Method and System for Predicting Probability of Radiation-Induced Normal Tissue Injury Based on Deep Learning

Abstract

The present invention relates to a method and system for predicting the probability of normal tissue damage based on deep learning, belonging to the technical field of medical data processing. The method includes: obtaining the tumor dose data, basic information, and tumor diagnosis information of a patient; the tumor dose data is formulated by a doctor based on the patient's tumor diagnosis information before radiotherapy for the patient; the basic information includes at least the patient's age; based on the tumor dose data, obtaining the dose volume histogram of the patient; based on the dose volume histogram, obtaining the volume parameters and biological equivalent dose of the patient's normal tissue; inputting the volume parameters and biological equivalent dose of the patient's normal tissue, the patient's basic information, and tumor diagnosis information into a normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissue. Compared with traditional statistical models or empirical formulas, the prediction accuracy of this method is higher, and it can more accurately evaluate the damage probability of normal tissue.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a method and system for predicting the probability of radioactive normal tissue damage based on deep learning. Background Art

[0002] In tumor treatment, radiotherapy is one of the important means. Radiotherapy is a local treatment method that uses various rays with different energies to irradiate tumors to inhibit and kill cancer cells. These rays can destroy the internal components of human tissue cells, directly or indirectly act on DNA molecules, inhibit or kill tumor cells, so as to achieve the treatment purpose. However, during the radiotherapy process, in addition to irradiating and eliminating tumors, the rays will also have a certain impact on normal human tissues.

[0003] However, research has found that for the radiotherapy of patients' tumors at present, most are formulated by doctors based on experience, or the probability of damage is predicted according to a statistical model as an auxiliary. In other words, the existing methods are all based on fixed statistical formulas or doctor experience rules to formulate radiotherapy doses. However, the levels of doctors vary and are subjective, lacking consistency and standardization in radiotherapy dose formulation, and also lacking accuracy. And traditional statistical models are too simple to capture complex relationships, and the prediction accuracy is limited. In other words, the existing methods will all affect the accuracy of doctors in formulating radiotherapy doses. If the formulated radiotherapy dose is too high, it will increase the risk of normal tissue complications; while if the formulated radiotherapy dose is too low, it will affect the tumor control effect. Summary of the Invention

[0004] To solve the above-mentioned problems of the prior art, the present invention provides a method and system for predicting the probability of radioactive normal tissue damage based on deep learning.

[0005] In a first aspect, an embodiment of the present application provides a method for predicting the probability of radioactive normal tissue damage based on deep learning, including: obtaining the tumor dose data, basic information, and tumor diagnosis information of a patient; wherein, the tumor dose data is formulated by a doctor according to the tumor diagnosis information of the patient before radiotherapy for the patient; the basic information at least includes the age of the patient; based on the tumor dose data, obtaining the dose volume histogram of the patient; based on the dose volume histogram, obtaining the volume parameter and biological equivalent dose of the normal tissue of the patient; wherein, the conversion formula of the biological equivalent dose is:

[0006] ; wherein, represents the biological equivalent dose when the physical dose extracted from the dose volume histogram after conversion is divided by a dose of 2 Gy; represents the dose per fraction in the radiotherapy plan as , the total physical dose when the number of divisions is , represents the cell proliferation correction factor, and the value of is taken as 3 Gy; the volume parameter and biological equivalent dose of the normal tissue of the patient, as well as the basic information and tumor diagnosis information of the patient are input into a normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the normal tissue of the patient.

[0007] Optionally, the tumor diagnosis information includes the tumor type; the normal tissue damage probability prediction model constructed based on deep learning is a combined model group; different combined model groups are constructed in advance according to different tumor types to form a prediction model library; the step of inputting the volume parameter and biological equivalent dose of the normal tissue of the patient, as well as the basic information and tumor diagnosis information of the patient into a normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the normal tissue of the patient includes: based on the tumor type of the patient, determining a target combined model group from the prediction model library; inputting the volume parameter and biological equivalent dose of the normal tissue of the patient, as well as the basic information and tumor diagnosis information of the patient into the target combined model group to output the damage probabilities of multiple normal tissues associated with the tumor type of the patient through the target combined model group.

[0008] Optionally, the number of normal tissue damage probability prediction models constructed based on deep learning is multiple and they are network models of different types; the step of inputting the volume parameter and biological equivalent dose of the normal tissue of the patient, as well as the basic information and tumor diagnosis information of the patient into a normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the normal tissue of the patient includes: inputting the volume parameter and biological equivalent dose of the normal tissue of the patient, as well as the basic information and tumor diagnosis information of the patient into multiple network models to output the damage probabilities of the normal tissue of the patient through each network model; obtaining the weights of each network model; based on the weights of each network model and the damage probabilities of the normal tissue of the patient output by each network model, obtaining a combined damage probability; wherein, the combined damage probability is the finally obtained damage probability of the normal tissue of the patient.

[0009] Optionally, the step of obtaining the weights of each network model includes: determining the age interval to which the patient belongs based on the basic information of the patient; determining the range interval of the tumor size of the patient based on the tumor diagnosis information of the patient; determining the weights of each network model according to the age interval to which the patient belongs, the range interval of the tumor size of the patient, and the model accuracy of each network model in this age interval and for this range interval of the tumor size.

[0010] Optionally, obtaining the weights of each of the network models includes: determining the weights of each of the network models according to the input data type of the tumor diagnosis information of the patient.

[0011] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a convolutional neural network model; the structure of the convolutional neural network model includes: a plurality of convolutional layers, a batch normalization layer, a Dropout layer, a max pooling layer, and a fully connected layer.

[0012] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a deep neural network model; the structure of the deep neural network model includes: being constructed with a plurality of fully connected layers as the core; and simultaneously including a batch normalization layer and a Dropout layer.

[0013] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a residual network model; the structure of the residual network model includes: being constructed with residual blocks as the core; and simultaneously including a convolutional layer, a batch normalization layer, an activation function layer, and a fully connected layer.

[0014] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a long short-term memory network model; the structure of the long short-term memory network model includes: being constructed with a plurality of LSTM layers as the core; and simultaneously including a batch normalization layer and a Dropout layer.

[0015] In a second aspect, an embodiment of the present application provides a deep learning-based radioactive normal tissue damage probability prediction system, including: a first acquisition module, configured to acquire tumor dose data, basic information, and tumor diagnosis information of a patient; wherein, the tumor dose data is formulated by a doctor according to the tumor diagnosis information of the patient before radiotherapy for the patient; the basic information includes at least the age of the patient; a second acquisition module, configured to acquire a dose volume histogram of the patient based on the tumor dose data; a third acquisition module, configured to acquire volume parameters and biological equivalent dose of normal tissue of the patient based on the dose volume histogram; wherein, the conversion formula of the biological equivalent dose is:

[0016] ; wherein, represents the biological equivalent dose when the physical dose extracted from the dose volume histogram after conversion is divided into 2 Gy as the segmentation dose; represents the total physical dose when the segmentation dose in the radiotherapy plan is and the number of segmentation times is , represents the cell proliferation correction factor, The value is taken as 3 Gy; a prediction module, configured to input the volume parameter and biological equivalent dose of the normal tissue of the patient, as well as the basic information and tumor diagnosis information of the patient, into a normal tissue damage probability prediction model constructed based on deep learning, so as to obtain the damage probability of the normal tissue of the patient.

[0017] The beneficial effects of the present invention include:

[0018] First, in the embodiments of the present application, by introducing a normal tissue damage probability prediction model based on deep learning, high-dimensional features can be extracted from complex data, and the potential relationships among the tumor dose data, tumor diagnosis information, and individual information of the patient can be fully explored. Compared with traditional statistical models or empirical formulas, the prediction accuracy of this method is higher, and the damage probability of normal tissue can be evaluated more accurately.

[0019] Second, a prediction method provided by the embodiments of the present application through a deep learning prediction model and combined with multi-dimensional information fusion analysis can improve the data utilization rate and make up for the problem that traditional models can only rely on single data, resulting in low prediction accuracy.

[0020] Third, considering the consistency of data utilization of the deep learning prediction model, in the embodiments of the present application, the physical dose directly obtained from the dose volume histogram is pre-converted, and the physical dose is converted into a biological equivalent dose with a 2 Gy fractional dose, which can more scientifically input the patient dose distribution into the deep learning prediction model, enhance the prediction ability and clinical relevance of the model, and further standardize the dose distribution of different radiotherapy plans, making the dose distributions between different treatment plans comparable.

[0021] Fourth, the above solution has a wide adaptability and can be applied to various tumor types and their corresponding normal tissue damage prediction requirements, with strong versatility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flowchart of the steps of the first method for predicting the probability of radioactive normal tissue damage based on deep learning provided by the embodiments of the present invention;

[0023] Figure 2 It is a schematic diagram of the prediction effect of the convolutional neural network model provided by the embodiments of the present invention;

[0024] Figure 3 It is a schematic diagram of the prediction effect of the deep neural network model provided by the embodiments of the present invention;

[0025] Figure 4 It is a schematic diagram of the prediction effect of the residual network model provided by the embodiments of the present invention;

[0026] Figure 5 Schematic diagram of the prediction effect of the long short-term memory network model provided by the embodiment of the present invention;

[0027] Figure 6 Flowchart of the steps of the second deep learning-based method for predicting the probability of normal tissue damage caused by radiotherapy provided by the embodiment of the present invention;

[0028] Figure 7 Flowchart of the steps of the third deep learning-based method for predicting the probability of normal tissue damage caused by radiotherapy provided by the embodiment of the present invention;

[0029] Figure 8 Flowchart of the steps of the fourth deep learning-based method for predicting the probability of normal tissue damage caused by radiotherapy provided by the embodiment of the present invention;

[0030] Figure 9 Block diagram of the modules of a system for predicting the probability of normal tissue damage caused by radiotherapy based on deep learning provided by the embodiment of the present invention;

[0031] Figure 10 Block diagram of the modules of an electronic device provided by the embodiment of the present invention. Detailed implementation manners

[0032] In the following description, specific details such as specific system structures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0033] In addition, in the description of the specification of the present application and the appended claims, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0034] Research has found that for the radiotherapy of tumors in current patients, most radiotherapy doses are formulated by doctors based on experience, or the probability of damage is predicted according to a statistical model as an aid. In other words, the existing methods all formulate radiotherapy doses according to fixed statistical formulas or doctor experience rules. However, the levels of doctors vary and are subjective, lacking consistency and standardization in radiotherapy dose formulation, and also lacking accuracy. Moreover, traditional statistical models are too simple to capture complex relationships, and the prediction accuracy is limited. In other words, the existing methods will all affect the accuracy of doctors in formulating radiotherapy doses. If the formulated radiotherapy dose is too high, the risk of normal tissue complications will increase; while if the formulated radiotherapy dose is too low, the problem of affecting the tumor control effect will occur.

[0035] In view of the above problems, the present application proposes the following embodiments to solve the above technical problems.

[0036] Please refer to Figure 1 , an embodiment of the present application provides a method for predicting the probability of normal tissue injury due to radiation based on deep learning, including: Step 101 to Step 104.

[0037] Step 101: Obtain the tumor dose data, basic information, and tumor diagnosis information of the patient.

[0038] Among them, the tumor dose data is formulated by a doctor according to the patient's tumor diagnosis information before radiotherapy for the patient, and the tumor dose data includes the radiotherapy dose. Here, the patient's tumor diagnosis information may include, but is not limited to, tumor type, tumor size, tumor location, etc. After obtaining the patient's tumor diagnosis information, the doctor formulates the tumor dose data. The above basic information of the patient includes at least the patient's age.

[0039] Step 102: Based on the tumor dose data, obtain the dose-volume histogram of the patient.

[0040] Among them, the dose-volume histogram (DVH, Dose and Volume Histogram) corresponding to the tumor dose data can be directly generated automatically based on the radiotherapy treatment planning system.

[0041] Step 103: Based on the dose-volume histogram, obtain the volume parameters and biological equivalent dose of the patient's normal tissue.

[0042] It should be noted that since the dose-volume histogram is a physical dose, it is not convenient to unify the standard and subsequent prediction. Therefore, after extracting the physical dose from the dose-volume histogram, it is converted into a biological equivalent dose.

[0043] Among them, the conversion formula for the biological equivalent dose is:

[0044] ; where represents the biological equivalent dose when the physical dose extracted from the dose-volume histogram is divided into 2 Gy; represents the total physical dose when the fractionation dose in the radiotherapy plan is , and the number of fractions is , represents the cell proliferation correction factor. Considering that the normal tissue is normal late-responding tissue, the value of is taken as 3 Gy.

[0045] Step 104: Input the volume parameters and biological equivalent doses of the patient's normal tissues, as well as the patient's basic information and tumor diagnosis information, into the normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissues.

[0046] In this application, a normal tissue damage probability prediction model is pre-constructed based on deep learning to predict the damage probability of the patient's normal tissues. The deep learning model can automatically extract high-order features from large-scale data and adapt to complex non-linear relationships.

[0047] It should be noted that this solution can be applied to the damage probability of most tumors and a corresponding normal tissue of the tumor.

[0048] Exemplarily, the normal tissue damage probability prediction model constructed based on deep learning can be a pre-constructed damage prediction model for the lung tissue corresponding to lung cancer. Then an application scenario of the above solution is that it can obtain the damage probability prediction for the lung tissue based on the tumor dose data, basic information, and tumor diagnosis information of lung cancer patients.

[0049] Exemplarily, the normal tissue damage probability prediction model constructed based on deep learning can be a pre-constructed damage prediction model for the throat soft tissue corresponding to nasopharyngeal carcinoma. Then another application scenario of the above solution is that it can obtain the damage probability prediction for the throat soft tissue based on the tumor dose data, basic information, and tumor diagnosis information of nasopharyngeal carcinoma patients.

[0050] Of course, in other application scenarios, the type of tumor can also be brain tumor, esophageal cancer, etc., and the corresponding normal tissues can also be brainstem tissue, optic nerve tissue, lung tissue, and so on.

[0051] In summary, the deep learning-based radioactive normal tissue damage probability prediction method provided by the embodiments of this application has the following beneficial effects:

[0052] First, in the embodiments of this application, by introducing a normal tissue damage probability prediction model based on deep learning, high-dimensional features can be extracted from complex data, and the potential relationships among the patient's tumor dose data, tumor diagnosis information, and individual information can be fully explored. Compared with traditional statistical models or empirical formulas, the prediction accuracy of this method is higher, and the damage probability of normal tissues can be evaluated more accurately.

[0053] Second, a prediction method provided by the embodiments of this application through a deep learning prediction model and combined with multi-dimensional information fusion analysis can improve the data utilization rate and make up for the problem that traditional models can only rely on single data, resulting in low prediction accuracy.

[0054] Third, considering the consistency of data utilization in the prediction model using deep learning, in the embodiments of the present application, the physical dose directly obtained from the dose volume histogram is pre-converted, and the physical dose is converted into a biological equivalent dose with a 2Gy fractional dose, which can more scientifically input the patient dose distribution into the prediction model of deep learning, enhance the prediction ability and clinical relevance of the model, and further standardize the dose distribution of different radiotherapy plans, making the dose distributions between different treatment plans comparable.

[0055] Fourth, the above solution has a wide range of adaptability and can be applied to various tumor types and their corresponding normal tissue damage prediction requirements, with strong versatility and scalability.

[0056] In the embodiments of the present application, multiple different types of network models are pre-constructed to construct the normal tissue damage probability model based on deep learning in the above embodiments.

[0057] Next, various types of network models provided by the embodiments of the present application will be described.

[0058] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a convolutional neural network model (CNN model).

[0059] The structure of the convolutional neural network model includes: multiple convolutional layers, batch normalization layers, Dropout layers, max pooling layers, and fully connected layers.

[0060] That is, in the embodiments of the present application, a complex convolutional neural network model is constructed, which includes multiple convolutional layers, batch normalization layers, Dropout layers, max pooling layers, and fully connected layers. By defining the conv_block function to construct convolutional blocks, within each convolutional block, the input data sequentially undergoes convolutional layer, batch normalization, and activation function operations, and this structure helps to extract and enhance the local features of the data. In the entire model construction, first, an initial convolutional layer, batch normalization, and activation function are set. Then, different groups of convolutional blocks are added in multiple loops, and the number of convolutional blocks in each group is different. Max pooling layers are added at appropriate positions to reduce the data dimension and highlight important features. Finally, the features are converted into vectors through a global average pooling layer and then classified through a fully connected layer, which can be used to process one-dimensional data.

[0061] Since the convolutional neural network model has the characteristic of being good at capturing spatial correlations, it can be better applied to scenarios where tumors are close to complex normal tissues, such as head and neck tumors and brain tumors.

[0062] Please refer to Figure 2 , Figure 2This is a prediction effect display image for a kind of tumor based on a convolutional neural network model provided by an embodiment of the present application, where the abscissa is the tumor dose data of the patient, and the ordinate is the damage probability of the patient's normal tissue.

[0063] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a deep neural network model (DNN model).

[0064] The structure of the deep neural network model includes: being constructed with multiple fully connected layers as the core; and also including a batch normalization layer and a Dropout layer.

[0065] The model building of this deep neural network model adopts a sequential manner. First, a reshaping layer is used to adjust the input data into a suitable shape to prepare for the processing of the fully connected layer. Each fully connected layer uses the ReLU activation function to introduce non-linearity, enabling the network to learn complex functional relationships. A batch normalization layer is added after each fully connected layer to accelerate the training process and improve the model stability, reducing the problem of internal covariate shift. The Dropout layer randomly discards neuron outputs at a certain ratio to prevent the model from overfitting and enhance the generalization ability. Finally, through the fully connected output layer, the features extracted previously are mapped to a suitable output dimension, and this deep neural network model can process data of various dimensions.

[0066] Please refer to Figure 3 , Figure 3 This is a prediction effect display image for a kind of tumor based on a deep neural network model provided by an embodiment of the present application, where the abscissa is the tumor dose data of the patient, and the ordinate is the damage probability of the patient's normal tissue.

[0067] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a residual network model (ResNet model).

[0068] The structure of the residual network model includes: being constructed with residual blocks as the core; and also including a convolutional layer, a batch normalization layer, an activation function layer, and a fully connected layer.

[0069] Specifically, the residual network model constructs a neural network model with residual blocks as the core, covering convolutional layers, batch normalization layers, activation function layers, and fully connected layers. First, the residual_block function is defined to construct the residual block. Inside the function, the input is saved as the shortcut first. After passing through two convolutional layers and corresponding batch normalization and activation function operations, the dimension of the shortcut can be adjusted according to the change in stride or the number of channels. Then, the convolutional result and the shortcut are added together and activated again to achieve the residual connection, effectively solving the gradient problem in the training of deep networks. In the model construction, an initial convolutional layer, batch normalization, and activation function, as well as a max pooling layer, are set first. Then, different groups of residual blocks are added in multiple loops according to the design. The number and parameters of the residual blocks in each group are different, gradually extracting the complex features of the image. Finally, the features are converted into vectors through the global average pooling layer and then classified through the fully connected layer, which is suitable for processing data with complex hierarchical structures such as images.

[0070] Please refer to Figure 4 , Figure 4 which is the image showing the prediction effect of a tumor based on the residual network model provided by the embodiment of the present application. The abscissa is the tumor dose data of the patient, and the ordinate is the damage probability of the patient's normal tissue.

[0071] Optionally, the normal tissue damage probability prediction model constructed based on deep learning is a long short-term memory network model (LSTM model).

[0072] The structure of the long short-term memory network model includes: being constructed with multiple LSTM layers as the core; and including a batch normalization layer and a Dropout layer at the same time.

[0073] Specifically, first, a neural network model with long short-term memory units as the core is constructed, which includes multiple LSTM layers, batch normalization layers, and Dropout layers. The LSTM module is constructed by defining the lstm_module function. Within each module, the input data first passes through an LSTM layer, which uses its special gate structure (input gate, forget gate, and output gate) to process the long-term dependencies in the sequence data. Then, the data is normalized by the batch normalization layer to stabilize the training process and reduce the impact of data distribution changes on model training. Next is the Dropout layer, which randomly discards neuron outputs with a certain probability to prevent overfitting. During the model construction process, the shape of the input data is specified first, including the sequence length and the feature dimension at each time step. Then, the lstm_module is added multiple times to construct LSTM structures at different levels. The return_sequences of the first few lstm_modules is set to True to continue processing the sequence information, and the return_sequences of the last lstm_module is set to False to extract the final representation of the entire sequence. Finally, the features output by the LSTM module are mapped to the appropriate output dimension through a fully connected layer, which is specifically used to process sequence data.

[0074] Please refer to Figure 5 , Figure 5 which is the image showing the prediction effect of a tumor based on the long short-term memory network model provided by the embodiment of the present application. The abscissa is the tumor dose data of the patient, and the ordinate is the damage probability of the patient's normal tissue.

[0075] It should be noted that the above network models provided by the embodiments of the present application all need to perform a conversion of the biological equivalent dose. That is, it is converted into the equivalent uniform dose of each network model.

[0076] The equivalent uniform dose of each network model can be calculated as follows:

[0077] ;

[0078] In this formula, represents the equivalent uniform dose; The biological equivalent dose of the normal tissue is divided into equal parts, is the dose of the th equal part, then is the corresponding normal tissue volume; is the Lebesgue norm.

[0079] The training process of the normal tissue damage probability prediction model based on deep construction provided by the embodiments of the present application will be described below. This training process can be applied to any of the above types of network models, and of course, it can also be applied to other types of network models not listed.

[0080] Specifically, the training of the network model is carried out through the following steps, including: obtaining sample data; wherein, the sample data is the detection data of tumor patients before and after radiotherapy; the detection data of tumor patients after radiotherapy is the detection data of normal tissues of tumor patients in the sixty months after radiotherapy; based on the sample data, the initial network model is trained, and the model parameters of the initial network model are adjusted to obtain a network model in the mathematical model library.

[0081] The above sample data is the detection data of tumor patients before and after radiotherapy. Among them, the detection data before radiotherapy is mainly the tumor dose data, basic information and tumor diagnosis information of the patient, while the detection data after radiotherapy mainly includes the detection of the impact on the normal tissue damage of the patient. The detection data after radiotherapy is used as a label to perform supervised training on the network model.

[0082] It is found that in the prediction of the probability of normal tissue damage, the difficulty in the construction process of the network model lies in the insufficient sample size. Since usually the damage of normal tissues is a gradual process, some damages may not be obvious in the short term after radiotherapy, but will gradually appear in a longer period. For example, radiotherapy (affecting the temporal lobe) leads to the gradual decline of cognitive functions such as memory and language, and this change can only be manifested in the years after radiotherapy. Relying only on short-term detection data may not be able to accurately predict these side effects and the impact brought by normal tissue damage. Therefore, in the embodiments of the present application, the detection data of at least 500 patients in the sixty months after radiotherapy is used as sample data to train the initial network model. The above method can provide richer data features for the network model. The data in the sixty months after radiotherapy can provide rich information in many aspects, including the patient's physiological state, side effects after radiotherapy, tissue repair process, etc. These data features can make the model training more comprehensive and avoid missing key factors that may affect damage prediction. Compared with relying only on short-term data, long-term follow-up data can make the prediction results more accurate and comprehensive. At the same time, the impact of radiotherapy on patients with different individualities (such as age) varies greatly. Long-term data can help identify the unique response patterns of patients to radiotherapy, thus providing more basis for personalized treatment plans.

[0083] It should be noted that for the detection data of tumor patients after radiotherapy, it can include Computed Tomography (CT), Magnetic Resonance Imaging (MRI), blood tests, and so on.

[0084] In the embodiments of the present application, preprocessing of sample data is also performed, including: removing outliers, noise data, and duplicate data to ensure the accuracy and consistency of the data; performing efficient data integration; integrating various types of data from different devices and different time points together, and unifying the data format into the standardized Digital Imaging and Communications in Medicine (DICOM) format and the International System of Units (SI) for subsequent processing; performing reasonable data transformation: selecting appropriate transformation methods according to the characteristics of the data, such as logarithmic transformation, standardization, normalization, etc. For example, logarithmic transformation is performed on some data to reduce the skewness of the data, and standardization is performed on some other data to make its mean 0 and standard deviation 1. To make the data meet the input requirements of the model, improve the stability and convergence speed of the model; performing scientific data reduction: reducing the dimension and complexity of the data through methods such as feature selection or dimensionality reduction, reducing the computational cost, improving the computational efficiency, and avoiding overfitting problems at the same time.

[0085] In addition, to make the results more accurate, based on the multiple cross-validation method, the preprocessed sample data is grouped; the data is divided into a training set, a validation set, and a test set; 70% of the data is randomly selected as the training set, 20% as the validation set, and 10% as the test set. The training set is used for the training and parameter adjustment of the model, and its scale should be large enough to ensure that the model can fully learn the characteristics and laws of the data; the validation set is used to evaluate the performance of the model during the training process and select the best model parameters. Its data should be representative and able to reflect the distribution characteristics of the overall data; the test set is used to finally evaluate the generalization ability of the model, and its data should be independent of the training set and the validation set to truly test the performance of the model on new data; the test set can assist clinicians in predicting the temporal lobe injury caused by radiotherapy dose and observe the stability and accuracy of the model. By randomly dividing the data set multiple times, the result deviation caused by the randomness of data division is reduced, and the reliability and stability of the model are improved.

[0086] For the verification of the model, in the embodiments of the present application, the above-mentioned validation set is used to verify the model through multiple evaluation indicators; the model is verified using multiple evaluation indicators, including but not limited to the evaluation index and the AUC area evaluation index to comprehensively evaluate the performance of the model.

[0087] It reflects the proportion of the variation of the dependent variable that the model can explain. It is a relative measure and can be used to compare with similar models trained on the same data. When is close to 1, it indicates that the model can well explain the variability of the dependent variable and has a high degree of fitting. When is close to 0, it indicates that the model cannot explain the variability of the dependent variable and has a low degree of fitting. Reflects the model's ability to interpret data. The higher the value, the better the model can explain the variation in the data.

[0088] The AUC area is used to evaluate the classification performance of the model. Especially in binary classification problems, the larger the AUC area, the stronger the discrimination ability of the model. For the AUC area, by plotting the Receiver Operating Characteristic (ROC) curve and calculating the area under the curve, a model with a larger AUC area is selected.

[0089] In addition, it should be noted that during the training process of the above four network models, a weight and bias optimization module is designed. The backpropagation algorithm is used in all four network models, that is, starting from the output layer, the partial derivatives of the weights and biases of each layer are calculated according to the loss function, and the parameters are updated according to a certain optimization algorithm until the optimal model configuration is optimized;

[0090] The differences in the optimization algorithms of the above four deep learning-based network models are as follows:

[0091] The convolutional neural network model contains special layers such as convolutional layers and pooling layers; the weight update of the convolutional layer is achieved by sliding the convolutional kernel over the input data to calculate the gradient.

[0092] The deep neural network model is mainly composed of fully connected layers, and the weight update is achieved based on the neuron connection method of the fully connected layer; for a fully connected layer with n input neurons and m output neurons, the weight matrix is an m*n matrix, and the bias is an m-dimensional vector; when updating the weights, it is necessary to consider the contribution of each input neuron to each output neuron, and the weight matrix and bias vector are updated through gradient calculation.

[0093] The core of the residual network model is the residual block, and the weight and bias updates are affected by the skip connections in the residual block; when calculating the gradient, by directly adding the input to the output of the subsequent layer, the gradient can be propagated more directly between different layers, alleviating the gradient vanishing problem in deep networks; enabling the network to more easily learn the identity mapping, that is, the output is the input itself.

[0094] The long short-term memory network model has special gate structures, namely the input gate, forget gate, and output gate; the forget gate determines which information to discard from the cell state, and the updates of its weights and biases comprehensively consider multiple factors such as the current input, the previous hidden state, and the cell state. During the backpropagation process, it is necessary to calculate the partial derivatives of these gate operations according to the loss function, and then update the corresponding weights and biases to learn the long-term dependencies in the sequence data;

[0095] The model parameters are continuously adjusted using different optimization algorithms to achieve the optimal model configuration, enabling the model to accurately fit the relationship between the input data such as the dose and volume data of the patient's radiotherapy and the probability of normal tissue complications.

[0096] Please refer to Figure 6 , Figure 6 which shows the specific training process of the four network models provided by the embodiments of the present application. First, data is collected, that is, sample data, and then the data is preprocessed. After that, 70% of the data is randomly divided as the training set, 20% as the validation set, and 10% as the test set. Four network models are trained using the training set. Then, after the iterative training of the training set is completed, the model is evaluated using the validation set. According to the evaluation results, it is determined whether the optimal model configuration has been reached. If not, the model continues to be iteratively trained; if so, the model is predicted using the test set. Finally, the prediction results are presented.

[0097] Optionally, the tumor diagnosis information includes the tumor type; the normal tissue damage probability prediction model constructed based on deep learning is a joint model group; different joint model groups are pre-constructed according to different tumor types to form a prediction model library.

[0098] Exemplarily, when the tumor type is a brain tumor, a joint model group A is constructed, and this joint model group A includes a prediction model for the brainstem tissue, a prediction model for the temporal lobe tissue, and a prediction model for the salivary gland tissue.

[0099] Exemplarily, when the tumor type is lung cancer, a joint model group B is constructed, and this joint model group B includes a prediction model for the lung tissue, a prediction model for the esophagus tissue, and a prediction model for the heart.

[0100] The joint model group A and the joint model group B in the above examples are both stored in the prediction model library. Different joint model groups can be understood as multi-task prediction model groups separately constructed for different tumor types, and the damage probabilities of normal tissues output by different joint model groups are all associated with the tumor type.

[0101] Specifically, please refer to Figure 7 , where the volume parameters and biological equivalent doses of the patient's normal tissues, as well as the patient's basic information and tumor diagnosis information, are input into the normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissues, including: step 701 to step 702.

[0102] Step 701: Based on the patient's tumor type, determine the target joint model group from the prediction model library.

[0103] According to the foregoing example, assuming that the tumor type of the patient is a brain tumor, the combined model group A is screened out from the prediction model library.

[0104] Step 702: Input the volume parameters and biological equivalent doses of the patient's normal tissues, as well as the patient's basic information and tumor diagnosis information, into the target combined model group, so as to output the damage probabilities of multiple normal tissues associated with the patient's tumor type through the target combined model group.

[0105] Then, input the volume parameters and biological equivalent doses of the patient's normal tissues, as well as the patient's basic information and tumor diagnosis, into the determined target combined model, so as to realize the prediction of the damage probabilities of multiple associated normal tissues of this tumor type.

[0106] Continuing with the above example, the damage probabilities of the brainstem tissue, the temporal lobe tissue, and the salivary gland tissue are output through the combined model group A.

[0107] It can be seen that the embodiment of the present application provides a multi-task prediction method for the tumor type of a patient, which can obtain the prediction results of the damage probabilities of multiple associated normal tissues based on the tumor type of the patient. However, in the traditional method, the damage probability prediction is usually only performed on a certain normal tissue that is easily affected, lacking the evaluation of other normal tissues of tumor patients. Here, through one input, the prediction of the damage probabilities of multiple associated normal tissues can be obtained, which can further assist doctors in more comprehensively understanding the potential impacts after radiotherapy. In summary, through the construction of combined model groups for different tumor types, the combined prediction of the damage probabilities of multiple normal tissues is realized, achieving personalization and high efficiency. It can not only comprehensively evaluate the impacts of radiotherapy on multiple normal tissues, but also help doctors optimize treatment plans, providing better treatment experiences and effects for patients.

[0108] Optionally, the number of normal tissue damage probability prediction models constructed based on deep learning is multiple, and they are network models of different types.

[0109] Please refer to Figure 8 , the above steps of inputting the volume parameters and biological equivalent doses of the patient's normal tissues, as well as the patient's basic information and tumor diagnosis information, into the normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissues include: Step 801 to Step 803.

[0110] Step 801: Input the volume parameters and biological equivalent doses of the patient's normal tissues, as well as the patient's basic information and tumor diagnosis information, into multiple network models, so as to output the damage probability of the patient's normal tissues through each network model.

[0111] In one embodiment, four network models as introduced in the foregoing embodiments can be constructed for the same normal tissue. For example, these four models are all applicable to predicting the damage probability of lung tissue. Then, the volume parameters and biological equivalent doses of the patient's lung tissue, as well as the patient's basic information and tumor diagnosis information, can be input into these four network models, so that each network model outputs the damage probability of the patient's lung tissue.

[0112] Step 802: Obtain the weights of each network model.

[0113] Step 803: Based on the weights of each network model and the damage probability of the patient's normal tissue output by each network model, obtain the combined damage probability.

[0114] Among them, the combined damage probability is the finally obtained damage probability of the patient's normal tissue.

[0115] Here, it can be understood that weights are assigned to the damage probabilities output by each network model, and then the sum is obtained to get the final combined damage probability.

[0116] That is to say, the embodiments of the present application provide a way of combined prediction by combining multiple different types of network models. The damage probabilities of the patient's normal tissue are independently predicted by multiple network models, and then the outputs of different models are fused using weights, integrating the advantages of different network models, thereby improving the accuracy and credibility of the overall prediction. This way can reduce the limitations of the output results caused by a single type of model.

[0117] It should be noted that the weights of each of the above network models can be static and can be preset.

[0118] Of course, in some embodiments, the weights of each of the above network models can also be dynamically adjusted.

[0119] Optionally, the embodiments of the present application provide a way to dynamically adjust the weights. That is, the step of obtaining the weights of each network model can specifically include: determining the age range to which the patient belongs based on the patient's basic information; determining the range of the patient's tumor size based on the patient's tumor diagnosis information; and determining the weights of each network model according to the age range to which the patient belongs, the range of the patient's tumor size, and the model accuracy of each network model in this age range and for this range of tumor sizes.

[0120] Exemplarily, the age ranges can be preset as 20 - 30 years old, 30 - 40 years old, 40 - 50 years old, 50 - 60 years old, and so on. Then determine the age range to which the patient belongs. Suppose the patient is 35 years old, then this patient belongs to the age range of 30 - 40 years old.

[0121] Exemplarily, the size range of the tumor can be preset, and the range is determined according to different tumor types. For example, the conventional settings can be 0-3 cm, 3-5 cm, 5-10 cm, etc. Then, determine the range interval to which the size of the patient's tumor belongs. Suppose the size of the patient's tumor is 7 cm, then it belongs to the 5-10 cm interval.

[0122] After determining the age interval of 30 to 40 years old and the tumor size interval of 5 to 10 cm, according to the model accuracies of the above four network models in this age interval and the range interval of the tumor size, determine the weights of each network model. The determination of the model accuracy is the accuracy of the model prediction in this interval, and the determination of this accuracy can be obtained from the test data of the test set in the foregoing embodiments.

[0123] Initially, the weights of each network model can be set to 0.25. Suppose the convolutional neural network model has the highest accuracy in the range of 30 to 40 years old, and the deep neural network model has the lowest accuracy, then the corresponding weight of the convolutional neural network model can be increased, and the corresponding weight of the deep neural network model can be decreased.

[0124] Also suppose that in the interval of the tumor size of 5 to 10 cm, the accuracies are arranged from high to low as follows: residual network model, long short-term memory network model, deep neural network model, convolutional neural network model. Then, the corresponding weights of the residual network model and the long short-term memory network model can be increased, and the corresponding weights of the deep neural network model and the convolutional neural network model can be decreased. And increase and decrease according to the gradient. For example, the weight of the final residual network model can be 0.4, the weight of the long short-term memory network model can be 0.3, the weight of the deep neural network model can be 0.2, and the weight of the convolutional neural network model can be 0.1 (this is for example and does not consider the age influence).

[0125] It can be seen that the embodiment of the present application provides a more accurate prediction, introducing a mechanism for dynamically adjusting weights, combining the patient's age interval, tumor size range, and the prediction accuracy of each network model under these conditions, dynamically optimizing the weight allocation of each network model, and thus can ensure that under different patient characteristic conditions, the weight allocation matches the model performance, thereby significantly improving the accuracy of the final prediction result.

[0126] Optionally, the embodiment of the present application provides another way to dynamically adjust weights, that is, the above step of obtaining the weights of each network model can specifically include: determining the weights of each network model according to the input data type of the patient's tumor diagnosis information.

[0127] The input data type here can include time series data, image data, etc.

[0128] Exemplarily, since the residual network model is more suitable for feature extraction of complex image data, especially for maintaining feature fidelity in deep networks and is suitable for processing high-dimensional tumor diagnosis information (such as multi-modal image data). Therefore, when the patient's tumor diagnosis information involves image data (such as CT), the weight of the residual network model can be increased.

[0129] Exemplarily, since the long short-term memory network model performs excellently on time series data, when the patient's tumor diagnosis information involves data such as tumor diagnosis progress records, the weight of the long short-term memory network model can be increased.

[0130] In summary, the above method dynamically determines the weight of each network model according to the data type of the tumor diagnosis information, achieving the best match between the model characteristics and data features, and thus can comprehensively improve the accuracy, reliability and adaptability of the prediction. This method not only makes full use of the unique advantages of each network model, but also can dynamically adapt to different tumor types and data scenarios, supporting personalized and multi-modal processing, and has significant clinical application value.

[0131] Please refer to Figure 9 , based on the same inventive concept, an embodiment of the present application provides a radioactive normal tissue damage probability prediction system 900 based on deep learning, including: a first acquisition module 901 for acquiring the patient's tumor dose data, basic information, and tumor diagnosis information; wherein, the tumor dose data is formulated by a doctor according to the patient's tumor diagnosis information before the patient's radiotherapy; the basic information at least includes the patient's age; a second acquisition module 902 for acquiring the dose volume histogram of the patient based on the tumor dose data; a third acquisition module 903 for acquiring the volume parameter and biological equivalent dose of the patient's normal tissue based on the dose volume histogram; wherein, the conversion formula of the biological equivalent dose is:

[0132] ; wherein, represents the biological equivalent dose when the physical dose extracted from the dose volume histogram after conversion is divided by 2 Gy; represents the total physical dose when the fractionation dose in the radiotherapy plan is and the number of fractions is ; represents the cell proliferation correction factor, and the value of is taken as 3 Gy; a prediction module 904 for inputting the volume parameter and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnosis information into a normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissue.

[0133] Please refer to Figure 10 , based on the same inventive concept, an embodiment of the present application provides a module housing of an electronic device 1000 that applies the above method. The electronic device 1000 includes: at least one processor 1001 ( Figure 10 only one is shown in the figure), a memory 1002, and a computer program 1003 stored in the memory 1002 and executable on at least one processor 1001. When the processor 1001 executes the computer program 1003, the steps of the method in any of the foregoing embodiments are implemented.

[0134] The electronic device 1000 may be a server, a personal computer, a laptop computer, etc.

[0135] Those skilled in the art can understand that Figure 10 merely examples of the electronic device 1000, which do not constitute a limitation on the electronic device 1000, and may include more or fewer components than shown in the figure, or combine some components, or different components.

[0136] The so-called processor 1001 may be a central processing unit (CPU), and the processor 1001 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0137] The memory 1002 may be an internal storage unit of the electronic device 1000 in some embodiments, such as the hard disk or memory of the electronic device 1000. The memory 1002 may also be an external storage device of the electronic device 1000 in other embodiments, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 1000. Further, the memory 1002 may also include both the internal storage unit and the external storage device of the electronic device 1000.

[0138] It should be noted that for the above-mentioned systems, devices, etc., since they are based on the same concept as the method embodiments of the present application, the modules designed by the system, as well as the steps executed by the device and the technical effects brought about, can be referred to the method embodiment part, and will not be elaborated here.

[0139] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiment and will not be elaborated here.

[0140] The embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0141] The embodiment of the present application provides a computer program product. When the computer program product runs on a mobile terminal, the mobile terminal can execute the steps in the above-mentioned method embodiments.

[0142] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-described embodiment methods of this application, a computer program can be used to instruct relevant hardware to complete. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the photographing device / electronic device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc, etc.

[0143] In the above embodiments, the descriptions of the respective embodiments have their own focuses. For parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0144] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0145] In the embodiments provided in this application, it should be understood that the disclosed device / network device and method can be implemented in other ways. For example, the device / network device embodiments described above are merely illustrative. For example, the division of the modules or units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in an electrical, mechanical, or other form.

[0146] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or may be distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0147] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for predicting the probability of radiation normal tissue damage based on deep learning, characterized in that: include: Obtaining the patient's tumor dose data, basic information and tumor diagnosis information; wherein the tumor dose data includes the dose data formulated by the doctor based on the patient's tumor diagnosis information before radiotherapy; the basic information includes at least the patient's age; Based on the tumor dose data, obtaining a dose volume histogram of the patient; Based on the dose-volume histogram, the volume parameters and biological equivalent dose of the patient's normal tissue are obtained; wherein the conversion formula of the biological equivalent dose is: ;in, represents the biological equivalent dose when the physical dose extracted from the dose-volume histogram after conversion is 2 Gy as the fractional dose; Indicates that the fractionated dose in the radiotherapy plan is , the number of divisions is The total physical dose at represents the cell proliferation correction factor, The value of is 3Gy; Inputting the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnosis information into a normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissue; Among them, the tumor diagnostic information includes tumor type; the normal tissue damage probability prediction model constructed based on deep learning is a joint model group; different joint model groups are pre-constructed according to different tumor types to form a prediction model library; the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnostic information are input into the normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissue, including: based on the patient's tumor type, determining a target joint model group from the prediction model library; inputting the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnostic information into the target joint model group, so as to output the damage probabilities of multiple normal tissues associated with the patient's tumor type through the target joint model group.

2. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 1, characterized in that: There are multiple normal tissue damage probability prediction models constructed based on deep learning, and they are network models of different types; the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the basic information and tumor diagnosis information of the patient are input into the normal tissue damage probability prediction model constructed based on deep learning to obtain the damage probability of the patient's normal tissue, including: Inputting the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnosis information into a plurality of network models, so that each network model outputs the damage probability of the patient's normal tissue; Get the weights of each network model; Based on the weight of each network model and the damage probability of the normal tissue of the patient output by each network model, a combined damage probability is obtained; wherein the combined damage probability is the final obtained damage probability of the normal tissue of the patient.

3. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 2, characterized in that: The obtaining of the weight of each network model includes: Based on the basic information of the patient, determining the age range to which the patient belongs; Based on the patient's tumor diagnosis information, determining the range interval to which the patient's tumor size belongs; The weight of each network model is determined according to the age range of the patient, the range of the patient's tumor size, and the model accuracy of each network model in the age range and for the range of the tumor size.

4. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 2, characterized in that: The obtaining of the weight of each of the network models comprises: The weight of each network model is determined according to the input data type of the patient's tumor diagnosis information.

5. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 1, characterized in that: The normal tissue damage probability prediction model built based on deep learning is a convolutional neural network model; The structure of the convolutional neural network model includes: multiple convolutional layers, batch normalization layers, dropout layers, maximum pooling layers and fully connected layers.

6. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 1, characterized in that: The normal tissue damage probability prediction model built based on deep learning is a deep neural network model; The structure of the deep neural network model includes: being constructed with multiple fully connected layers as the core; and also including a batch normalization layer and a Dropout layer.

7. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 1, characterized in that: The normal tissue damage probability prediction model built based on deep learning is a residual network model; The structure of the residual network model includes: building with the residual block as the core; and including a convolutional layer, a batch normalization layer, an activation function layer and a fully connected layer.

8. The method for predicting the probability of radiation normal tissue damage based on deep learning according to claim 1, characterized in that: The normal tissue damage probability prediction model built based on deep learning is a long short-term memory network model; The structure of the long short-term memory network model includes: building with multiple LSTM layers as the core; and also including a batch normalization layer and a Dropout layer.

9. A radiation normal tissue damage probability prediction system based on deep learning, characterized in that: include: The first acquisition module is used to acquire the patient's tumor dose data, basic information and tumor diagnosis information; wherein the tumor dose data includes the data set by the doctor based on the patient's tumor diagnosis information before radiotherapy; the basic information at least includes the patient's age; A second acquisition module, configured to acquire a dose volume histogram of the patient based on the tumor dose data; The third acquisition module is used to acquire the volume parameters and biological equivalent dose of the normal tissue of the patient based on the dose volume histogram; wherein the conversion formula of the biological equivalent dose is: ;in, represents the biological equivalent dose when the physical dose extracted from the dose-volume histogram after conversion is 2 Gy as the fractional dose; Indicates that the fractionated dose in the radiotherapy plan is , the number of divisions is The total physical dose at represents cell proliferation correction factor, The value of is 3Gy; A prediction module, used to input the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnosis information into a normal tissue damage probability prediction model constructed based on deep learning, to obtain the damage probability of the patient's normal tissue; Among them, the tumor diagnostic information includes tumor type; the normal tissue damage probability prediction model constructed based on deep learning is a joint model group; different joint model groups are constructed in advance according to different tumor types to form a prediction model library; the prediction module is also specifically used to determine the target joint model group from the prediction model library based on the patient's tumor type; the volume parameters and biological equivalent dose of the patient's normal tissue, as well as the patient's basic information and tumor diagnostic information are input into the target joint model group, so as to output the damage probability of multiple normal tissues associated with the patient's tumor type through the target joint model group.

Images