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

Abstract

The application provides a cervical cancer brachytherapy segmentation mode prediction method based on a graph neural network. The method comprises the following steps: obtaining a cervical cancer data set; the cervical cancer data set is an electronic medical record containing the 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 to generate synthetic data; evaluating the quality of the synthetic data and selecting the synthetic data with the highest similarity to the preprocessed cervical cancer data set; selecting the synthetic data with the highest similarity to the preprocessed cervical cancer data set as the basis data, using a dynamic feature aggregation graph neural network to capture the treatment feature relationship of the cervical cancer patient, and predicting the corresponding treatment scheme. The application uses synthetic data generated based on the GPT2 model to solve the problem of insufficient medical data, and uses the method of dynamically capturing the relationship between patient features by the graph neural network to more accurately and effectively predict the treatment scheme of the patient.

Description

Technical Field

[0001] This invention relates to the technical field of segmentation pattern prediction for cervical cancer treatment, and more specifically, to a method for predicting segmentation patterns for brachytherapy of cervical cancer based on graph neural networks. Background Technology

[0002] Cervical cancer is one of the four most common cancers among women worldwide, and its early stages can be treated surgically. As the disease progresses, radiation therapy becomes the primary treatment. The combination of external beam radiation therapy (EBRT) and intracavitary brachytherapy is now considered the most effective treatment for most cervical cancer patients. Brachytherapy for cervical cancer involves placing a radiation source directly into or near the tumor area. This method plays a crucial role in the treatment of cervical cancer, especially in locally advanced or recurrent cases. Therefore, studying the relationship between patient clinical characteristics and radiation therapy protocols is essential for developing personalized brachytherapy modalities.

[0003] However, challenges in cervical cancer treatment include the scarcity of chronic disease data, the lengthy treatment process, and the significant time required for experienced planners to accumulate data. Training models based solely on individual planners may inadvertently incorporate their subjective biases, making it extremely difficult to obtain a comprehensive set of high-quality radiotherapy plans in a clinical setting. Developing dose allocation schemes for brachytherapy of cervical cancer requires consideration of multiple factors, such as treatment extent, duration, and dose selection. While existing evidence-based clinical guidelines, such as the NCCN guidelines, provide basic treatment benchmarks, the diversity and complexity of patient conditions necessitate individualized adjustments, which presents a significant challenge in clinical practice.

[0004] In recent years, graph neural networks (GNNs) have seen significant advancements in healthcare, particularly in predicting treatment plans based on clinical indicators. These progress stems from GNNs' ability to interpret complex relationships within 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, assigning gene expression data to nodes. Therefore, Graph-CNNs can identify gene expression patterns specific to particular health conditions. For example, layer-level correlation propagation (GLRP) has been adapted for Graph-CNNs, enabling the generation of interpretable, patient-specific molecular subnetworks. These subnetworks are crucial for identifying drivers of tumor progression and helping to tailor precision medicine strategies. On a broad breast cancer dataset, this approach played a key role in predicting distant metastasis and identifying patient-specific treatment vulnerabilities. In bioinformatics, GNNs are widely used in node classification, link prediction, and graph generation, demonstrating exceptional performance, especially in biological network data analysis. Their applications span multiple fields, including disease prediction, drug discovery, and biomedical imaging. The inherent advantages of GNNs, such as their 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 described in the related technology (patent number: CN202110108182.4) integrates Bayesian networks and regression analysis techniques, achieving improved accuracy and robustness in treatment cost prediction even with limited data. However, it has limitations in capturing complex relationships between features. Another related technology (patent number: CN201811407145.8) describes a treatment plan generation method that utilizes a feedforward deep neural network with temporal connections to generate the patient's current treatment plan and predict future treatment plans using a sequence-to-sequence model. However, this method requires a large amount of training data. The above technologies suffer from the following technical problems:

[0006] (1) The nonlinear relationship between multiple factors in patients remains a challenge in clinical practice. When the relationship between data is very complex and difficult to represent 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) Training deep learning models requires a lot of data, but for some rare and chronic diseases, the data is scarce and this technology cannot be used.

[0009] Therefore, we urgently need a novel method for predicting the fractionation pattern of brachytherapy for cervical cancer to solve the technical problems that the above-mentioned techniques have not yet solved. Summary of the Invention

[0010] This invention aims to provide a method that uses synthetic data generated based on the generative pre-trained model GPT2 to solve the problem of insufficient medical data, and employs graph neural networks to dynamically capture the relationships between patient features, thereby more accurately and effectively predicting the patient's treatment plan.

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

[0012] To achieve the above objectives, the present invention provides a method for predicting segmentation patterns in brachytherapy for cervical cancer based on graph neural networks. The method includes: Step S1: acquiring a cervical cancer dataset; the cervical cancer dataset is an electronic medical record containing clinical visit information for each cervical cancer patient; Step S2: preprocessing the cervical cancer dataset; Step S3: inputting the preprocessed cervical cancer dataset into a GPT2 model 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: using the synthetic data with the highest similarity to the preprocessed cervical cancer dataset as the base data, utilizing a dynamic feature aggregation graph neural network to capture the visit feature relationships of cervical cancer patients and predict the corresponding treatment plan.

[0013] Preferably, 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 to balance the cervical cancer dataset after KNN interpolation.

[0014] Preferably, step S3 specifically includes: step S3.1: creating a vocabulary for the input data using the GPT2 model, mapping the unique tokens in each column of the input data to a unique token ID, and then applying the mapping in the vocabulary to each row of the input data to generate a series of token 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 conditional distribution P(x) of the column values ​​in each row of the input data.ij |x i1 , ..., x i(j-1) ); where x ij Let be the value in the i-th row and j-th column of the dataset;

[0016] Step S3.3: Generate synthetic data using a strategy of target masking and overfitting detection; 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 equation (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 sizes were generated to determine the optimal size of the synthesized data.

[0020] Preferably, step S4 specifically includes: Step S4.1: Using PCD, KLD, and MMD as three metrics to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset; Step S4.2: Using a triangular radar chart plotted on a logarithmic scale to represent the average value of the three metrics, PCD, KLD, and MMD; wherein, in the triangular radar chart, the smaller triangle indicates a 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, 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 expressive power of the features; as shown in the following formula:

[0022] X out1 =σ(W1(Unsqueeze(X))) (2)

[0023] In equation (2), the low-dimensional input X is the preprocessed synthetic data in step S4; Unsqueeze represents expanding the dimension of X; W1 represents the feature-rich weights; σ represents the activation function; the above process transforms the original input data, facilitating the transition from the original space to the feature-rich space; Xout1 This 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, so as to flexibly transform the input into a graph structure:

[0025] X out2 =W2(DFL(Reshape(X) out1 ))) (3)

[0026] In equation (3), Reshape represents the dimensionality transformation operation using DFL(·) based on two-dimensional convolution, DFL(·) represents the use of dynamically adjusted feature aggregation, and W2 represents the weights used for feature-rich 1×1 convolutions; X out2 This is the output data after step S5.2;

[0027] For DFL, suppose there are two hidden feature labels x i and x j Their amplitude is z and their phase is θ. Their superposition can be calculated as follows:

[0028]

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

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

[0031] θ δ =θ j -θ i

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

[0033] In equation (5), θ δ To hide feature xi and hidden features x j The phase difference; θ r This is the phase aggregation result of two hidden features; F is the two-parameter arctangent function.

[0034] Step S5.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The last two dimensions are interpreted as node unit dimensions; transposing the feature-related dimensions c1×c2 and c3 will produce c2×c3 nodes, each with a feature vector of length c1 as its feature; the resulting transformed nodes for hiding features are represented as... The similarity between these feature vectors is then calculated, and connections between nodes are established using the k-nearest neighbor method, as shown in the following formula:

[0035]

[0036] In equation (6), Metric represents the calculation of the Euclidean distance between the feature vectors of two nodes, which is used to evaluate the node set. The graph represents the relationship between each pair of nodes and determines which nodes should have edges established between them; E represents the edges between neighboring nodes; G represents the graph representation of hidden features. Indicates the global node;

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

[0038]

[0039] In equation (7), To represent the i-th subgraph The i-th node; For nodes The kth adjacent neighbor node;

[0040] Step S5.4: Based on step 5.3, perform feature aggregation to obtain the following formula:

[0041]

[0042] In Equation (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network, respectively; This represents a single node updated after the graph convolutional network. This represents the set of all nodes in the graph; The node set represents the new graph formed by concatenating the nodes of a subset;

[0043] Step S5.5: Use fully connected layers composed of one-dimensional convolutions to enhance the richness of the connection node features, and utilize 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 equation (9), W out Indicates the weights of the fully connected layer used; This is the final output of the graph convolutional network.

[0046] Preferably, 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.

[0047] Preferably, the clinical 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.

[0048] The beneficial effects of this invention are:

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

[0050] (2) The method for predicting the segmentation pattern of cervical cancer brachytherapy based on graph neural networks provided by this invention captures the relationship of patient visit features through dynamic feature aggregation graph neural networks, thereby more accurately and effectively predicting 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; and feature aggregation is performed to enhance feature richness.

[0051] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0052] Figure 1A schematic flowchart of a segmentation pattern prediction method for cervical cancer brachytherapy based on graph neural networks is shown as an embodiment of the present invention.

[0053] Figure 2 A schematic flowchart illustrating the generation of synthetic data using the GPT2 model according to an embodiment of the present invention is shown.

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

[0055] Figure 4 A pie chart illustrating different proportions of real and synthetic data according to an embodiment of the present invention is shown.

[0056] Figure 5a A heatmap of the correlation matrix between data is shown in one embodiment of the present invention;

[0057] Figure 5b A heatmap of the correlation matrix between data is shown, representing another embodiment of the present invention. Detailed Implementation

[0058] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0059] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0060] Figure 1 A schematic flowchart illustrating a segmentation pattern prediction method for cervical cancer brachytherapy based on a graph neural network, according to an embodiment of the present invention, is shown. Figure 1 As shown, this method for predicting segmentation patterns in brachytherapy for cervical cancer based on graph neural networks includes:

[0061] Step S1: Obtain the cervical cancer dataset;

[0062] Step S2: Preprocess the cervical cancer dataset;

[0063] Step S3: Input the preprocessed cervical cancer dataset into the GPT2 model to generate synthetic data;

[0064] Step S4: Evaluate the quality of the synthetic data and select the synthetic data with the highest similarity to the preprocessed cervical cancer dataset;

[0065] Step S5: Select the synthetic data with the highest similarity to the preprocessed cervical cancer dataset as the base data, use the dynamic feature aggregation graph neural network to capture the relationship of cervical cancer patients' medical visit characteristics, and predict the corresponding treatment plan.

[0066] In this embodiment, the cervical cancer dataset is an electronic medical record containing clinical visit information for each cervical cancer patient.

[0067] In this embodiment, the cervical cancer brachytherapy segmentation pattern prediction method based on graph neural networks provided by the present invention generates synthetic data through the GPT2 model to simulate the feature distribution of medical data, thereby expanding the dataset and increasing the diversity of data. This solves the limitation problem of insufficient data for some rare and chronic diseases, while preserving 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.

[0068] In this embodiment, the cervical cancer brachytherapy segmentation pattern prediction method based on graph neural networks provided by the present invention captures the patient's visit feature relationships through dynamic feature aggregation graph neural networks, thereby more accurately and effectively predicting the patient's treatment plan.

[0069] In one embodiment of the present invention, 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 to balance the cervical cancer dataset after KNN interpolation.

[0070] In one embodiment of the present invention, step S3 specifically includes: step S3.1: creating a vocabulary for the input data using the GPT2 model, mapping the unique markers in each column of the input data to a unique marker ID, and then applying the mapping in the vocabulary to each row of the input data to generate a series of marker 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;

[0071] Step S3.2: Use the GPT2 model to simulate the conditional distribution P(x) of the column values ​​in each row of the input data. ij |x i1 , ..., x i(j-1) ); where x ij Let be the value in the i-th row and j-th column of the dataset;

[0072] Step S3.3: Generate synthetic data using a strategy of target masking and overfitting detection; wherein the synthetic data consists of a series of observation values ​​O i Composition; the observed value O i The expression is:

[0073] O i =[x i1 x i2 , ..., x ij , ..., x in (1)

[0074] In equation (1), i represents the i-th row; j represents the j-th column;

[0075] Step S3.4: Based on the preprocessed cervical cancer dataset, five different data sizes were generated to determine the optimal size of the synthesized data.

[0076] In one embodiment of the present invention, step S4 specifically includes: Step S4.1: Using PCD, KLD, and MMD as three metrics to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset; Step S4.2: Using a triangular radar chart plotted on a logarithmic scale to represent the average value of the three metrics, PCD, KLD, and MMD; wherein, in the triangular radar chart, the smaller triangle indicates a 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.

[0077] In one embodiment of the present invention, 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 expressive power of the features; as shown in the following formula:

[0078] X out1 =σ(W1(Unsqueeze(X))) (2)

[0079] In equation (2), the low-dimensional input X is the preprocessed synthetic data in step S4; Unsqueeze represents expanding the dimension of X; W1 represents the feature-rich weights; σ represents the activation function; the above process transforms the original input data, facilitating the transition from the original space to the feature-rich space; X out1 This is the output data after step S5.1;

[0080] Step S5.2: Introduce a dynamic adjustment method to improve the feature generalization ability of the input features, so as to flexibly transform the input into a graph structure:

[0081] X out2 =W2(DFL(Reshape(X) out1 ))) (3)

[0082] In equation (3), Reshape represents the dimensionality transformation operation using DFL(·) based on two-dimensional convolution, DFL(·) represents the use of dynamically adjusted feature aggregation, and W2 represents the weights used for feature-rich 1×1 convolutions; X out2 This is the output data after step S5.2;

[0083] For DFL, suppose there are two hidden feature labels x i and x j Their amplitude is z and their phase is θ. Their superposition can be calculated as follows:

[0084]

[0085] In equation (4), θ i To hide feature x i phase; θ j To hide feature x j Phase; ⊙ represents element-wise multiplication; |z i | For hidden features x i The amplitude; |z j | For hidden features x i The amplitude; |z mix | represents the amplitude aggregation result of the two hidden features;

[0086] Aggregate phase term θ r The calculation is as follows:

[0087] θ δ =θ j -θ i

[0088] θ r =θ i +F(|z j |⊙sin(θ δ ), |z i |+|z j |⊙cos(θ δ (5)

[0089] In equation (5), θ δ To hide feature x i and hidden features x j The phase difference; θ rThis is the phase aggregation result of two hidden features; F is the two-parameter arctangent function.

[0090] Step S5.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The last two dimensions are interpreted as node unit dimensions; transposing the feature-related dimensions c1×c2 and c3 will produce c2×c3 nodes, each with a feature vector of length c1 as its feature; the resulting transformed nodes for hiding features are represented as... The similarity between these feature vectors is then calculated, and connections between nodes are established using the k-nearest neighbor method, as shown in the following formula:

[0091]

[0092] In equation (6), Metri c This represents calculating the Euclidean distance between the feature vectors of two nodes, used to evaluate the node set. The graph represents the relationship between each pair of nodes and determines which nodes should have edges established between them; E represents the edges between neighboring nodes; G represents the graph representation of hidden features. Indicates the global node;

[0093] The node set and its neighbors are obtained as follows:

[0094]

[0095] In equation (7), To represent the i-th subgraph The i-th node; For nodes The kth adjacent neighbor node;

[0096] Step S5.4: Based on step 5.3, perform feature aggregation to obtain the following formula:

[0097]

[0098] In Equation (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network, respectively; This represents a single node updated after the graph convolutional network. This represents the set of all nodes in the graph; The node set represents the new graph formed by concatenating the nodes of a subset;

[0099] Step S5.5: Use fully connected layers composed of one-dimensional convolutions to enhance the richness of the connection node features, and utilize standard residual connections to obtain the final output of the graph convolutional network, and predict the corresponding treatment plan for cervical cancer patients:

[0100]

[0101] In equation (9), W out Indicates the weights of the fully connected layer used; This is the final output of the graph convolutional network.

[0102] In one embodiment of the present invention, 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.

[0103] In one embodiment of the present invention, the clinical medical 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.

[0104] The technical solution of the present invention will be illustrated below with a specific embodiment.

[0105] Combination Figure 2 , Figure 3 , Figure 4 , Figure 5a ,and Figure 5b This invention demonstrates how a graph neural network-based method for predicting segmentation patterns in brachytherapy for cervical cancer can accurately and effectively predict a patient's treatment plan. The specific implementation steps of this prediction method are as follows:

[0106] Step 1: Obtain the cervical cancer dataset.

[0107] The cervical cancer dataset is an electronic medical record containing clinical visit information for each cervical cancer patient. Each cervical cancer patient's clinical visit information 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.

[0108] Step 2: Preprocess the cervical cancer dataset.

[0109] The missing values ​​in the cervical cancer dataset were handled using the K-Nearest Neighbors (KNN) interpolation method. 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 were used to balance the data.

[0110] 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; and using the SMOTE balancing algorithm and the ADASYN balancing algorithm to balance the cervical cancer dataset after KNN interpolation.

[0111] Specifically, Euclidean distance is used to measure the similarity between samples, and cross-validation is used to determine the optimal value of k, which represents the k most similar samples in the dataset for each sample with missing values.

[0112] Furthermore, for each sample containing missing values, the k nearest neighbor samples are found based on the selected similarity measure. For continuous features, the missing values ​​can be estimated using the average or weighted average of these nearest neighbor samples; for categorical features, the most common category is selected as the estimate.

[0113] Furthermore, this method is applied to the entire dataset to estimate all missing values ​​one by one.

[0114] Furthermore, after using the KNN interpolation method to handle missing values ​​in the cervical cancer dataset, the SMOTE balancing algorithm and the ADASY balancing algorithm were used to balance the data.

[0115] Step 3: Input the preprocessed cervical cancer dataset into the GPT2 model to generate synthetic data.

[0116] Step 3 specifically includes:

[0117] Step 3.1: Create a vocabulary for the input data using the GPT2 model, map the unique token in each column of the input data to a unique token ID, and then apply the mapping in the vocabulary to each row of the input data to generate a series of token 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;

[0118] Here, the input tabular data is converted into a format suitable for the GPT-2 model to process, so that the model can be used to train autoregressive tasks.

[0119] Step 3.2: As Figure 2 As shown, the GPT2 model is used to simulate the conditional distribution P(x) of the column values ​​in each row of the input data. ij |x i1 , ..., x i(j-1) ); where x ij Let be the value in the i-th row and j-th column of the dataset;

[0120] Step 3.3: Generate synthetic data using a target masking and overfitting detection strategy; wherein, the synthetic data consists of a series of observations O i Composition; Observation value O i The expression is:

[0121] O i =[x i1 , xi 2, ..., x ij , ..., x in (1)

[0122] In equation (1), i represents the i-th row; j represents the j-th column;

[0123] This process treats a single observation as a (row) O i It is a sequence, where a single observation in a tabular data T with n columns is determined by O. i =[x i1 x i2 , ..., x ij , ..., x in The definition is that each line of observations is similar to a sentence in text, thereby learning the conditional distribution and sequentially generating the next value in the sequence, eventually generating a complete set of observations.

[0124] A strategy of target masking and overfitting detection is employed to reduce the risk of data duplication and improve the rate of generating valid observations. Target masking is achieved by randomly replacing special masking labels in the target or label markers of the training data, aiming to minimize the possibility that the training data records are "remembered" and copied by the model. The statistical method of Distance to Closest Record (DCR) is used to detect data overfitting.

[0125] Step 3.4: Based on the preprocessed cervical cancer dataset, five different data sizes were generated to determine the optimal size for the synthesized data. For example... 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; where the original dataset is the preprocessed cervical cancer dataset.

[0126] 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.

[0127] Step 4 specifically includes:

[0128] Step 4.1: Use PCD, KLD, and MMD as three metrics to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset;

[0129] The similarity between synthetic and original data is measured using three metrics: Pairwise Cor-relation Difference (PCD), Kullback-Leibler Divergence (KLD), and Maximum Mean Discrepancy (MMD). The lower the value, the higher the similarity.

[0130] Step 4.2: Use a triangle radar chart plotted on a logarithmic scale to represent the average values ​​of the three metrics, PCD, KLD, and MMD; wherein, in the triangle radar chart, the smaller triangle indicates that the synthetic data has a higher similarity to the preprocessed cervical cancer dataset, that is, the synthetic data has a higher quality.

[0131] This triangular radar chart visually describes the similarity between the synthetic and original datasets by the size of the triangular regions representing the performance of different algorithms.

[0132] Figure 5a The image shows the original data. Figure 5b The image shows the synthesized data. From Figure 5a and Figure 5b It can be seen that the synthetic data still retains the data correlation of the original data.

[0133] 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.

[0134] Specifically, the variability of two different balancing algorithms was compared through 100 iterations, and the results were visually displayed using box plots. "Closer" box plots, characterized by lower interquartile ranges (IQRs) or shorter box shapes, indicate that the algorithm exhibits minimal change in results across multiple iterations, reflecting lower variability and higher robustness. This effectively demonstrates and evaluates the effectiveness of different data balancing and synthesis techniques in preserving the characteristics of the original data.

[0135] Step 5: Select the synthetic data with the highest similarity to the preprocessed cervical cancer dataset as the base data, such as... Figure 3 As shown, a dynamic feature aggregation graph neural network is used to capture the relationship of medical visit characteristics of cervical cancer patients and predict the corresponding treatment plan.

[0136] Step 5 specifically includes:

[0137] Step S5.1: For a two-dimensional structured data table of size m×n, the traditional method treats each feature value as a point and constructs a graph by calculating their similarity. Such a graph structure cannot well represent the complex relationships in the original data. To overcome this problem, we use a graph structure learning method with hidden features to obtain a more expressive learning method. First, we expand the feature dimension of the low-dimensional input X, and then use a learnable matrix for non-linear mapping to enhance the expressive power of the features; as shown in the following formula:

[0138] X out1 =σ(W1(Unsqueeze(X))) (2)

[0139] In equation (2), the low-dimensional input X is the preprocessed synthetic data from step 4; Unsqueeze represents expanding the dimension of X; W1 represents the feature-rich weights; σ represents the activation function; the above process transforms the original input data, facilitating the transition from the original space to the feature-rich space; X out1 This is the output data after step 5.1;

[0140] Step 5.2: To obtain dynamic feature aggregation and avoid the insufficient generalization caused by fixed-weight aggregation, we introduce a dynamic adjustment method to improve the feature generalization ability of the input features, so as to flexibly transform the input into a graph structure:

[0141] X out2 =W2(DFL(Reshape(X) out1 ))) (3)

[0142] In equation (3), Reshape represents the dimensionality transformation operation using DFL(·) based on two-dimensional convolution, DFL(·) represents the use of dynamically adjusted feature aggregation, and W2 represents the weights used for feature-rich 1×1 convolutions; X out2 This is the output data after step 5.2;

[0143] In signal processing, amplitude and phase are fundamental properties describing waveform characteristics, and this method has been applied to feature aggregation of medical tabular data. For example, when analyzing blood glucose levels, amplitude represents the magnitude of blood glucose fluctuations, while phase reflects the temporal pattern of these fluctuations. For DFL, the hidden features learned after dimensional expansion can be represented as vectors, assuming two hidden feature labels x i and x j Their amplitude is z and their phase is θ. Their superposition can be calculated as follows:

[0144]

[0145] In equation (4), θ i To hide feature x i phase; θ j To hide feature x j Phase; ⊙ represents element-wise multiplication; |z i | For hidden features x i The amplitude; |z j | For hidden features x i The amplitude; |z mix | represents the amplitude aggregation result of the two hidden features;

[0146] Aggregate phase term θ r The calculation is as follows:

[0147] θ δ =θ j -θ i

[0148] θ r =θ i +F(|z j |⊙sin(θ δ ), |z i |+|z j |⊙cos(θ δ (5)

[0149] In equation (5), θ δ To hide feature xi and hidden features x j The phase difference; θ r This is the phase aggregation result of two hidden features; F is the two-parameter arctangent function.

[0150] Step S5.3: In order to use a graph neural network to model the acquired multidimensional hidden features for effective feature learning, the following steps are performed: The last two dimensions are interpreted as node unit dimensions; transposing the feature-related dimensions c1×c2 and c3 will produce c2×c3 nodes, each with a feature vector of length c1 as its feature; the resulting transformed nodes for hiding features are represented as... The similarity between these feature vectors is then calculated, and connections between nodes are established using the k-nearest neighbor method, as shown in the following formula:

[0151]

[0152] In Equation (6), Metric represents the calculation of the Euclidean distance between the feature vectors of two nodes, which is used to evaluate the relationship between each pair of nodes in the node set ν and determine which nodes should have edges established; E represents the edge between neighboring nodes; G represents the graph representation of the hidden features; Indicates the global node;

[0153] The node set and its neighbors are obtained as follows:

[0154]

[0155] In equation (7), To represent the i-th subgraph The i-th node; For nodes The kth adjacent neighbor node;

[0156] Step 5.4: Based on Step 5.3, perform feature aggregation to obtain the following formula:

[0157]

[0158] In Equation (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network, respectively; This represents a single node updated after the graph convolutional network. This represents the set of all nodes in the graph; The node set represents the new graph formed by concatenating the nodes of a subset;

[0159] Step 5.5: Use fully connected layers composed of one-dimensional convolutions to enhance the richness of connection node features, and utilize standard residual connections to obtain the final output of the graph convolutional network, and predict the corresponding treatment plan for cervical cancer patients:

[0160]

[0161] In equation (9), W out Indicates the weights of the fully connected layer used; This is the final output of the graph convolutional network.

[0162] from Figure 3 It can be seen that after data preprocessing, by extracting the hidden feature amplitude and phase information, more representative features are dynamically aggregated and extracted for the final output.

[0163] This specific embodiment generates synthetic data using the GPT2 model to simulate the characteristic distribution of medical data, thereby expanding the dataset and increasing its diversity. This addresses the limitation of insufficient data for some rare and chronic diseases, while preserving the correlations between the original data (e.g., Figure 5a and Figure 5b As shown in the figure, this enables data that could only be predicted using traditional machine learning models to be predicted using deep learning models.

[0164] Furthermore, by employing a dynamic feature capture method using graph neural networks, the complex relationships between patient features can be captured. This specific embodiment compares the prediction results of seven models on raw cervical cancer data: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), TabNet (attention-interpretable table learning), and SAINT (a neural network that improves tabular data through row attention and contrastive pre-training). Table 1 shows the prediction results of different models on the raw data, and Table 2 shows the prediction results of different models on data synthesized using the SMOTE and ADASYN oversampling algorithms, respectively. The conclusion is that the present invention achieves the highest prediction accuracy, precision, recall, and F1 score.

[0165] Table 1. Prediction results based on original cervical cancer data.

[0166]

[0167] Table 2. Prediction results based on synthetic cervical cancer data.

[0168]

[0169]

[0170] In summary, the key technical point of this invention lies in generating synthetic data through the GPT2 model to simulate the feature distribution of medical data, thereby expanding the dataset and increasing data diversity. 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. Further, a dynamic feature aggregation graph neural network is used to capture the relationship of patient visit features, enabling more accurate and effective prediction of patient treatment plans. 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 fully connected layers are used to enhance the richness of connection node features.

[0171] Furthermore, the intended protection point of this invention lies in the construction and training method of a dynamic feature aggregation graph neural network, including feature transformation, dynamic adjustment, graph neural network modeling, feature aggregation, and the design of fully connected layers.

[0172] These key technologies and protection points constitute the innovation and uniqueness of this invention in predicting segmentation patterns for brachytherapy of cervical cancer, ensuring its high efficiency and accuracy in practical applications.

[0173] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting segmentation patterns in brachytherapy for cervical cancer based on graph neural networks, characterized in that, include: Step S1: Obtain the cervical cancer dataset; The cervical cancer dataset is an electronic medical record containing clinical visit information for each cervical cancer patient. Step S2: Preprocess the cervical cancer dataset; Step S3: Input the preprocessed cervical cancer dataset into the GPT2 model to generate synthetic data; Step S4: Evaluate the quality of the synthetic data and select the synthetic 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 dataset as the basic data, use the dynamic feature aggregation graph neural network to capture the relationship of the medical visit characteristics of cervical cancer patients, and predict the corresponding treatment plan; Step S5 specifically includes: Step S5.1: Expanding the low-dimensional input The feature dimensions are then used to perform non-linear mapping using a learnable matrix to enhance the expressive power of the features; as shown in the following formula: （2） In equation (2), the low-dimensional input The synthesized data after preprocessing in step S4; Representation extension The dimension; Weights representing rich features; This represents the activation function; the above process transforms the original input data, facilitating the transition from the original space to a feature-rich space. This 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, so as to flexibly transform the input into a graph structure: （3） In equation (3), Indicates the use of two-dimensional convolution-based Perform dimensional transformation operations. This indicates the use of dynamically adjusted feature aggregation. Represents the weights used for feature-rich 1×1 convolutions; This is the output data after step S5.2; For DFL, assume two hidden feature labels. and Its amplitude is Phase is Their superposition can be calculated as follows: （4） In equation (4), To hide features The phase; To hide features Phase; For element-wise multiplication; To hide features The amplitude; To hide features The amplitude; The amplitude aggregation result of two hidden features; Aggregated phase terms The calculation is as follows: （5） In equation (5), To hide features and hidden features The phase difference; The result of phase aggregation of two hidden features; It is a two-parameter arctangent function; Step S5.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require The last two dimensions are interpreted as node unit dimensions; transpose and feature-related dimensions. and This will produce There are nodes, each with a length of . The eigenvectors are features; The obtained transformed prepared hidden feature nodes are represented as follows: Then, the similarity between these feature vectors is calculated, and used The nearest neighbor method establishes connections between nodes, as shown in the following formula: （6） In equation (6), This represents calculating the Euclidean distance between the feature vectors of two nodes, used to evaluate the node set. The relationship between each pair of nodes in the graph determines which nodes should have edges established between them. Represents the edges between neighboring nodes; A graph representation of hidden features; Indicates the global node; The node set and its neighbors are obtained as follows: （7） In equation (7), To represent the i-th subgraph The i-th node; For nodes The kth adjacent neighbor node; Step S5.4: Based on step 5.3, perform feature aggregation to obtain the following formula: （8） In equation (8), Aggregate and Update represent the aggregated neighbor features and updated node features in the graph convolutional network, respectively; This represents a single node updated after the graph convolutional network. This represents the set of all nodes in the graph; The node set represents the new graph formed by concatenating the nodes of a subset; Step S5.5: Use fully connected layers composed of one-dimensional convolutions to enhance the richness of the connection node features, and utilize standard residual connections to obtain the final output of the graph convolutional network, and predict the corresponding treatment plan for cervical cancer patients: （9） In equation (9), Indicates the weights of the fully connected layer used; This is the final output of the graph convolutional network.

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

3. The method for predicting segmentation patterns for cervical cancer brachytherapy based on graph neural networks according to claim 2, characterized in that, Step S3 specifically includes: Step S3.1: Create a vocabulary for the input data using the GPT2 model, mapping the unique tokens in each column of the input data to a unique token ID, and then applying the mapping from the vocabulary to each row of the input data to generate a series of token IDs; wherein, the input data is the preprocessed cervical cancer dataset, and the preprocessed cervical cancer dataset has OK, Columns of table data ; Step S3.2: Use the GPT2 model to simulate the conditional distribution of column values ​​in each row of the input data. ; in, For is the dataset number line, number The value of the column; Step S3.3: Generate synthetic data by employing a strategy of target masking and overfitting detection; wherein the synthetic data consists of a series of observations. Composition; the observed values The expression is: （1） In equation (1), Indicates the first OK; Indicates the first List; Step S3.4: Based on the preprocessed cervical cancer dataset, five different data sizes were generated to determine the optimal size of the synthesized data.

4. The method for predicting segmentation patterns for cervical cancer brachytherapy based on graph neural networks according to claim 3, characterized in that, Step S4 specifically includes: Step S4.1: Use PCD, KLD and MMD as three metrics to measure the similarity between the synthetic data and the preprocessed cervical cancer dataset; Step S4.2: Use a triangular radar chart plotted on a logarithmic scale to represent the average values ​​of the three metrics, PCD, KLD, and MMD; wherein, in the triangular radar chart, the smaller triangle indicates a higher similarity between the synthetic data and the preprocessed cervical cancer dataset, that is, the synthetic data is of 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 for cervical cancer brachytherapy based on graph neural networks according to claim 3, characterized in that, 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.

6. The method for predicting segmentation patterns for cervical cancer brachytherapy based on graph neural networks according to any one of claims 1 to 5, characterized in that, The clinical information for 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.

Images