1: obtaining a cervical cancer dataset; the cervical cancer dataset is an electronic medical record containing clinical consultation information of each cervical cancer patient; step S2: preprocessing the cervical cancer dataset; step S3: inputting the preprocessed cervical cancer dataset into the GPT2 model for data generation to generate synthetic data; step S4: evaluating the quality of the synthetic data, and selecting the synthetic data with the highest similarity to the preprocessed cervical cancer dataset; step S5: selecting the synthetic data with the highest similarity to the preprocessed cervical cancer dataset as the basic data, using the dynamic feature aggregation graph neural network to capture the consultation feature relationship of cervical cancer patients, and predicting the corresponding treatment plan.

[0013] Preferably, the step S2 specifically includes: step S2.1: using the KNN interpolation method to process the missing values ​​in the cervical cancer dataset to obtain an estimated value corresponding to the missing value of each sample in the cervical cancer dataset, so as to fill the missing value with the estimated value; step S2.2: using the SMOTE balancing algorithm and the ADASYN balancing algorithm respectively to balance the cervical cancer dataset after KNN interpolation processing.

[0014] Preferably, the step S3 specifically comprises: step S3.1: using the GPT2 model to create a vocabulary for the input data, mapping the unique tag in each column of the input data to a unique tag ID, and then applying the mapping in the vocabulary to each row in the input data to generate a series of tag IDs; wherein the input data is the preprocessed cervical cancer dataset, and the preprocessed cervical cancer dataset is a table data T with m rows and n columns;

[0015] Step S3.2: Use the GPT2 model to simulate the column value conditional distribution P(xij |x i1 , ..., x i(j-1) ), where x ij is the value of the i-th row and j-th column of the data set;

[0016] Step S3.3: Generate synthetic data by using the target masking and overfitting monitoring strategies; wherein the synthetic data consists of a series of observation values ​​O i Composition; the observed value O i The expression is:

[0017] O i =[x i1 , x i2 , ..., x ij , .., x in ] (1)

[0018] In formula (1), i represents the i-th row; j represents the j-th column;

[0019] Step S3.4: Based on the preprocessed cervical cancer dataset, five different data scales are generated to determine the optimal size of the synthetic data.

[0020] Preferably, the step S4 specifically includes: step S4.1: using the three measurement indicators of PCD, KLD and MMD to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset; step S4.2: using a triangular radar chart drawn on a logarithmic scale to represent the average values ​​of the three measurement indicators PCD, KLD and MMD; wherein, in the triangular radar chart, triangles with smaller areas represent higher similarity between the synthetic data and the preprocessed cervical cancer dataset, that is, the synthetic data has higher quality; step S4.3: comparing the variability of the SMOTE balancing algorithm and the ADASYN balancing algorithm through several iterations, and displaying the results using a box plot.

[0021] Preferably, the step S5 specifically includes: step S5.1: expanding the feature dimension of the low-dimensional input X, and then using a learnable matrix for nonlinear mapping to enhance the expressiveness of the feature; as shown in the following formula:

[0022] X out1 =σ(W 1 (Unsqueeze(X))) (2)

[0023] In formula (2), the low-dimensional input X is the synthetic data preprocessed in step S4; Unsqueeze means expanding the dimension of X; W1 means the feature-rich weight; σ means the activation function; the original input data is converted through the above process to promote the transition from the original space to the feature-rich space; X out1 is the output data after step S5.1;

[0024] Step S5.2: Introduce a dynamic adjustment method to improve the feature generalization ability of the input features to flexibly convert the input into a graph structure:

[0025] X out2 =W 2 (DFL(Reshape(X out1 ))) (3)

[0026] In formula (3), Reshape means using DFL(·) based on two-dimensional convolution to perform dimension transformation operation, DFL(·) means using dynamically adjusted feature aggregation, W 2 represents the weight of the 1×1 convolution used for feature enrichment; X out2 is the output data after step S5.2;

[0027] For DFL, assume that two hidden feature labels x i and x j , whose amplitude is z and phase is θ, their superposition can be calculated as follows:

[0028]

[0029] In formula (4), θ i is the hidden feature x i The phase of θ j is the hidden feature x j Phase; ⊙ is element-wise multiplication; |z i | is the hidden feature x i The amplitude of |z j | is the hidden feature x j The amplitude of |z mix | is the amplitude aggregation result of two hidden features;

[0030] Aggregate phase term θ r The calculation is as follows:

[0031] θ δ =θj-θ i

[0032] θ r =θ i +F(|z j |⊙sin(θ δ ),|zi |+|z j |⊙cos(θ δ )) (5)

[0033] In formula (5), θ δ is the hidden feature x i and hidden features x j The phase difference of θ r is the phase aggregation result of two hidden features; F is the two-parameter inverse tangent function;

[0034] Step S5.3: The last two dimensions of are interpreted as node unit dimensions; transpose the feature-related dimensions c 1 ×c 2 and c 3 , which will produce a c 2 ×c 3 nodes, each with a length of c 1 The feature vector of is the feature; the transformed hidden feature node obtained is represented as The similarity between these feature vectors is then calculated, and connections are established between nodes using the k-nearest neighbor method, as shown in the following formula:

[0035]

[0036] In formula (6), Metric represents the calculation of the Euclidean distance between two node feature vectors, which is used to evaluate the relationship between each pair of nodes in the node set V and determine which nodes should establish edges; E represents the edge between neighboring nodes; G represents the graph representation of hidden features; Indicates the total node;

[0037] The node set and its neighbors are obtained as follows:

[0038]

[0039] In formula (7), To represent the subgraph The i-th node of For the node The adjacent k-th neighbor node;

[0040] Step S5.4: Based on step 5.3, feature aggregation is performed and the formula is as follows:

[0041]

[0042] In formula (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network respectively; represents a single node updated after the graph convolutional network, Represents the set of all nodes in the graph; A node set representing a new graph formed by connecting subset nodes in series;

[0043] Step S5.5: Use a fully connected layer composed of one-dimensional convolutions to enhance the richness of the connected node features, and use standard residual connections to obtain the final output of the graph convolutional network and predict the corresponding treatment plan for cervical cancer patients:

[0044]

[0045] In formula (9), W out represents the weights of the fully connected layer used; is the final output of the graph convolutional network.

[0046] Preferably, the five different data sizes are: half the size of the original data set, the same size as the original data set, twice the size of the original data set, four times the size of the original data set, and eight times the size of the original data set; wherein the original data set is the preprocessed cervical cancer data set.

[0047] Preferably, the clinical visit information of each cervical cancer patient includes one of the following or a combination thereof: pathological factors, cancer stage, blood sugar level, white blood cell count, lymphocyte count, granulocyte count, hemoglobin level, platelet count, squamous cell carcinoma antigen, and tumor volume.

[0048] Beneficial effects of the present invention:

[0049] (1) The graph neural network-based method for predicting the segmentation pattern of cervical cancer brachytherapy provided by the present invention generates synthetic data through the GPT2 model, simulates the characteristic distribution of medical data, expands the data set and increases the diversity of the data, solves the limitation of insufficient data on some rare diseases and chronic diseases, and retains the correlation between the original data, so that data that could only be predicted by traditional machine learning models can also be predicted using deep learning models.

[0050] (2) The method for predicting the segmentation pattern of cervical cancer brachytherapy based on graph neural network provided by the present invention captures the patient's visit feature relationship through dynamic feature aggregation graph neural network, and predicts the patient's treatment plan more accurately and effectively. Specifically, the structured data is converted into a graph format to enrich feature learning; a dynamic adjustment method is introduced to improve the generalization ability of input features; a graph neural network is used to model multidimensional hidden features; and feature aggregation is performed to enhance feature richness.

[0051] Additional aspects and advantages of the invention will become apparent from the following description, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic flow chart of a method for predicting segmentation patterns of cervical cancer brachytherapy based on a graph neural network according to an embodiment of the present invention is shown;

[0053] Figure 2 A schematic flowchart of generating synthetic data by a GPT2 model according to an embodiment of the present invention is shown;

[0054] Figure 3 A schematic block diagram showing a graph neural network dynamically capturing patient visit feature relationships according to an embodiment of the present invention;

[0055] Figure 4 A pie-shaped diagram showing different proportions of real data and synthetic data according to an embodiment of the present invention;

[0056] Figure 5a A heat map of correlation matrix between data of one embodiment of the present invention is shown; Figure 5b A heat map of the correlation matrix between data according to another embodiment of the present invention is shown. DETAILED DESCRIPTION

[0057] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0058] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0059] Figure 1 FIG. 1 is a schematic flow chart of a method for predicting segmentation patterns of cervical cancer brachytherapy based on a graph neural network according to an embodiment of the present invention. Figure 1 As shown, the method for predicting segmentation patterns of cervical cancer brachytherapy based on graph neural network includes:

[0060] Step S1: Obtain cervical cancer dataset;

[0061] Step S2: preprocessing the cervical cancer dataset;

[0062] Step S3: input the preprocessed cervical cancer dataset into the GPT2 model for data generation to generate synthetic data;

[0063] Step S4: Evaluate the quality of the synthesized data and select the synthesized data with the highest similarity to the preprocessed cervical cancer dataset;

[0064] Step S5: Select the synthetic data with the highest similarity to the preprocessed cervical cancer dataset as the basic data, use the dynamic feature aggregation graph neural network to capture the medical feature relationship of cervical cancer patients, and predict the corresponding treatment plan.

[0065] In this embodiment, the cervical cancer dataset is an electronic medical record containing clinical consultation information of each cervical cancer patient.

[0066] In this embodiment, the graph neural network-based cervical cancer brachytherapy segmentation pattern prediction method provided by the present invention generates synthetic data through the GPT2 model, simulates the characteristic distribution of medical data, expands the data set and increases the diversity of the data, solves the limitation of insufficient data for some rare diseases and chronic diseases, and retains the correlation between the original data, so that data that could only be predicted by traditional machine learning models can also be predicted by deep learning models.

[0067] In this embodiment, the graph neural network-based cervical cancer brachytherapy segmentation pattern prediction method provided by the present invention captures the patient's medical feature relationship through a dynamic feature aggregation graph neural network, and predicts the patient's treatment plan more accurately and effectively.

[0068] In one embodiment of the present invention, the step S2 specifically includes: step S2.1: using the KNN interpolation method to process the missing values ​​in the cervical cancer dataset to obtain an estimated value corresponding to the missing value of each sample in the cervical cancer dataset, so as to fill the missing value with the estimated value; step S2.2: using the SMOTE balancing algorithm and the ADASYN balancing algorithm respectively to balance the cervical cancer dataset after KNN interpolation processing.

[0069] In one embodiment of the present invention, the step S3 specifically includes: step S3.1: using the GPT2 model to create a vocabulary for the input data, mapping the unique tag in each column of the input data to a unique tag ID, and then applying the mapping in the vocabulary to each row in the input data to generate a series of tag IDs; wherein the input data is the preprocessed cervical cancer dataset, and the preprocessed cervical cancer dataset is a table data T with m rows and n columns;

[0070] Step S3.2: Use the GPT2 model to simulate the column value conditional distribution P(x ij |x i1 , ..., x i(j-1) ), where x ijis the value of the i-th row and j-th column of the data set;

[0071] Step S3.3: Generate synthetic data by using the target masking and overfitting monitoring strategies; wherein the synthetic data consists of a series of observation values ​​O i Composition; the observed value O i The expression is:

[0072] O i =[x i1 , x i2 , ..., x ij , ..., x in ] (1)

[0073] In formula (1), i represents the i-th row; j represents the j-th column;

[0074] Step S3.4: Based on the preprocessed cervical cancer dataset, five different data scales are generated to determine the optimal size of the synthetic data.

[0075] In one embodiment of the present invention, the step S4 specifically includes: step S4.1: using the three measurement indicators of PCD, KLD and MMD to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset; step S4.2: using a triangular radar chart drawn on a logarithmic scale to represent the average values ​​of the three measurement indicators PCD, KLD and MMD; wherein, in the triangular radar chart, a triangle with a smaller area indicates that the similarity between the synthetic data and the preprocessed cervical cancer dataset is higher, that is, the quality of the synthetic data is higher; step S4.3: comparing the variability of the SMOTE balancing algorithm and the ADASYN balancing algorithm through several iterations, and displaying the results using a box plot.

[0076] In one embodiment of the present invention, the step S5 specifically includes: step S5.1: expanding the feature dimension of the low-dimensional input X, and then using a learnable matrix for nonlinear mapping to enhance the expressiveness of the feature; as shown in the following formula:

[0077] X out1 =σ(W 1 (Unsqueeze(X))) (2)

[0078] In formula (2), the low-dimensional input X is the synthetic data preprocessed in step S4; Unsqueeze means expanding the dimension of X; W1 means the feature-rich weight; σ means the activation function; the original input data is converted through the above process to promote the transition from the original space to the feature-rich space; X out1 After step S 5 Output data after .1;

[0079] Step S5.2: Introduce a dynamic adjustment method to improve the feature generalization ability of the input features to flexibly convert the input into a graph structure:

[0080] X out2 =W 2 (DFL(Reshape(X out1 ))) (3)

[0081] In formula (3), Reshape means using DFL(·) based on two-dimensional convolution to perform dimension transformation operation, DFL(·) means using dynamically adjusted feature aggregation, W 2 represents the weight of the 1×1 convolution used for feature enrichment; X out2 is the output data after step S5.2;

[0082] For DFL, assume that two hidden feature labels x i and x j , whose amplitude is z and phase is θ, their superposition can be calculated as follows:

[0083]

[0084] In formula (4), θ i is the hidden feature x i The phase of θ j is the hidden feature x j Phase; ⊙ is element-wise multiplication; |z i | is the hidden feature x i The amplitude of |z j | is the hidden feature x j The amplitude of |z mix | is the amplitude aggregation result of two hidden features;

[0085] Aggregate phase term θ r The calculation is as follows:

[0086] θ δ =θj-θ i

[0087] θ r =θ i +F(|z j |⊙sin(θ δ ),|z i |+|z j |⊙cos(θ δ )) (5)

[0088] In formula (5), θ δ is the hidden feature x i and hidden features xj The phase difference of θ r is the phase aggregation result of two hidden features; F is the two-parameter inverse tangent function;

[0089] Step S5.3: The last two dimensions of are interpreted as node unit dimensions; transpose the feature-related dimensions c 1 ×c 2 and c 3 , which will produce a c 2 ×c 3 nodes, each with a length of c 1 The feature vector of is the feature; the transformed hidden feature node obtained is represented as The similarity between these feature vectors is then calculated, and connections are established between nodes using the k-nearest neighbor method, as shown in the following formula:

[0090]

[0091] In formula (6), Metric represents the calculation of the Euclidean distance between two node feature vectors, which is used to evaluate the node set The relationship between each pair of nodes in determines which nodes should establish edges; E represents the edges between neighboring nodes; G represents the graph representation of hidden features; Indicates the total node;

[0092] The node set and its neighbors are obtained as follows:

[0093]

[0094] In formula (7), To represent the subgraph The i-th node of For the node The adjacent k-th neighbor node;

[0095] Step S5.4: Based on step 5.3, feature aggregation is performed and the formula is as follows:

[0096]

[0097] In formula (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network respectively; represents a single node updated after the graph convolutional network, Represents the set of all nodes in the graph; A node set representing a new graph formed by connecting subset nodes in series;

[0098] Step S5.5: Use a fully connected layer composed of one-dimensional convolutions to enhance the richness of the connected node features, and use standard residual connections to obtain the final output of the graph convolutional network and predict the corresponding treatment plan for cervical cancer patients:

[0099]

[0100] In formula (9), W out represents the weights of the fully connected layer used; is the final output of the graph convolutional network.

[0101] In one embodiment of the present invention, the five different data sizes are: half the size of the original data set, the same size as the original data set, twice the size of the original data set, four times the size of the original data set, and eight times the size of the original data set; wherein the original data set is the preprocessed cervical cancer data set.

[0102] In one embodiment of the present invention, the clinical consultation information of each cervical cancer patient includes one of the following or a combination thereof: pathological factors, cancer stage, blood sugar level, white blood cell count, lymphocyte count, granulocyte count, hemoglobin level, platelet count, squamous cell carcinoma antigen, and tumor volume.

[0103] The technical solution of the present invention will be demonstrated with a specific embodiment below.

[0104] Combination Figure 2 , Figure 3 , Figure 4 , Figure 5a ,and Figure 5b , showing that the method for predicting the segmentation pattern of cervical cancer brachytherapy based on graph neural network provided by the present invention can accurately and effectively predict the treatment plan of the patient. The specific implementation steps of the prediction method are as follows:

[0105] Step 1: Get the cervical cancer dataset.

[0106] The cervical cancer dataset is an electronic medical record containing clinical visit information of each cervical cancer patient. The clinical visit information of each cervical cancer patient includes one or a combination of the following: pathological factors, cancer stage, blood glucose level, white blood cell count, lymphocyte count, granulocyte count, hemoglobin level, platelet count, squamous cell carcinoma antigen, and tumor volume.

[0107] Step 2: Preprocess the cervical cancer dataset.

[0108] The K-Nearest Neighbors (KNN) interpolation method is used to process the missing values ​​in the cervical cancer dataset. To address the data imbalance problem in the cervical cancer dataset, the Synthetic Minority Over-sampling Technique (SMOTE) and the Adaptive Synthetic Sampling Approach for Imbalanced Learning (ADASYN) algorithm are used to balance the data.

[0109] In short, step 2 specifically includes: using the KNN interpolation method to process the missing values ​​in the cervical cancer dataset to obtain an estimated value corresponding to the missing value of each sample in the cervical cancer dataset, so as to fill the missing value with the estimated value; using the SMOTE balancing algorithm and the ADASYN balancing algorithm to balance the cervical cancer dataset after KNN interpolation processing.

[0110] Specifically, the Euclidean distance is used to measure the similarity between samples, and the optimal k value is determined through cross-validation. This value means that for each sample with missing values, the k most similar samples in the data set will be found.

[0111] Furthermore, for each sample with missing values, the k nearest neighbor samples are found according to the selected similarity measure. For continuous features, the missing value can be estimated by the average or weighted average of these nearest neighbor samples; for categorical features, the most common category is selected as the estimated value.

[0112] Furthermore, this method is applied to the entire dataset to estimate all missing values ​​one by one.

[0113] Furthermore, after using the KNN interpolation method to process the missing values ​​in the cervical cancer dataset, the SMOTE balancing algorithm and the ADASY balancing algorithm were used to balance the data.

[0114] Step 3: Input the preprocessed cervical cancer dataset into the GPT2 model for data generation to generate synthetic data.

[0115] Step 3 specifically includes:

[0116] Step 3.1: Use the GPT2 model to create a vocabulary for the input data, map the unique tag in each column of the input data to a unique tag ID, and then apply the mapping in the vocabulary to each row in the input data to generate a series of tag IDs; wherein the input data is a preprocessed cervical cancer dataset, and the preprocessed cervical cancer dataset is a table data T with m rows and n columns;

[0117] Here, the input tabular data is converted into a format suitable for processing by the GPT-2 model, so that the model can be used to train autoregressive tasks.

[0118] Step 3.2: If Figure 2 As shown, the GPT2 model is used to simulate the conditional distribution P(x ij |x i1 , ..., x i(j-1) ), where x ij is the value of the i-th row and j-th column of the data set;

[0119] Step 3.3: Generate synthetic data using the target masking and overfitting monitoring strategies; the synthetic data consists of a series of observations O i Composition; Observed value O i The expression is:

[0120] O i =[x i1 , x i2 , ..., x ij , ..., x in ] (1)

[0121] In formula (1), i represents the i-th row; j represents the j-th column;

[0122] This process treats a single observation as a (row) i is a sequence, a single observation in a tabular data set T with n columns represented by O i =[x i1 , x i2 , ..., x ij , ..., x in ] definition, each line of observation is similar to a sentence in the text, thereby learning the conditional distribution and sequentially generating the next value in the sequence, and finally generating a complete observation.

[0123] The target masking and overfitting monitoring strategies are adopted to reduce the risk of data duplication and increase the rate of generating valid observations. Target masking is achieved by randomly replacing special masking tags in the target or label tags of the training data, aiming to minimize the possibility of training data records being "remembered" and copied by the model, and the distance to the closest record (DCR) statistical method is used to detect data overfitting.

[0124] Step 3.4: Based on the preprocessed cervical cancer dataset, five different data scales were generated to determine the optimal size of the synthetic data. Figure 4 As shown, the five different data sizes are: half the size of the original dataset, the same size as the original dataset, twice the size of the original dataset, four times the size of the original dataset, and eight times the size of the original dataset; wherein the original dataset is the preprocessed cervical cancer dataset.

[0125] Step 4: Evaluate the quality of the synthetic data generated in step 3 and select the synthetic data with the highest similarity to the preprocessed cervical cancer dataset.

[0126] Step 4 specifically includes:

[0127] Step 4.1: Use three metrics, namely PCD, KLD and MMD, to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset;

[0128] The similarity between the synthetic data and the original data is measured using three metrics: pairwise correlation difference (PCD), Kullback-Leibler Divergence (KLD), and maximum mean discrepancy (MMD). The lower the value, the higher the similarity.

[0129] Step 4.2: Use a triangular radar chart drawn on a logarithmic scale to represent the average values ​​of the three measurement indicators, namely, PCD, KLD and MMD; wherein, in the triangular radar chart, a triangle with a smaller area indicates that the synthetic data has a higher similarity with the preprocessed cervical cancer dataset, that is, the synthetic data has a higher quality;

[0130] This triangular radar chart visually describes the similarity between the synthetic dataset and the original dataset through the size of the triangular area representing the performance of different algorithms.

[0131] Figure 5a The raw data are shown. Figure 5b Synthetic data are shown. Figure 5a and Figure 5b It can be seen that the synthetic data still retains the data correlation of the original data.

[0132] Step S4.3: Compare the variability of the SMOTE balancing algorithm and the ADASYN balancing algorithm through 100 iterations and display the results using a box plot.

[0133] Specifically, the variability of two different balancing algorithms was compared through 100 iterations, and the results were intuitively displayed using box plots. The "tighter" box plot, characterized by a lower interquartile range (IQR) or shorter box, indicates that the results of the algorithm change the least over multiple iterations, reflecting lower variability and higher robustness. The effectiveness of different data balancing and synthesis techniques in preserving the characteristics of the original data is effectively demonstrated and evaluated.

[0134] Step 5: Select the synthetic data with the highest similarity to the preprocessed cervical cancer dataset as the basic data, such as Figure 3 As shown in the figure, a dynamic feature aggregation graph neural network is used to capture the medical feature relationship of cervical cancer patients and predict the corresponding treatment plan.

[0135] Step 5 specifically includes:

[0136] Step S5.1: For a two-dimensional structured data table of size m×n, the traditional method is to regard each eigenvalue as a point and construct a graph by calculating the similarity between them. Such a graph structure cannot well represent the complex relationship in the original data. To overcome the above problems, we use a graph structure learning method with hidden features to obtain a more expressive learning method. First, expand the feature dimension of the low-dimensional input X, and then use a learnable matrix for nonlinear mapping to enhance the expressiveness of the feature; as shown in the following formula:

[0137] X out1 =σ(W 1 (Unsqueeze(X))) (2)

[0138] In formula (2), the low-dimensional input X is the synthetic data preprocessed in step 4; Unsqueeze means expanding the dimension of X; W1 represents the feature-rich weight; σ represents the activation function; the original input data is transformed through the above process to promote the transition from the original space to the feature-rich space; X out1 is the output data after step 5.1;

[0139] Step 5.2: In order to obtain dynamic feature aggregation and avoid the lack of generalization caused by fixed weight aggregation, we introduce a dynamic adjustment method to improve the feature generalization ability of the input features to flexibly transform the input into a graph structure:

[0140] X out2 =W 2 (DFL(Reshape(X out1 ))) (3)

[0141] In formula (3), Reshape means using DFL(·) based on two-dimensional convolution to perform dimension transformation operation, DFL(·) means using dynamically adjusted feature aggregation, W 2 represents the weight of the 1×1 convolution used for feature enrichment; X out2 is the output data after step 5.2;

[0142] In the field of signal processing, amplitude and phase are the basic properties that describe waveform features. This method is applied to feature aggregation of medical table data. For example, when analyzing blood sugar levels, amplitude represents the magnitude of blood sugar fluctuations, while phase reflects the temporal pattern of these fluctuations. For DFL, the hidden features learned after expanding the dimension can be expressed in vector form. Assume that two hidden feature labels x i and x j , whose amplitude is z and phase is θ, their superposition can be calculated as follows:

[0143]

[0144] In formula (4), θ i is the hidden feature x i The phase of θ j is the hidden feature x j Phase; ⊙ is element-wise multiplication; |z i | is the hidden feature x i The amplitude of |z j | is the hidden feature x j The amplitude of |z mix | is the amplitude aggregation result of two hidden features;

[0145] Aggregate phase term θ r The calculation is as follows:

[0146] θ δ =θ j -θ i

[0147] θ r =θ i +F(|z j |⊙sin(θ δ ),|z i|+|z j |⊙cos(θ δ )) (5)

[0148] In formula (5), θ δ is the hidden feature x i and hidden features x j The phase difference of θ r is the phase aggregation result of two hidden features; F is the two-parameter inverse tangent function;

[0149] Step S5.3: In order to use graph neural networks to model the acquired multi-dimensional hidden features for effective feature learning, The last two dimensions of are interpreted as node unit dimensions; transpose the feature-related dimensions c 1 ×c 2 and c 3 , which will produce a c 2 ×c 3 nodes, each with a length of c 1 The feature vector of is the feature; the transformed hidden feature node obtained is represented as The similarity between these feature vectors is then calculated, and connections are established between nodes using the k-nearest neighbor method, as shown in the following formula:

[0150]

[0151] In formula (6), Metric represents the calculation of the Euclidean distance between two node feature vectors, which is used to evaluate the node set The relationship between each pair of nodes in determines which nodes should establish edges; E represents the edges between neighboring nodes; G represents the graph representation of hidden features; Indicates the total node;

[0152] The node set and its neighbors are obtained as follows:

[0153]

[0154] In formula (7), To represent the subgraph The i-th node of For the node The adjacent k-th neighbor node;

[0155] Step 5.4: Based on step 5.3, feature aggregation is performed and the formula is as follows:

[0156]

[0157] In formula (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network respectively; represents a single node updated after the graph convolutional network, Represents the set of all nodes in the graph; A node set representing a new graph formed by connecting subset nodes in series;

[0158] Step 5.5: Use a fully connected layer consisting of one-dimensional convolutions to enhance the richness of the connected node features, and use standard residual connections to obtain the final output of the graph convolutional network and predict the corresponding treatment plan for cervical cancer patients:

[0159]

[0160] In formula (9), W out represents the weights of the fully connected layer used; is the final output of the graph convolutional network.

[0161] from Figure 3 It can be seen that after the data is preprocessed, the hidden feature amplitude and phase information is extracted, and more representative features are extracted by dynamic aggregation for final output.

[0162] This specific embodiment generates synthetic data through the GPT2 model to simulate the characteristic distribution of medical data to expand the data set and increase the diversity of the data, solving the limitation of insufficient data on some rare diseases and chronic diseases while retaining the correlation between the original data (such as Figure 5a and Figure 5b As shown in the figure, data that could only be predicted by traditional machine learning models can also be predicted using deep learning models.

[0163] Furthermore, through the dynamic feature capture method of graph neural network, the complex relationship between patient characteristics can be captured. This specific embodiment compares the prediction results of seven models on the original cervical cancer data: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), TabNet (Interpretable Table Learning) and SAINT (Neural Network for Improving Tabular Data through Row Attention and Comparative Pre-training). Table 1 shows the prediction results of different models on the original data, and Table 2 shows the prediction results of different models on the data synthesized using SMOTE and ADASYN oversampling algorithms. It is concluded that the prediction accuracy, precision, recall rate and F1 score of the present invention are the highest.

[0164] Table 1 Prediction results on original cervical cancer data

[0165]

[0166] Table 2 Prediction results on synthetic cervical cancer data

[0167]

[0168]

[0169] In summary, the key technical point of the present invention is to generate synthetic data through the GPT2 model to simulate the characteristic distribution of medical data to expand the data set and increase the diversity of the data. Furthermore, indicators such as PCD, KLD and MMD are used to evaluate the similarity between the synthetic data and the original data, and the synthetic data with the highest similarity to the original data is selected. Furthermore, the patient's medical feature relationship is captured through a dynamic feature aggregation graph neural network to more accurately and effectively predict the patient's treatment plan. Specifically, structured data is converted into a graphical format to enrich feature learning; a dynamic adjustment method is introduced to improve the generalization ability of input features; a graph neural network is used to model multidimensional hidden features; feature aggregation is performed to enhance feature richness; and a fully connected layer is used to enhance the richness of connection node features.

[0170] In addition, the protection point of the present invention lies in the construction and training method of dynamic feature aggregation graph neural network, including feature conversion, dynamic adjustment, graph neural network modeling, feature aggregation and the design of fully connected layers.

[0171] The above-mentioned key technical points and protection points constitute the innovation and uniqueness of the present invention in the prediction of the segmentation pattern of cervical cancer brachytherapy, ensuring its high efficiency and accuracy in practical applications.

[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting segmentation patterns in cervical cancer brachytherapy based on graph neural network, characterized in that: include: Step S1: Obtain cervical cancer dataset; The cervical cancer dataset is an electronic medical record containing clinical consultation information of each cervical cancer patient; Step S2: preprocessing the cervical cancer dataset; Step S3: input the preprocessed cervical cancer dataset into the GPT2 model for data generation to generate synthetic data; Step S4: evaluating the quality of the synthesized data, and selecting the synthesized data with the highest similarity to the preprocessed cervical cancer dataset; Step S5: Select the synthetic data with the highest similarity to the preprocessed cervical cancer data set as basic data, use the dynamic feature aggregation graph neural network to capture the medical characteristic relationship of cervical cancer patients, and predict the corresponding treatment plan.

2. The method for predicting segmentation patterns of cervical cancer brachytherapy based on graph neural network according to claim 1, characterized in that: The step S2 specifically includes: Step S2.1: Use the KNN interpolation method to process the missing values ​​in the cervical cancer dataset to obtain an estimated value corresponding to the missing value of each sample in the cervical cancer dataset, so as to fill the missing value with the estimated value; Step S2.2: Use the SMOTE balancing algorithm and the ADASYN balancing algorithm to balance the cervical cancer dataset after KNN interpolation processing.

3. The method for predicting segmentation patterns of cervical cancer brachytherapy based on graph neural network according to claim 2, characterized in that: The step S3 specifically includes: Step S3.1: using the GPT2 model to create a vocabulary for the input data, mapping the unique tag in each column of the input data to a unique tag ID, and then applying the mapping in the vocabulary to each row in the input data to generate a series of tag IDs; wherein the input data is the preprocessed cervical cancer dataset, and the preprocessed cervical cancer dataset is a table data T with m rows and n columns; Step S3.2: Use the GPT2 model to simulate the column value conditional distribution P(x ij |x i1 …,x i(j-1) ) where x ij is the value of the i-th row and j-th column of the data set; Step S3.3: Generate synthetic data by using the target masking and overfitting monitoring strategies; wherein the synthetic data consists of a series of observation values ​​O i Composition; the observed value O i The expression is: O i =[x i1 ,x i2 ,...,x ij ,...,x in ] (1) In formula (1), i represents the i-th row; j represents the j-th column; Step S3.4: Based on the preprocessed cervical cancer dataset, five different data scales are generated to determine the optimal size of the synthetic data.

4. The method for predicting segmentation patterns of cervical cancer brachytherapy based on graph neural network according to claim 3, characterized in that: The step S4 specifically includes: Step S4.1: using three metrics, namely PCD, KLD and MMD, to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset; Step S4.2: using a triangular radar chart drawn on a logarithmic scale to represent the average values ​​of the three measurement indicators, namely, PCD, KLD and MMD; wherein, in the triangular radar chart, a triangle with a smaller area indicates that the synthetic data has a higher similarity with the preprocessed cervical cancer dataset, that is, the synthetic data has a higher quality; Step S4.3: Compare the variability of the SMOTE balancing algorithm and the ADASYN balancing algorithm through several iterations, and display the results using a box plot.

5. The method for predicting segmentation patterns of cervical cancer brachytherapy based on graph neural network according to claim 4, characterized in that: The step S5 specifically includes: Step S5.1: Expand the feature dimension of the low-dimensional input X, and then use a learnable matrix for nonlinear mapping to enhance the expressiveness of the feature; as shown in the following formula: X out1 =σ(W1(Unsqueeze(X))) (2) In formula (2), the low-dimensional input X is the synthetic data preprocessed in step S4; Unsqueeze means expanding the dimension of X; W1 means the feature-rich weight; σ means the activation function; the original input data is converted through the above process to promote the transition from the original space to the feature-rich space; X out1 is the output data after step S5.1; Step S5.2: Introduce a dynamic adjustment method to improve the feature generalization ability of the input features to flexibly convert the input into a graph structure: X out2 =W2(DFL(Reshape(X out1 ))) (3) In formula (3), Reshape indicates the dimension transformation operation using DFL() based on 2D convolution, DFL(·) indicates the feature aggregation using dynamic adjustment, W2 indicates the weight of 1×1 convolution for feature enrichment; X out2 is the output data after step S5.2; For DFL, assume that two hidden feature labels x i and xj, with amplitude z and phase θ, their superposition can be calculated as follows: In formula (4), θ i is the hidden feature x i The phase of θ j is the hidden feature x j Phase; ⊙ is element-wise multiplication; |z i | is the hidden feature x i The amplitude of |z j | is the hidden feature x j The amplitude of |Z mix | is the amplitude aggregation result of two hidden features; Aggregate phase term θ r The calculation is as follows: i δ =θ j -θ i i r =θ i +F(|z j |☉sin(θ δ ),|z i |+|z j |☉)cos(θ δ )) (5) In formula (5), θ δ is the hidden feature x i and hidden features x j The phase difference of θ r is the phase aggregation result of two hidden features; F is the two-parameter inverse tangent function; Step S5.3: The last two dimensions of are interpreted as node unit dimensions; transpose the feature-related dimensions c1×c2 and c3, which will produce c2×c3 nodes, each of which is characterized by a feature vector of length c1; the transformed prepared hidden feature node is represented as The similarity between these feature vectors is then calculated, and connections are established between nodes using the k-nearest neighbor method, as shown in the following formula: In formula (6), Metric represents the calculation of the Euclidean distance between two node feature vectors, which is used to evaluate the node set The relationship between each pair of nodes in determines which nodes should establish edges; E represents the edges between neighboring nodes; G represents the graph representation of hidden features; Indicates the total node; The node set and its neighbors are obtained as follows: In formula (7), To represent the subgraph The i-th node of For the node The adjacent k-th neighbor node; Step S5.4: Based on step 5.3, feature aggregation is performed and the formula is as follows: In formula (8), Aggrgate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network respectively; represents a single node updated after the graph convolutional network, Represents the set of all nodes in the graph; A node set representing a new graph formed by connecting subset nodes in series; Step S5.5: Use a fully connected layer composed of one-dimensional convolutions to enhance the richness of the connected node features, and use standard residual connections to obtain the final output of the graph convolutional network and predict the corresponding treatment plan for cervical cancer patients: In formula (9), W out represents the weights of the fully connected layer used; is the final output of the graph convolutional network.

6. The method for predicting segmentation patterns of cervical cancer brachytherapy based on graph neural network according to claim 3, characterized in that: The five different data sizes are: half the size of the original data set, the same size as the original data set, twice the size of the original data set, four times the size of the original data set, and eight times the size of the original data set; wherein the original data set is the preprocessed cervical cancer data set.

7. The method for predicting segmentation patterns for cervical cancer brachytherapy based on graph neural network according to any one of claims 1 to 6, characterized in that: The clinical visit information of each cervical cancer patient includes one of the following or a combination thereof: pathological factors, cancer stage, blood sugar level, white blood cell count, lymphocyte count, granulocyte count, hemoglobin level, platelet count, squamous cell carcinoma antigen, and tumor volume.

Images