Tumor radiotherapy response prediction model and method based on capsule graph structure neural network

Abstract

The application relates to the technical field of medicine and deep learning, in particular to a tumor radiotherapy reaction prediction model and method based on a capsule graph structure neural network. The method comprises the following steps: acquiring a tumor data set, and extracting tumor data features from the tumor data set; classifying the tumor data features according to risk grades, respectively constructing graph structure data sets, and marking whether the tumor reacts; constructing a deep learning model based on a capsule graph structure neural network; pre-training the deep learning model, evaluating the performance of the deep learning model, and obtaining a trained deep learning model; inputting a tumor data set to be predicted into the trained deep learning model, and predicting whether the whole tumor has a radiotherapy reaction. The application can utilize image data before tumor radiotherapy and a radiotherapy dose to predict whether radiotherapy reactions of tumor regions with different severities are effective, thereby assisting doctors in reasonably formulating a radiotherapy dose, realizing precise radiotherapy, and reducing radiotherapy side effects.

Description

Technical Field

[0001] The present application relates to the fields of medicine and deep learning technology, and in particular to a tumor radiotherapy response prediction model and method based on a capsule graph structure neural network. Background Art

[0002] Lung cancer, also known as primary bronchogenic carcinoma, is the most common lung malignancy originating from the trachea, bronchial mucosa, or glands. Globally, the incidence and mortality rates of lung cancer are extremely high and on the rise. Among men, both the incidence and mortality rates of lung cancer rank first among malignant tumors, and radiotherapy is a common tumor treatment. Currently, the setting of radiotherapy doses relies heavily on the subjective decision-making of doctors. Excessive radiotherapy doses can lead to side effects such as hair loss, skin damage, and physical weakness; while insufficient radiotherapy doses cannot produce the expected therapeutic effect. Therefore, it is crucial to set different doses for sub-regions with different risk levels and ensure that the dose set for each sub-region is reasonable and effective. Summary of the Invention

[0003] The present application provides a tumor radiotherapy response prediction model and method based on a capsule graph structure neural network, which can predict whether regional radiotherapy of tumors with different severities is effective, thereby intelligently assisting doctors to rationally formulate radiotherapy doses, achieve precise radiotherapy, reduce radiotherapy reaction side effects, and improve radiotherapy effects.

[0004] To solve the above technical problems, in the first aspect, an embodiment of the present application provides a method for predicting tumor radiotherapy response based on a capsule graph structure neural network, which includes the following steps: first, obtaining a tumor data set and extracting tumor data features from the tumor data set; then, classifying the tumor data features according to the risk level; and based on the classified data features, respectively constructing a graph structure data set, and marking whether the tumor in the graph structure data set responds; next, based on the graph structure data set, constructing a deep learning model based on a capsule graph structure neural network; then, using a leave-one-out cross-validation method, pre-training the deep learning model to obtain optimized model parameters; and based on the optimized model parameters, performing a performance evaluation on the deep learning model to obtain a trained deep learning model; finally, inputting the tumor data set to be predicted into the trained deep learning model to predict whether the entire tumor responds to radiotherapy.

[0005] In some exemplary embodiments, extracting tumor data features includes extracting X, Y, and Z coordinates of each pixel of the tumor, the early stage of radiotherapy response, the mid-stage of radiotherapy response, and the radiotherapy dose from the tumor data features.

[0006] In some exemplary embodiments, tumor data features are classified according to risk levels, including: using a K-means clustering method, with the X-coordinate, Y-coordinate, Z-coordinate, and standard uptake value of the pixel point before radiotherapy response as features, to cluster all pixels of each patient's tumor into three categories, namely high-risk area, medium-risk area, and low-risk area.

[0007] In some exemplary embodiments, graph-structured data sets are constructed based on the classified data features, including: taking the standard uptake value and radiotherapy dose of the early radiotherapy response of each tumor point as node feature values, connecting each point with its neighboring points to form edges, forming a graph-structured data for each patient's tumor, and converting the classified data features into a graph-structured data set.

[0008] In some exemplary embodiments, labeling whether a tumor in a graph structured dataset responds includes determining whether the standard uptake value of the entire tumor before radiotherapy satisfies a corresponding relationship with the standard uptake value during radiotherapy; if so, determining that the entire tumor responds, labeling the radiotherapy response effective, and adding a label "0"; if not, determining that the entire tumor does not respond, labeling the radiotherapy response ineffective, and adding a label "1"; the corresponding relationship is:

[0009]

[0010] in, It represents the average value of the standard uptake value of all pixels in the entire tumor before radiotherapy, which reflects the degree of tumor lesions; It represents the average standard uptake value of all pixels in the entire tumor at the middle stage of radiotherapy.

[0011] In some exemplary embodiments, a deep learning model based on a capsule graph structured neural network is constructed based on a graph structured dataset, including: first, constructing a main architecture of a capsule graph structured neural network based on the graph structured dataset; the main architecture of the capsule graph structured neural network includes a basic node capsule extraction block, a high-level graph capsule extraction block, and a graph classification block connected in sequence; then, constructing a pre-network architecture of the capsule graph structured neural network; the pre-network architecture of the capsule graph structured neural network is used to convert tumor data into graph structured data, and input the graph structured data into the main architecture of the capsule graph structured neural network for training; based on the graph structured dataset, constructing the main architecture of the capsule graph structured neural network includes: first, using a graph neural network to extract local vertex features with different receptive fields from the graph structured dataset, and then constructing a main node capsule in the basic node capsule extraction block; then, fusing the attention module and dynamic routing in the high-level graph capsule extraction block to generate multiple capsules for the graph; and finally, applying dynamic routing again in the graph classification block to generate class capsules for graph classification.

[0012] In some example embodiments, the deep learning model is pre-trained by using a leave-one-out cross-validation method to obtain optimized model parameters; and the deep learning model is evaluated based on the optimized model parameters to obtain a trained deep learning model, including: dividing the tumor data set into a test set and a training set by using the leave-one-out cross-validation method, training the deep learning model by using the training set, and testing the trained deep learning model by using the test set; adjusting parameters of the tested deep learning model to obtain an optimized deep learning model and optimized model parameters; and evaluating the deep learning model based on the optimized model parameters according to evaluation indexes of accuracy and area under the receiver operating characteristic curve to obtain the trained deep learning model.

[0013] In some example embodiments, the tumor data set to be predicted is input into the trained deep learning model to predict whether the whole tumor has a radiotherapy response, including: pre-processing the tumor data set to be predicted to obtain a pre-processed tumor graph structure data set; and inputting the pre-processed tumor graph structure data set into the trained deep learning model to predict whether the whole tumor has a radiotherapy response.

[0014] In some example embodiments, the tumor data set to be predicted is pre-processed to obtain a pre-processed tumor graph structure data set, including: sequentially performing feature extraction and pixel point risk level clustering and division pre-processing operations on the tumor data set to be predicted to obtain the pre-processed tumor graph structure data set.

[0015] In a second aspect, the embodiments of the present application further provide a tumor radiotherapy response prediction model based on a capsule graph structure neural network, including sequentially connected data preprocessing modules, model main body modules, model training modules, model evaluation modules, and prediction modules; the data preprocessing module is used to obtain a tumor data set, and extract tumor data features from the tumor data set; classify the tumor data features according to risk levels; and based on the classified data features, respectively construct graph structure data sets, and label whether the tumors in the graph structure data sets have a response; the model main body module is used to construct a deep learning model based on a capsule graph structure neural network according to the graph structure data sets; the model training module is used to pre-train the deep learning model by using a leave-one-out cross-validation method to obtain optimized model parameters; the model evaluation module is used to evaluate the performance of the deep learning model based on the optimized model parameters to obtain a trained deep learning model; and the prediction module is used to input a tumor data set to be predicted into the trained deep learning model to predict whether the whole tumor has a radiotherapy response.

[0016] The technical solutions provided by the embodiments of the present application have at least the following advantages:

[0017] An embodiment of the present application provides a tumor radiotherapy response prediction model and method based on a capsule graph structure neural network, the method comprising the following steps: first, obtaining a tumor data set and extracting tumor data features from the tumor data set; then, classifying the tumor data features according to risk levels; and based on the classified data features, constructing graph structure data sets respectively, and annotating whether the tumors in the graph structure data sets respond; next, constructing a deep learning model based on a capsule graph structure neural network based on the graph structure data set; then, using a leave-one-out cross-validation method, pre-training the deep learning model to obtain optimized model parameters; and based on the optimized model parameters, performing a performance evaluation on the deep learning model to obtain a trained deep learning model; finally, inputting the tumor data set to be predicted into the trained deep learning model to predict whether the entire tumor responds to radiotherapy.

[0018] The tumor radiotherapy response prediction model and method based on the capsule graph structure neural network provided in the embodiments of the present application have the following advantages:

[0019] (1) This application converts the original tumor image data into graph structure data, takes the pre-radiotherapy response standard uptake value (PreSUV) and radiotherapy dose (Dose) of each tumor point as node feature values, connects each point with its neighboring points to form edges, so that it can express the three-dimensional spatial relationship of tumor pixels and can be applied to graph structure neural networks.

[0020] (2) This application first uses K-means clustering to classify tumor pixels according to their severity, and predicts whether tumor subregions with different risk levels will respond to radiotherapy, thereby improving the accuracy of prediction and promoting the implementation of precision radiotherapy.

[0021] (3) This application uses evaluation indicators to compare with traditional network models to verify the effectiveness, accuracy and adjustability of the tumor radiotherapy response prediction model.

[0022] The model provided in this application can predict whether a tumor will respond to radiotherapy, thereby intelligently assisting doctors in rationally formulating radiotherapy doses, achieving precise radiotherapy, reducing the side effects of radiotherapy reactions, and improving radiotherapy effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] One or more embodiments are exemplarily described by the pictures in the corresponding drawings. These exemplifications do not constitute limitations on the embodiments. Unless otherwise stated, the pictures in the drawings do not constitute proportional limitations.

[0024] Figure 1 A schematic flow chart of a method for predicting tumor radiotherapy response based on a capsule graph neural network according to an embodiment of the present application;

[0025] Figure 2A schematic diagram of a data preprocessing process for a method for predicting tumor radiotherapy response based on a capsule graph structure neural network provided in one embodiment of the present application;

[0026] Figure 3 A network architecture diagram of a capsule graph structure neural network provided in one embodiment of the present application;

[0027] Figure 4 A schematic diagram of the structure of a tumor radiotherapy response prediction model based on a capsule graph structure neural network provided in one embodiment of the present application. DETAILED DESCRIPTION

[0028] Currently, radiotherapy dose setting is primarily based on physician experience, which is subject to individual differences and accidental factors. This approach makes it difficult to ensure the scientific, rational, and effective treatment of patients. Therefore, scientifically and rationally setting the radiotherapy dose for the tumor target is crucial for achieving precision radiotherapy.

[0029] To address the shortcomings of the prior art, the present invention provides a tumor radiotherapy response prediction model and method based on a capsule graph-structured neural network. The method comprises the following steps: first, obtaining a tumor dataset and extracting tumor data features from the tumor dataset; then, classifying the tumor data features according to risk level; and based on the classified data features, constructing graph-structured datasets and labeling whether the tumors in the graph-structured datasets have responded; next, constructing a deep learning model based on the capsule graph-structured neural network based on the graph-structured dataset; then, using a leave-one-out cross-validation method, pre-training the deep learning model to obtain optimized model parameters; and based on the optimized model parameters, performing a performance evaluation on the deep learning model to obtain a trained deep learning model; finally, inputting the tumor dataset to be predicted into the trained deep learning model to predict whether the entire tumor will respond to radiotherapy. The model provided in the present invention can predict whether radiotherapy is effective for regions with different tumor severities, thereby intelligently assisting doctors in rationally formulating radiotherapy doses, achieving precise radiotherapy, reducing radiotherapy reaction side effects, and improving radiotherapy effects.

[0030] The following detailed description of the various embodiments of the present application is provided in conjunction with the accompanying drawings. However, those skilled in the art will appreciate that many technical details are provided in the various embodiments of the present application to facilitate a better understanding of the present application. However, even without these technical details and the various variations and modifications based on the following embodiments, the technical solutions claimed in the present application can still be implemented.

[0031] See Figure 1 The present invention provides a method for predicting tumor radiotherapy response based on a capsule graph neural network. The method comprises the following steps:

[0032] Step S1, obtain a tumor data set, and extract tumor data features from the tumor data set.

[0033] Step S2, classify the tumor data features according to the risk level; and based on the classified data features, respectively construct a graph structure data set, and label whether the tumor is reactive in the graph structure data set.

[0034] Step S3, based on the graph structure data set, a deep learning model based on a capsule graph structure neural network is constructed.

[0035] Step S4, using the leave-one-out cross-validation method, the deep learning model is pre-trained to obtain optimized model parameters; and based on the optimized model parameters, the performance of the deep learning model is evaluated to obtain a trained deep learning model.

[0036] Step S5, input the tumor data set to be predicted into the trained deep learning model to predict whether the whole tumor has a radiotherapy reaction.

[0037] The capsule graph neural network is used to solve the shortcomings of the existing graph embedding algorithm based on the graph neural network (GNN), and the attention module contained can be used to process graphs of various sizes, so that the model can focus on processing the key part of the graph. Deep neural network, it uses graph structure data in the pytorch deep learning framework to realize the classification task. It can input the graph structure data converted from the original tumor image data, take the pre-radiotherapy standard uptake value (PreSUV) and radiotherapy dose (Dose) of each point of the tumor as the node feature value, and connect each point with its adjacent points to form an edge. The data can express the three-dimensional spatial relationship of the tumor pixel points, highlight the features of the tumor data, and avoid the feature loss caused by image segmentation. Based on this, the present application proposes a tumor radiotherapy reaction prediction model and method based on a capsule graph structure neural network, which can predict whether the tumor reacts after radiotherapy, thereby intelligently assisting doctors to reasonably formulate radiotherapy doses, realizing precise radiotherapy, reducing radiotherapy side effects, and improving radiotherapy effect.

[0038] In some embodiments, the tumor data features extracted in step S1 include: extracting tumor pixel point coordinates X, Y, Z, pre-radiotherapy standard uptake value (PreSUV), mid-radiotherapy standard uptake value (MidSUV), and radiotherapy dose (Dose) from the tumor data features.

[0039] In some embodiments, in step S2, tumor data features are classified according to risk levels. Specifically, the data is divided into three categories with different risk levels. The classification process is: using the K-means clustering method, with X coordinates, Y coordinates, Z coordinates, and pixel point early standard uptake value (PreSUV) as features, all pixel points of each patient's tumor are clustered into three categories, namely high-risk area, medium-risk area, and low-risk area.

[0040] In some embodiments, based on the classified data features, graph structured data sets are constructed separately, including: taking the PreSUV and Dose of each tumor point as node feature values, connecting each point with its adjacent points to form edges, forming a graph structured data for each patient's tumor, and converting the classified data features into graph structured data, that is, converting the original tumor data into a graph structured data set.

[0041] In some embodiments, labeling whether a tumor in a graph structure dataset responds includes: determining whether the standard uptake value of the entire tumor before radiotherapy satisfies a corresponding relationship with the standard uptake value during radiotherapy; if so, determining that the entire tumor responds, labeling the radiotherapy response effective, and adding a label "0"; if not, determining that the entire tumor does not respond, labeling the radiotherapy response invalid, and adding a label "1"; the corresponding relationship is shown in formula (1):

[0042]

[0043] in, It represents the average value of the standard uptake value of all pixels in the entire tumor before radiotherapy, which reflects the degree of tumor lesions; It represents the average standard uptake value of all pixels in the entire tumor at the middle stage of radiotherapy.

[0044] It should be noted that the Standard Uptake Value (SUV) is a semiquantitative indicator commonly used in positron emission computed tomography (PET) tumor diagnosis. It is the ratio of the radioactivity of the imaging agent taken up by local tissues to the average systemic injected activity. Currently, the SUV is widely used to differentiate benign and malignant tumors, evaluate therapeutic efficacy, and predict prognosis.

[0045] Specifically, in step S3, based on the graph structure dataset, a deep learning model based on a capsule graph structure neural network is constructed, including:

[0046] Step S301: construct the main architecture of the capsule graph structure neural network based on the graph structure dataset; the main architecture of the capsule graph structure neural network includes a basic node capsule extraction block, a high-level graph capsule extraction block and a graph classification block connected in sequence.

[0047] Step S302 , constructing a pre-network architecture of a capsule graph structure neural network; the pre-network architecture of the capsule graph structure neural network is used to convert tumor data into graph structure data, and input the graph structure data into the main architecture of the capsule graph structure neural network for training.

[0048] Specifically, in step S301, the main architecture of the capsule graph structure neural network is constructed based on the graph structure dataset, including:

[0049] Step S3011: Use GNN to extract local vertex features with different receptive fields.

[0050] Step S3012: Construct a master node capsule in the basic node capsule extraction block.

[0051] Step S3013: Then, in the high-level graph capsule extraction block, the attention module and dynamic routing are fused to generate multiple capsules for the graph.

[0052] Step S3014: Dynamic routing is applied again in the graph classification block to generate class capsules for graph classification.

[0053] Specifically, in step S302, a pre-network architecture of a capsule graph structure neural network is constructed. Based on the pre-network architecture of the capsule graph structure neural network, the original tumor data is converted into graph structure data, and the graph structure data is input into the main architecture of the capsule graph structure neural network for training. Leave-one-out cross-validation is performed, and network parameters are adjusted to obtain the optimal deep learning model.

[0054] In some embodiments, step S4 uses a leave-one-out cross-validation method to pre-train the deep learning model to obtain optimized model parameters; and based on the optimized model parameters, a performance evaluation is performed on the deep learning model to obtain a trained deep learning model, including:

[0055] Step S401: Use the leave-one-out cross-validation method to divide the tumor dataset into a test set and a training set, use the training set to train the deep learning model, and use the test set to test the trained deep learning model.

[0056] Step S402: Adjust the parameters of the tested deep learning model to obtain an optimized deep learning model and optimized model parameters.

[0057] Step S403: Based on the optimized model parameters, the deep learning model is evaluated according to the evaluation indicators of accuracy and area under the receiver operating characteristic curve to obtain a trained deep learning model.

[0058] Specifically, step S4 mainly uses the leave-one-out cross-validation method to pre-train the capsule graph structure neural network model, adjust the network parameters, and evaluate the performance of the model. When the tumor dataset is divided into a test set and a training set, the leave-one-out cross-validation method is used to use each tumor data in the tumor dataset as a test set, and the rest as a training set. A small model is trained for each small model, and the error of the overall model is the average of the errors of each model. During the training process of each small model, the marginal loss function is used to calculate the classification loss, where k is the number of classifications, and λ is used to stop the initial learning to reduce the length of all class capsules. Especially when k is large, formula (2) is used for calculation:

[0059]

[0060] Where c is the capsule vector, m and T are constants.

[0061] The model parameters are adjusted based on the loss after model training convergence to minimize the loss after convergence during model validation. Finally, the model parameters that minimize the loss after convergence are selected and imported into the deep learning model. The model is then tested using the test set. The parameters of the tested deep learning model are then adjusted to obtain an optimized deep learning model and optimized model parameters. Finally, based on the optimized model parameters, the performance of the deep learning model is evaluated using evaluation metrics such as accuracy and the area under the receiver operating characteristic curve (AUC) to obtain a trained deep learning model.

[0062] Specifically, in step S5, the tumor dataset to be predicted is input into the trained deep learning model to predict whether the entire tumor will respond to radiotherapy, including:

[0063] Step S501: preprocess the tumor dataset to be predicted to obtain a preprocessed tumor graph structure dataset.

[0064] Step S502: Input the preprocessed tumor graph structure dataset into the trained deep learning model to predict whether the entire tumor responds to radiotherapy.

[0065] In some embodiments, step S501 pre-processes the tumor dataset to be predicted to obtain a pre-processed tumor graph structure dataset, including:

[0066] The tumor dataset to be predicted is subjected to preprocessing operations of feature extraction and pixel risk level clustering in sequence to obtain a preprocessed tumor graph structure dataset.

[0067] It should be noted that the format of the preprocessed tumor graph structure dataset can be correctly read by the deep learning model.

[0068] See Figure 2 The embodiment of the present application provides a data preprocessing flow diagram of a method for predicting tumor radiotherapy response based on a capsule graph structure neural network, comprising the following steps:

[0069] First, data features are extracted. The X, Y, and Z coordinates of each pixel of the tumor, the early standard uptake value of radiotherapy response (PreSUV), the mid-standard uptake value of radiotherapy response (MidSUV), and the radiotherapy dose (Dose) are extracted from the tumor data features.

[0070] Then, the data was divided into three categories with different risk levels. Using the K-means clustering method, all pixels of each patient's tumor were clustered into three categories: high-risk area, medium-risk area, and low-risk area, based on the X-coordinate, Y-coordinate, Z-coordinate, and pre-radiotherapy response standard uptake value (PreSUV) of the pixel.

[0071] Next, a graph-structured dataset is constructed, with the PreSUV and Dose of each tumor point as node feature values, and each point and its neighboring points forming edges. Each patient's tumor forms a graph-structured data set, and the original tumor data is converted into a graph-structured dataset.

[0072] Finally, the tumor response is marked by determining whether the standard uptake value of the entire tumor before radiotherapy satisfies the corresponding relationship with the standard uptake value during radiotherapy. If so, the entire tumor is judged to have responded, the radiotherapy response is effective, and a label of "0" is added. If not, the entire tumor is judged to have not responded, the radiotherapy response is ineffective, and a label of "1" is added. The corresponding relationship is:

[0073]

[0074] in, It represents the average value of the standard uptake value of all pixels in the entire tumor before radiotherapy, which reflects the degree of tumor lesions; It represents the average standard uptake value of all pixels in the entire tumor at the middle stage of radiotherapy.

[0075] See Figure 3, an embodiment of the present application provides a network architecture diagram of a capsule graph structure neural network, including a network main architecture of a capsule graph structure neural network and a network pre-architecture of a capsule graph structure neural network. The network main architecture first uses GNN to extract local vertex features with different receptive fields, and then constructs a main node capsule in the basic node capsule extraction block; then, in the high-level graph capsule extraction block, the attention module and dynamic routing are integrated to generate multiple capsules for the graph; finally, dynamic routing is applied again in the graph classification block to generate class capsules for graph classification. The network pre-architecture of the capsule graph structure neural network is constructed, the original tumor data is converted into graph structure data, and the graph structure data is input into the network main architecture of the capsule graph structure neural network for training, and leave-one-out cross-validation is performed to adjust the network parameters to obtain the optimal deep learning model.

[0076] See Figure 4 , the embodiment of the present application also provides a tumor radiotherapy response prediction model based on a capsule graph structure neural network, including a data preprocessing module 101, a model main body module 102, a model training module 103, a model evaluation module 104 and a prediction module 105 connected in sequence; the data preprocessing module 101 is used to obtain a tumor data set and extract tumor data features from the tumor data set; classify the tumor data features according to the risk level; and based on the classified data features, respectively construct a graph structure data set, and mark whether the tumor in the graph structure data set responds; the model main body module 102 is used to construct a deep learning model based on the capsule graph structure neural network according to the graph structure data set; the model training module 103 is used to pre-train the deep learning model using a leave-one-out cross-validation method to obtain optimized model parameters; the model evaluation module 104 is used to perform performance evaluation on the deep learning model based on the optimized model parameters to obtain a trained deep learning model; the prediction module 105 is used to input the tumor data set to be predicted into the trained deep learning model to predict whether the entire tumor responds to radiotherapy.

[0077] The following describes in detail the tumor radiotherapy response prediction model based on capsule graph structure neural network provided by the present application with respect to a specific embodiment.

[0078] First, the tumor radiotherapy response prediction model based on capsule graph structure neural network proposed in the embodiment of the present application is based on Windows 11 system and implemented using Python 3.9.

[0079] In addition, it should be noted that the examples of the present application analyze and evaluate the radiotherapy response of non-small cell lung cancer.

[0080] First, data features were extracted. The X, Y, and Z coordinates of each tumor pixel, the pre-radiotherapy response standardized uptake value (PreSUV), the mid-radiotherapy response standardized uptake value (MidSUV), and the radiotherapy dose (Dose) were extracted from the tumor data. The data was then divided into three risk categories. Using the K-means clustering method, all pixels in each patient's tumor were clustered into three categories: high-risk, medium-risk, and low-risk areas, using the X, Y, and Z coordinates and the PreSUV as features. A graph-structured dataset was constructed, with the PreSUV and Dose values ​​of each tumor point used as node features. Each point formed edges with its neighbors, creating a graph-structured dataset for each patient's tumor. The original tumor data was converted into a graph-structured dataset. Each patient's tumor was labeled as responding or not. If the entire tumor responded, the response was considered effective and labeled "0." If the entire tumor did not respond, the response was considered ineffective and labeled "1."

[0081] Then, a deep learning model based on a capsule graph neural network was constructed. First, a GNN was used to extract local vertex features with different receptive fields. Then, a master node capsule was constructed in the basic node capsule extraction block. Then, in the advanced graph capsule extraction block, an attention module and dynamic routing were integrated to generate multiple capsules for the graph. Finally, dynamic routing was applied again in the graph classification block to generate class capsules for graph classification. The graph data was input into the main architecture of the capsule graph neural network for training. The tumor data of each patient in the tumor dataset was used as the test set, and the remaining data served as the training set. A model was trained for each patient. The overall model error was calculated as the average of the errors of each model. Parameter adjustments were performed to obtain an optimized deep learning model and optimized model parameters.

[0082] Next, the performance of the deep learning model was evaluated using an evaluation module. Evaluation metrics included accuracy and the area under the receiver operating characteristic curve (AUC). For the entire tumor region of the patient, the model achieved an AUC of 0.6929 on the test set; for the high-risk tumor region, the model achieved an AUC of 0.5952 on the test set; and for the intermediate-risk tumor region, the model achieved an AUC of 0.5317 on the test set. This indicates that the model performs best for predicting the entire tumor region, followed by high-risk tumor regions.

[0083] Finally, the pre-processed tumor dataset that needs to be predicted is input into the trained deep learning model. The trained deep learning model gives the prediction results through the prediction module, which can predict whether the tumor responds after radiotherapy.

[0084] According to the technical scheme, the embodiment of the present application provides a tumor radiotherapy reaction prediction model and method based on a capsule graph structure neural network, which comprises the following steps: first, a tumor data set is obtained, and tumor data features are extracted from the tumor data set; then, the tumor data features are classified according to risk levels; and based on the classified data features, graph structure data sets are respectively constructed, and whether the tumor reacts is labeled in the graph structure data sets; next, based on the graph structure data sets, a deep learning model based on the capsule graph structure neural network is constructed; then, the deep learning model is pre-trained by using a leave-one-out cross-validation method, and optimized model parameters are obtained; and based on the optimized model parameters, the performance of the deep learning model is evaluated, and a trained deep learning model is obtained; finally, a tumor data set to be predicted is input into the trained deep learning model, and whether the whole tumor has a radiotherapy reaction is predicted.

[0085] The present application provides a tumor radiotherapy reaction prediction model and method based on a capsule graph structure neural network. First, the present application converts original tumor image data into graph structure data, takes each point PreSUV and Dose of the tumor as a node feature value, and connects each point with its adjacent points to form an edge, so that it can express the three-dimensional spatial relationship of the tumor pixel points and can be applied to a graph structure neural network. Second, the present application first classifies tumor pixel points according to their severity by using K-means clustering, respectively predicts whether the tumor sub-regions of different risk levels have a radiotherapy reaction, improves the accuracy of prediction, and promotes the implementation of precise radiotherapy. Finally, the present application compares the evaluation index with the traditional network model, which can verify the effectiveness, accuracy and adjustability of the built tumor radiotherapy reaction prediction model. The model provided by the present application can predict whether the tumor has a reaction after radiotherapy, thereby intelligently assisting doctors to reasonably formulate radiotherapy doses, realizing precise radiotherapy, reducing the side effects of radiotherapy, and improving the effect of radiotherapy.

[0086] Those skilled in the art can understand that the above embodiments are specific examples of the present application, and in actual application, various changes can be made in form and details without departing from the spirit and scope of the present application. Any person skilled in the art can make changes and modifications without departing from the spirit and scope of the present application, therefore the protection scope of the present application should be limited by the scope defined in the claims.

Claims

1. A method for predicting tumor radiotherapy response based on capsule graph structure neural network, characterized in that: The following steps are involved: Acquiring a tumor dataset, and extracting tumor data features from the tumor dataset; Classifying the tumor data features according to risk levels; constructing graph structure data sets based on the classified data features, and marking whether the tumors in the graph structure data sets are responsive; Based on the graph structure dataset, a deep learning model based on a capsule graph structure neural network is constructed; Pre-training the deep learning model using a leave-one-out cross-validation method to obtain optimized model parameters; and performing a performance evaluation on the deep learning model based on the optimized model parameters to obtain a trained deep learning model; The tumor dataset to be predicted is input into the trained deep learning model to predict whether the entire tumor will respond to radiotherapy; Extract tumor data features, including: Extract the X, Y, and Z coordinates of each pixel of the tumor, the standard uptake value (SEV) in the early stage of radiotherapy response, the SEV in the mid stage of radiotherapy response, and the radiotherapy dose from the tumor data features; Based on the classified data features, graph structure datasets are constructed, including: The standard uptake value and radiotherapy dose of each tumor point in the early stage of radiotherapy response are used as node feature values. Each point is connected to its neighboring points to form edges. Each patient's tumor forms a graph structure data. The classified data features are converted into a graph structure data set. Based on the graph structure dataset, a deep learning model based on a capsule graph structure neural network is constructed, including: Based on the graph structure data set, the main architecture of the capsule graph structure neural network is constructed; the main architecture of the capsule graph structure neural network includes a basic node capsule extraction block, a high-level graph capsule extraction block and a graph classification block connected in sequence; Constructing a pre-network architecture of a capsule graph structure neural network; the pre-network architecture of the capsule graph structure neural network is used to convert tumor data into graph structure data, and input the graph structure data into the main architecture of the capsule graph structure neural network for training; Based on the graph structure dataset, the main architecture of the capsule graph structure neural network is constructed, including: First, a graph neural network is used to extract local vertex features with different receptive fields from the graph structure dataset, and then the main node capsule is constructed in the basic node capsule extraction block. Then, the attention module and dynamic routing are fused in the high-level graph capsule extraction block to generate multiple capsules for the graph. Finally, dynamic routing is applied again in the graph classification block to generate class capsules for graph classification.

2. The method for predicting tumor radiotherapy response based on capsule graph structure neural network according to claim 1, characterized in that: The tumor data features are classified according to risk levels, including: Using the K-means clustering method, all pixels of each patient's tumor were clustered into three categories: high-risk area, medium-risk area, and low-risk area, based on the X-coordinate, Y-coordinate, Z-coordinate, and standard uptake value of the pixel point before radiotherapy response.

3. The method for predicting tumor radiotherapy response based on capsule graph structure neural network according to claim 1, characterized in that: Annotating whether the tumor in the graph structure dataset responds includes: Determine whether the standard uptake value of the entire tumor before radiotherapy and the standard uptake value during radiotherapy satisfy the corresponding relationship; If yes, the whole tumor is judged to have responded, the radiotherapy response is marked as effective, and a label "0" is added; If not, it is determined that there is no response to the entire tumor, and the radiotherapy response is marked as ineffective, with an additional label of "1"; The corresponding relationship is: in, It represents the average value of the standard uptake value of all pixels in the entire tumor before radiotherapy, which reflects the degree of tumor lesions; It represents the average standard uptake value of all pixels in the entire tumor at the middle stage of radiotherapy.

4. The method for predicting tumor radiotherapy response based on capsule graph structure neural network according to claim 1, characterized in that: The deep learning model is pre-trained using a leave-one-out cross-validation method to obtain optimized model parameters; and based on the optimized model parameters, the deep learning model is performance evaluated to obtain a trained deep learning model, including: Using a leave-one-out cross-validation method, the tumor dataset is divided into a test set and a training set, the training set is used to train the deep learning model, and the trained deep learning model is tested using the test set; Adjust the parameters of the tested deep learning model to obtain an optimized deep learning model and optimized model parameters; Based on the optimized model parameters, the performance of the deep learning model is evaluated according to the evaluation indicators of accuracy and area under the receiver operating characteristic curve to obtain a trained deep learning model.

5. The method for predicting tumor radiotherapy response based on capsule graph structure neural network according to claim 1, characterized in that: The tumor dataset to be predicted is input into the trained deep learning model to predict whether the entire tumor will respond to radiotherapy, including: Preprocessing the tumor data set to be predicted to obtain a preprocessed tumor graph structure data set; The preprocessed tumor graph structure dataset is input into the trained deep learning model to predict whether the entire tumor responds to radiotherapy.

6. The method for predicting tumor radiotherapy response based on capsule graph structure neural network according to claim 5, characterized in that: The tumor dataset to be predicted is preprocessed to obtain a preprocessed tumor graph structure dataset, including: The tumor dataset to be predicted is subjected to preprocessing operations of feature extraction and pixel risk level clustering in sequence to obtain a preprocessed tumor graph structure dataset.

7. A tumor radiotherapy response prediction model based on a capsule graph structure neural network, the model being used to implement the tumor radiotherapy response prediction method based on a capsule graph structure neural network as claimed in any one of claims 1 to 6, characterized in that: It includes a data preprocessing module, a model main module, a model training module, a model evaluation module and a prediction module connected in sequence; The data preprocessing module is used to obtain a tumor data set and extract tumor data features from the tumor data set; Classifying the tumor data features according to risk levels; constructing graph structure data sets based on the classified data features, and marking whether the tumors in the graph structure data sets are responsive; The model main body module is used to construct a deep learning model based on a capsule graph structure neural network according to the graph structure dataset; The model training module is used to pre-train the deep learning model using a leave-one-out cross-validation method to obtain optimized model parameters; The model evaluation module is used to perform performance evaluation on the deep learning model based on the optimized model parameters to obtain a trained deep learning model; The prediction module is used to input the tumor data set to be predicted into the trained deep learning model to predict whether the entire tumor will respond to radiotherapy.

Images