A method and system for predicting radiation pneumonitis after thoracic tumor radiotherapy based on machine learning

Through a machine learning-based method, combining the patient's CT imaging, clinical and dosimetry data, the improved support vector machine model is trained, which solves the problem of difficult prediction of radiopneumonia after radiotherapy in the prior art, and achieves more accurate and rapid predictive effect.

CN114266729BActive Publication Date: 2024-12-03THE FIRST AFFILIATED HOSPITAL OF XIAMEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111434814.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-29
Publication Date
2024-12-03
Estimated Expiration
2041-11-29

AI Technical Summary

Technical Problem

The prior art has not been able to establish a simple and effective model to comprehensively predict the occurrence of radiopneumonia after thoracic tumor radiotherapy. The main factors include the patient's age, gender, KPS functional status score, whether to suffer from hypertension, whether to chemotherapy, and chemotherapy cycle before radiotherapy Number of tumor target area, tumor location, lung radiation dose, etc.

Method used

Using a machine learning-based approach, we can obtain patients' lung CT imaging data, clinical data and radiotherapy plan dosimetry data, and train an improved support vector machine model through feature loss value prediction and redundant feature removal, using Gaussian kernel function and soft interval, combined with multiple factors to predict.

Benefits of technology

It achieves rapid and accurate prediction of the occurrence of radiopneumonia in patients with chest tumors after radiotherapy, which is more accurate than single-factor prediction. It can combine multiple parameters to improve the accuracy and efficiency of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114266729B_ABST
    Figure CN114266729B_ABST
Patent Text Reader

Abstract

The present invention provides a method for predicting radiation pneumonitis after radiotherapy of thoracic tumors based on machine learning, including: obtaining patient data; digitizing non-digital features in the patient data, and predicting feature missing values ​​and removing redundant features to obtain pre-processed patient data; dividing the pre-processed patient data into a training set and a test set, and training to generate an improved support vector machine model; using the test set to calculate the area under the curve AUC of the operating characteristic curve ROC curve to evaluate the accuracy of the model; inputting the patient-related data to be predicted to verify the model, and predicting the occurrence of radiation pneumonitis. The method provided by the present invention can combine a large number of medical case databases, and combine multiple factors such as the patient's clinical information, dose information, and CT imaging genomics to more quickly and intuitively predict the occurrence of radiation pneumonitis after radiotherapy of thoracic tumor patients.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a prediction method for radiation pneumonia after chest tumor radiotherapy based on machine learning. Background Art

[0002] Radiation pneumonitis is a part of radiation lung injury and is the most important and serious complication of radiotherapy for chest tumors. It is common within 1 to 3 months after the start of radiotherapy. 10%-30% of chest radiotherapy patients will experience radiation pneumonitis, which not only limits the implementation of treatment and affects the efficacy of treatment, but also reduces the patient's quality of life and even survival rate. Therefore, reducing the incidence of radiation pneumonitis is of great clinical significance. Accurate prediction of the occurrence of radiation pneumonitis is conducive to timely clinical intervention and reducing the risk of radiotherapy for patients.

[0003] Current studies have found that the occurrence of acute radiation pneumonitis is related to many factors, such as: patient age, gender, KPS functional status score, hypertension, chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, lymph node location, average lung dose, lung volume receiving 5Gy, 10Gy, 20Gy, 30Gy doses, daily patient dose, and total course dose. In clinical practice, the volume of the lungs receiving 5Gy and 20Gy doses (V5Gy, V20Gy) and the average lung dose in the radiotherapy plan are often used to evaluate the occurrence of radiation pneumonitis. However, a simple and effective model that can comprehensively consider various factors to predict the occurrence of radiation pneumonitis has not yet been established. Summary of the invention

[0004] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art and propose a method for predicting radiation pneumonia after radiotherapy for chest tumors based on machine learning. The method can combine a large number of medical case databases, the patient's clinical information, dose information, CT imaging genomics and other factors to more quickly and intuitively predict the occurrence of radiation pneumonia in patients with chest tumors after radiotherapy.

[0005] The present invention adopts the following technical solution:

[0006] A method for predicting radiation pneumonitis after thoracic tumor radiotherapy based on machine learning, comprising:

[0007] Acquiring patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data;

[0008] Digitize the non-numeric features in the patient data, predict the missing values ​​of the features and remove the redundant features to obtain the preprocessed patient data;

[0009] The preprocessed patient data is divided into a training set and a test set, and an improved support vector machine model is generated through training;

[0010] The area under the ROC curve (AUC) is calculated using the test set to evaluate the accuracy of the model. If the area under the curve is greater than or equal to 0.9, the model is verified and the output is a verified model. If the area under the curve is less than 0.9, the support vector machine model is regenerated.

[0011] The patient-related data that needs to be predicted is input into the model for verification and prediction of the occurrence of radiation pneumonitis.

[0012] Specifically, the patient data includes the patient's lung CT imaging data, the patient's clinical data, and the patient's radiotherapy plan dosimetry data, specifically:

[0013] The patient's lung CT radiomics data include: the minimum value of the first-order texture feature under LLH wavelet transform, the minimum value of the first-order texture feature and the autocorrelation of the gray-level co-occurrence matrix under HHL wavelet transform, and the autocorrelation of the gray-level co-occurrence matrix under HHH wavelet transform;

[0014] The clinical data of patients included: age, gender, KPS functional status score, whether they had hypertension, whether they had chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, and lymph node location of patients undergoing radiotherapy for chest tumors;

[0015] The dosimetric data of the patient's radiotherapy plan include: the average lung dose, the volume of the lung receiving 5Gy, 10Gy, 20Gy, and 30Gy doses, the patient's daily radiation dose, and the total treatment course radiation dose.

[0016] Specifically, feature missing value prediction and redundant feature removal are performed, specifically:

[0017] The K nearest neighbor algorithm was used to fit the non-missing patient data and predict the missing values ​​of features;

[0018] According to the variance maximization principle, a set of linearly independent and mutually orthogonal new vector groups are used to represent the rows / columns of the original data matrix to compress the number of features and eliminate redundant features.

[0019] Specifically, the training generates an improved support vector machine model, including:

[0020] Training sample set D = {(x1, y1), (x2, y2), ..., (x m ,y m )},y i ∈{-1,+1}; m is the number of features;

[0021] Introducing "soft margin" to allow some samples not to meet the constraints

[0022] y i (ω T x i +b)≥1

[0023] The optimization objective can be written as

[0024]

[0025] sty i (ω T x i +b)≥1-i

[0026] i≥0, i=1,2,…,m.

[0027] That is the improved support vector machine model;

[0028] Gaussian kernel function in the model:

[0029]

[0030] Among them, ω and b are model parameters, C is called the penalty coefficient, γ is the width coefficient, i is the slack variable, i≥0.

[0031] The embodiment of the present invention also provides a system for predicting radiation pneumonia after thoracic tumor radiotherapy based on machine learning, comprising:

[0032] Acquiring patient data module: acquiring patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data;

[0033] Data preprocessing module: digitizes non-digital features in patient data, predicts missing values ​​of features and removes redundant features to obtain preprocessed patient data;

[0034] Model training module: divide the preprocessed patient data into training set and test set, and train and generate an improved support vector machine model;

[0035] Model evaluation module: Use the test set to calculate the area under the curve (AUC) of the operating characteristic curve (ROC) to evaluate the accuracy of the model. If the area under the curve is greater than or equal to 0.9, the verification is passed and the model is output as having passed the verification. If the area under the curve is less than 0.9, the support vector machine model is regenerated.

[0036] Prediction module: The patient-related data to be predicted is input into the model for verification and prediction of the occurrence of radiation pneumonia.

[0037] Specifically, the patient data includes the patient's lung CT imaging data, the patient's clinical data, and the patient's radiotherapy plan dosimetry data, specifically:

[0038] The patient's lung CT radiomics data include: the minimum value of the first-order texture feature under LLH wavelet transform, the minimum value of the first-order texture feature and the autocorrelation of the gray-level co-occurrence matrix under HHL wavelet transform, and the autocorrelation of the gray-level co-occurrence matrix under HHH wavelet transform;

[0039] The clinical data of patients included: age, gender, KPS functional status score, whether they had hypertension, whether they had chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, and lymph node location of patients undergoing radiotherapy for chest tumors;

[0040] The dosimetric data of the patient's radiotherapy plan include: the average lung dose, the volume of the lung receiving 5Gy, 10Gy, 20Gy, and 30Gy doses, the patient's daily radiation dose, and the total treatment course radiation dose.

[0041] Specifically, in the data preprocessing module, feature missing value prediction and redundant feature removal are performed, specifically:

[0042] The K nearest neighbor algorithm was used to fit the non-missing patient data and predict the missing values ​​of features;

[0043] According to the variance maximization principle, a set of linearly independent and mutually orthogonal new vector groups are used to represent the rows / columns of the original data matrix to compress the number of features and eliminate redundant features.

[0044] Specifically, in the model training module, training and generating an improved support vector machine model specifically includes:

[0045] Training sample set D = {(x1, y1), (x2, y2), ..., (x m ,y m )}, i ∈{-1,+1}; m is the number of features;

[0046] Introducing "soft margin" to allow some samples not to meet the constraints

[0047] y i (ω T x i +b)≥1

[0048] The optimization objective can be written as

[0049]

[0050] sty i (ω T x i +b)≥1-i

[0051] i≥0,i=1,2,…,m.

[0052] That is the improved support vector machine model;

[0053] Gaussian kernel function in the model:

[0054]

[0055] Among them, ω and b are model parameters, C is called the penalty coefficient, γ is the width coefficient, i is the slack variable, i≥0.

[0056] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) The method for predicting radiation pneumonitis after radiotherapy of thoracic tumors based on machine learning provided by the present invention first obtains patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data; the non-digital features in the patient data are digitized, and feature missing value prediction and redundant feature removal are performed to obtain pre-processed patient data; the pre-processed patient data are divided into a training set and a test set, and an improved support vector machine model is generated by training; the area under the curve (AUC) of the operating characteristic curve (ROC) is calculated using the test set to evaluate the accuracy of the model, and if the area under the curve is greater than or equal to 0.9, the verification is passed, and the verification model is output; if the area under the curve is less than 0.9, the support vector machine model is regenerated; the patient-related data to be predicted is input into the verification model to predict the occurrence of radiation pneumonitis; the method provided by the present invention can combine a large number of medical case databases, and combine multiple factors such as the patient's clinical information, dose information, and CT imaging data to more quickly and intuitively predict the occurrence of radiation pneumonitis after radiotherapy of thoracic tumor patients. Compared with the previous single-factor prediction and dosimetry prediction, machine learning technology simultaneously inputs multiple parameters, and the prediction of radiation pneumonitis is more accurate.

[0058] (2) The method for predicting radiation pneumonia after radiotherapy for thoracic tumors based on machine learning provided by the present invention adopts an improved support vector machine model, in which a "soft interval" is introduced to achieve fast and effective prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flow chart of a method for predicting radiation pneumonitis after thoracic tumor radiotherapy based on machine learning provided in an embodiment of the present invention;

[0060] Figure 2 A structural diagram of a system for predicting radiation pneumonitis after chest tumor radiotherapy based on machine learning provided in an embodiment of the present invention.

[0061] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0062] Radiomics features combine the shape, intensity, and texture features of the original lesion image and the image transformed by various filters such as wavelet and Laplacian of Gaussian. Combining feature selection methods and machine learning algorithms, a prediction model can be built on the training dataset and further evaluated on the test dataset.

[0063] Machine learning is the ability of machines to learn and predict future events and outcomes based on large data sets. In healthcare, machine learning aims to improve the interpretation of medical data, thereby speeding up workflows, reducing errors, eliminating unnecessary expenses, and improving human health. As risk factors for the development of radiation pneumonitis, patient imaging features, clinical and treatment parameters, and dosimetric parameters are intertwined, and it is not possible to link radiation pneumonitis to a single parameter. The prediction of radiation pneumonitis using machine learning methods can help reduce patient toxicity, improve the patient's quality of life, and also reduce the patient's medical expenses.

[0064] like Figure 1 , is a flow chart of a method for predicting radiation pneumonia after chest tumor radiotherapy based on machine learning provided by an embodiment of the present invention, specifically comprising:

[0065] S1: Acquire patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data;

[0066] Specifically, the patient data includes the patient's lung CT imaging data, the patient's clinical data, and the patient's radiotherapy plan dosimetry data, specifically:

[0067] The patient's lung CT radiomics data include: the minimum value of the first-order texture feature under LLH wavelet transform, the minimum value of the first-order texture feature and the autocorrelation of the gray-level co-occurrence matrix under HHL wavelet transform, and the autocorrelation of the gray-level co-occurrence matrix under HHH wavelet transform;

[0068] The clinical data of patients included: age, gender, KPS functional status score, whether they had hypertension, whether they had chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, and lymph node location of patients undergoing radiotherapy for chest tumors;

[0069] The dosimetric data of the patient's radiotherapy plan include: the average lung dose, the volume of the lung receiving 5Gy, 10Gy, 20Gy, and 30Gy doses, the patient's daily radiation dose, and the total treatment course radiation dose.

[0070] In addition, when there are new patients, the corresponding data will be updated to the patient data, increasing the sample size and improving the prediction accuracy.

[0071] S2: Digitize the non-numeric features in the patient data, predict the missing values ​​of the features and remove the redundant features to obtain the preprocessed patient data;

[0072] Digitize the non-numeric features in the patient data. For example, the lymph node locations are divided into mediastinum, hilum, and supraclavicular. In preprocessing, the lymph node locations are defined as numbers: 1 - mediastinum, 2 - hilum, 3 - supraclavicular. The same process is performed on other non-numeric features.

[0073] Specifically, feature missing value prediction and redundant feature removal are performed, specifically:

[0074] The K nearest neighbor algorithm was used to fit the non-missing patient data and predict the missing values ​​of features;

[0075] According to the variance maximization principle, a set of linearly independent and mutually orthogonal new vector groups are used to represent the rows / columns of the original data matrix to compress the number of features and eliminate redundant features.

[0076] S3: Divide the preprocessed patient data into a training set and a test set, and train and generate an improved support vector machine model;

[0077] 70% of the data is randomly selected as model training data, and the remaining 30% of the data is used as model testing data. The training data is used to generate a support vector machine model.

[0078] Given a training sample set D = {(x1,y1),(x2,y2),…,(x m ,y m )},y i ∈{-1,+1}, find a partitioning hyperplane in the sample space based on the training set D to separate different categories. For some training samples, it is impossible to classify them linearly. The samples can be mapped from the original space to a higher-dimensional feature space so that the samples can be linearly separable in this feature space.

[0079] Represents the feature vector after mapping x, so the model corresponding to the hyperplane divided in the feature space can be expressed as

[0080]

[0081] Where ω and b are model parameters. In order to maximize the distance from any point x in the sample space to the hyperplane (ω, b), that is, to know the partitioning hyperplane with the "maximum interval", that is,

[0082]

[0083] sty i (ω T φ(xi )+b)≥1, i=1,2,…,m.

[0084] Its dual problem is

[0085]

[0086]

[0087] α i ≥0, i=1,2,…,m.

[0088] Where φ(x i ) T φ(x j ) is the sample x i With x j The inner product after mapping to the feature space. Since the dimension of the feature space may be very high, it is easier to directly calculate φ(x i ) T φ(x j ) is difficult. We can define the kernel function:

[0089] κ(x i ,x j )=<φ(x i ),φ(x j )>=φ(x i ) T φ(x j )

[0090] The dual problem can be rewritten as

[0091]

[0092]

[0093] α i ≥0, i=1,2,…,m.

[0094] After solving, we get

[0095]

[0096] κ(·,·) is the “kernel function” (Kernel), and the above formula is called the “support vector expansion”.

[0097] In real-world problems, it is often difficult to determine a suitable kernel function that makes the training samples linearly separable in the feature space. In order to avoid overfitting of the model, the support vector machine is allowed to make mistakes on some samples and a "soft margin" is introduced to allow some samples not to meet the constraints.

[0098] y i (ω T x i+b)≥1

[0099] While maximizing the margin, the number of samples that do not satisfy the constraints should be as small as possible. Therefore, the optimization objective can be written as

[0100]

[0101] Where C>0 is a constant, called the penalty coefficient. When C takes a finite value, the above formula allows some samples to not meet the constraints. 0 / 1 is the "0 / 1 loss function":

[0102]

[0103] However, due to l 0 / 1 Non-convex, non-continuous, poor mathematical properties, usually other functions are used to replace l 0 / 1 , called "alternative loss". This model uses hinge loss: hinge (z)=max(0,1-z), the optimization objective can be written as

[0104]

[0105] Introducing the "slack variable" i≥0, each sample has a corresponding slack variable to characterize the degree to which each sample does not satisfy the constraints. The above formula can be rewritten as

[0106]

[0107] sty i (ω T x i +b)≥1-i

[0108] i≥0, i=1,2,…,m.

[0109] This is the "improved support vector machine". The dual problem of the above equation can be obtained by Lagrange multiplier method.

[0110]

[0111]

[0112] 0≤α i ≤C, i=1,2,…,m.

[0113] The choice of kernel function mentioned above is related to the type of hyperplane. The Gaussian kernel function (RBF function) suitable for the case of linear inseparability is selected in the model:

[0114]

[0115] The width coefficient γ defines the influence range of a single sample. The larger the γ, the more support vectors there are.

[0116] The penalty coefficient (regularization parameter) C mentioned above defines the tolerance for "foul" samples.

[0117] Using the data in the training set, grid search cross validation GridSearchCV is used, and the value of C ranges from 10 -4 to 10 10 , and pick 15 numbers in the middle. The value range of γ is from 10 -4 to 10 ^3 , take 24 numbers in the middle. Traverse all given parameter combinations of γ and C, train the data to find the best matching values ​​of parameters γ and C to achieve the highest prediction accuracy. And generate a support vector machine model.

[0118] S4: Use the test set to calculate the area under the curve (AUC) of the operating characteristic curve (ROC) to evaluate the accuracy of the model. If the area under the curve is greater than or equal to 0.9, the verification is passed and the model is output as having passed the verification. If the area under the curve is less than 0.9, the support vector machine model is regenerated.

[0119] S5: The patient-related data to be predicted is input into the model for verification and prediction of the occurrence of radiation pneumonitis.

[0120] According to the prediction results, for patients whose prediction results are "occurrence", the characteristic parameters can be adjusted in reverse, such as reducing the prescribed dose of radiotherapy, modifying the radiotherapy plan, strictly controlling the lung irradiation dose, adjusting the number of chemotherapy cycles, etc. The new characteristic parameters are then input into the output model to predict the occurrence of radiation pneumonitis.

[0121] like Figure 2 Another embodiment of the present invention further provides a system for predicting radiation pneumonia after thoracic tumor radiotherapy based on machine learning, comprising:

[0122] Acquiring patient data module 201: Acquiring patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data;

[0123] Specifically, the patient data includes the patient's lung CT imaging data, the patient's clinical data, and the patient's radiotherapy plan dosimetry data, specifically:

[0124] The patient's lung CT radiomics data include: the minimum value of the first-order texture feature under LLH wavelet transform, the minimum value of the first-order texture feature and the autocorrelation of the gray-level co-occurrence matrix under HHL wavelet transform, and the autocorrelation of the gray-level co-occurrence matrix under HHH wavelet transform;

[0125] The clinical data of patients included: age, gender, KPS functional status score, whether they had hypertension, whether they had chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, and lymph node location of patients undergoing radiotherapy for chest tumors;

[0126] The dosimetric data of the patient's radiotherapy plan include: the average lung dose, the volume of the lung receiving 5Gy, 10Gy, 20Gy, and 30Gy doses, the patient's daily radiation dose, and the total treatment course radiation dose.

[0127] In addition, when there are new patients, the corresponding data will be updated to the patient data, increasing the sample size and improving the prediction accuracy.

[0128] Data preprocessing module 202: digitizes non-digital features in the patient data, predicts feature missing values ​​and removes redundant features to obtain preprocessed patient data;

[0129] Digitize the non-numeric features in the patient data. For example, the lymph node locations are divided into mediastinum, hilum, and supraclavicular. In preprocessing, the lymph node locations are defined as numbers: 1 - mediastinum, 2 - hilum, 3 - supraclavicular. The same process is performed on other non-numeric features.

[0130] Specifically, feature missing value prediction and redundant feature removal are performed, specifically:

[0131] The K nearest neighbor algorithm was used to fit the non-missing patient data and predict the missing values ​​of features;

[0132] According to the variance maximization principle, a set of linearly independent and mutually orthogonal new vector groups are used to represent the rows / columns of the original data matrix to compress the number of features and eliminate redundant features.

[0133] Model training module 203: dividing the pre-processed patient data into a training set and a test set, and training and generating an improved support vector machine model;

[0134] 70% of the data is randomly selected as model training data, and the remaining 30% of the data is used as model testing data. The training data is used to generate a support vector machine model.

[0135] Given a training sample set D = {(x1, y1), (x2, y2), ..., (x m ,y m )},y i ∈{-1,+1}, find a partitioning hyperplane in the sample space based on the training set D to separate different categories. For some training samples, it is impossible to classify them linearly. The samples can be mapped from the original space to a higher-dimensional feature space so that the samples can be linearly separable in this feature space.

[0136] Represents the feature vector after mapping x, so the model corresponding to the hyperplane divided in the feature space can be expressed as

[0137]

[0138] Where ω and b are model parameters. In order to maximize the distance from any point x in the sample space to the hyperplane (ω, b), that is, to know the partitioning hyperplane with the "maximum interval", that is,

[0139]

[0140] sty i (ω T φ(x i )+b)≥1, i=1,2,…,m.

[0141] Its dual problem is

[0142]

[0143]

[0144] α i ≥0, i=1,2,…,m.

[0145] Where φ(x i ) T φ(x j ) is the sample x i With x j The inner product after mapping to the feature space. Since the dimension of the feature space may be very high, it is easier to directly calculate φ(x i ) T φ(x j ) is difficult. We can define the kernel function:

[0146] κ(x i ,x j )=<φ(x i ),φ(x j )>=φ(x i ) T φ(x j )

[0147] The dual problem can be rewritten as

[0148]

[0149]

[0150] α i ≥0, i=1,2,…,m.

[0151] After solving, we get

[0152]

[0153] κ(·,·) is the “kernel function” (Kernel), and the above formula is called the “support vector expansion”.

[0154] In real-world problems, it is often difficult to determine a suitable kernel function that makes the training samples linearly separable in the feature space. In order to avoid overfitting of the model, the support vector machine is allowed to make mistakes on some samples and a "soft margin" is introduced to allow some samples not to meet the constraints.

[0155] y i (ω T x i +b)≥1

[0156] While maximizing the margin, the number of samples that do not satisfy the constraints should be as small as possible. Therefore, the optimization objective can be written as

[0157]

[0158] Where C>0 is a constant, called the penalty coefficient. When C takes a finite value, the above formula allows some samples to not meet the constraints. 0 / 1 is the "0 / 1 loss function":

[0159]

[0160] However, due to l 0 / 1 Non-convex, non-continuous, poor mathematical properties, usually other functions are used to replace l 0 / 1 , called "alternative loss". This model uses hinge loss: hinge (z)=max(0,1-z), the optimization objective can be written as

[0161]

[0162] Introducing the "slack variable" i≥0, each sample has a corresponding slack variable to characterize the degree to which each sample does not satisfy the constraints. The above formula can be rewritten as

[0163]

[0164] sty i (ω T x i +b)≥1-i

[0165] i≥0, i=1,2,…,m.

[0166] This is the "improved support vector machine". The dual problem of the above equation can be obtained by Lagrange multiplier method.

[0167]

[0168]

[0169] 0≤α i ≤C, i=1,2,…,m.

[0170] The choice of kernel function mentioned above is related to the type of hyperplane. The Gaussian kernel function (RBF function) suitable for the case of linear inseparability is selected in the model:

[0171]

[0172] The width coefficient γ defines the influence range of a single sample. The larger the γ, the more support vectors there are.

[0173] The penalty coefficient (regularization parameter) C mentioned above defines the tolerance for "foul" samples.

[0174] Using the data in the training set, grid search cross validation GridSearchCV is used, and the value of C ranges from 10 -4 to 10 10 , and pick 15 numbers in the middle. The value range of γ is from 10 -4 to 10 ^3 , take 24 numbers in the middle. Traverse all given parameter combinations of γ and C, train the data to find the best matching values ​​of parameters γ and C to achieve the highest prediction accuracy. And generate a support vector machine model.

[0175] Model evaluation module 204: using the test set to calculate the area under the curve (AUC) of the operating characteristic curve (ROC) to evaluate the accuracy of the model. If the area under the curve is greater than or equal to 0.9, the model is verified to be passed and the verified model is output; if the area under the curve is less than 0.9, the support vector machine model is regenerated;

[0176] Prediction module 205: inputs the patient-related data to be predicted into the model for verification, and predicts the occurrence of radiation pneumonia.

[0177] According to the prediction results, for patients whose prediction results are "occurrence", the characteristic parameters can be adjusted in reverse, such as reducing the prescribed dose of radiotherapy, modifying the radiotherapy plan, strictly controlling the lung irradiation dose, adjusting the number of chemotherapy cycles, etc. The new characteristic parameters are then input into the output model to predict the occurrence of radiation pneumonitis.

[0178] The method for predicting radiation pneumonia after radiotherapy of thoracic tumors based on machine learning provided by the present invention first obtains patient data, and the patient data includes patient lung CT imaging group data, patient clinical data, and patient radiotherapy plan dosimetry data; the non-digital features in the patient data are digitized, and feature missing value prediction and redundant feature removal are performed to obtain pre-processed patient data; the pre-processed patient data are divided into a training set and a test set, and the improved support vector machine model is generated by training; the area under the curve AUC of the operating characteristic curve ROC curve is calculated using the test set to evaluate the accuracy of the model, if the area under the curve is greater than or equal to 0.9, the verification is passed, and the verification model is output; if the area under the curve is less than 0.9, the support vector machine model is regenerated; the patient-related data to be predicted is input into the verification model to predict the occurrence of radiation pneumonia; the method provided by the present invention can be combined with a large number of medical case databases, combined with the patient's clinical information, dose information, CT imaging group and other factors, to more quickly and intuitively predict the occurrence of radiation pneumonia after radiotherapy of thoracic tumor patients. Compared with the previous single-factor prediction and dosimetry prediction, machine learning technology inputs multiple parameters at the same time, and the prediction of radiation pneumonia is more accurate.

[0179] The method for predicting radiation pneumonia after chest tumor radiotherapy based on machine learning provided by the present invention adopts an improved support vector machine model, and a "soft interval" is introduced into the model to achieve fast and effective prediction.

[0180] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited to this. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A method for predicting radiation pneumonitis after thoracic tumor radiotherapy based on machine learning, characterized in that: include: Acquiring patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data; Digitize the non-numeric features in the patient data, predict the missing values ​​of the features and remove the redundant features to obtain the preprocessed patient data; The preprocessed patient data is divided into a training set and a test set, and an improved support vector machine model is generated through training; The training generates an improved support vector machine model, specifically comprising: Training sample set D = {(x1, y1), (x2, y2), ..., (x m ,y m )),y i ∈{-1, +1}; m is the number of features; Introducing "soft margin" to allow some samples not to meet the constraints y i (ω T x i +b)≥1 The optimization objective can be written as s.t.y i (ω T x i +b)≥1-ε i e i ≥0, i=1,2,…,m. That is the improved support vector machine model; Gaussian kernel function in the model: Among them, ω and b are model parameters, C is called the penalty coefficient, γ is the width coefficient, ε i is the slack variable, ε i ≥0; The area under the ROC curve (AUC) is calculated using the test set to evaluate the accuracy of the model. If the area under the curve is greater than or equal to 0.9, the model is verified and the output is a verified model. If the area under the curve is less than 0.9, the support vector machine model is regenerated. The patient-related data that needs to be predicted is input into the model for verification and prediction of the occurrence of radiation pneumonitis.

2. The method for predicting radiation pneumonitis after thoracic tumor radiotherapy based on machine learning according to claim 1, characterized in that: The patient data includes the patient's lung CT imaging data, the patient's clinical data, and the patient's radiotherapy plan dosimetry data, specifically: The patient's lung CT radiomics data include: the minimum value of the first-order texture feature under LLH wavelet transform, the minimum value of the first-order texture feature and the autocorrelation of the gray-level co-occurrence matrix under HHL wavelet transform, and the autocorrelation of the gray-level co-occurrence matrix under HHH wavelet transform; The clinical data of patients included: age, gender, KPS functional status score, whether they had hypertension, whether they had chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, and lymph node location of patients undergoing radiotherapy for chest tumors; The dosimetric data of the patient's radiotherapy plan included the average lung dose, the volume of the lung receiving 5 Gy, 10 Gy, 20 Gy, and 30 Gy doses, the patient's daily radiation dose, and the total radiation dose during the treatment.

3. The method for predicting radiation pneumonitis after thoracic tumor radiotherapy based on machine learning according to claim 1, characterized in that: And perform feature missing value prediction and redundant feature removal, specifically: The K nearest neighbor algorithm was used to fit the non-missing patient data and predict the missing values ​​of features; According to the variance maximization principle, a set of linearly independent and mutually orthogonal new vector groups are used to represent the rows / columns of the original data matrix to compress the number of features and eliminate redundant features.

4. A machine learning-based radiation pneumonitis prediction system for chest tumors after radiotherapy, characterized in that: include: Acquiring patient data module: acquiring patient data, wherein the patient data includes patient lung CT imaging data, patient clinical data, and patient radiotherapy plan dosimetry data; Data preprocessing module: digitizes non-digital features in patient data, predicts missing values ​​of features and removes redundant features to obtain preprocessed patient data; Model training module: divide the preprocessed patient data into training set and test set, and train and generate an improved support vector machine model; The training generates an improved support vector machine model, specifically comprising: Training sample set D = {(x1, y1), (x2, y2), ..., (x m ,y m )},y i ∈{-1, +1}; m is the number of features; Introducing "soft margin" to allow some samples not to meet the constraints y i (ω T x i +b)≥1 The optimization objective can be written as s.t.y i (ω T x i +b)≥1-ε i e i ≥0, i=1, 2,…, m. That is the improved support vector machine model; Gaussian kernel function in the model: Among them, ω and b are model parameters, C is called the penalty coefficient, γ is the width coefficient, ε i is the slack variable, ε i ≥0; Model evaluation module: Use the test set to calculate the area under the curve (AUC) of the operating characteristic curve (ROC) to evaluate the accuracy of the model. If the area under the curve is greater than or equal to 0.9, the verification is passed and the model is output as having passed the verification. If the area under the curve is less than 0.9, the support vector machine model is regenerated. Prediction module: The patient-related data to be predicted is input into the model for verification and prediction of the occurrence of radiation pneumonia.

5. The machine learning-based radiation pneumonitis prediction system after thoracic tumor radiotherapy according to claim 4, characterized in that: The patient data includes the patient's lung CT imaging data, the patient's clinical data, and the patient's radiotherapy plan dosimetry data, specifically: The patient's lung CT radiomics data include: the minimum value of the first-order texture feature under LLH wavelet transform, the minimum value of the first-order texture feature and the autocorrelation of the gray-level co-occurrence matrix under HHL wavelet transform, and the autocorrelation of the gray-level co-occurrence matrix under HHH wavelet transform; The clinical data of patients included: age, gender, KPS functional status score, whether they had hypertension, whether they had chemotherapy, number of chemotherapy cycles before radiotherapy, tumor target volume, tumor location, and lymph node location of patients undergoing radiotherapy for chest tumors; The dosimetric data of the patient's radiotherapy plan include: the average lung dose, the volume of the lung receiving 5Gy, 10Gy, 20Gy, and 30Gy doses, the patient's daily radiation dose, and the total treatment course radiation dose.

6. The machine learning-based radiation pneumonitis prediction system after thoracic tumor radiotherapy according to claim 4, characterized in that: In the data preprocessing module, feature missing value prediction and redundant feature removal are performed, specifically: The K nearest neighbor algorithm was used to fit the non-missing patient data and predict the missing values ​​of features; According to the variance maximization principle, a set of linearly independent and mutually orthogonal new vector groups are used to represent the rows / columns of the original data matrix to compress the number of features and eliminate redundant features.

7. The machine learning-based radiation pneumonitis prediction system after thoracic tumor radiotherapy according to claim 4, characterized in that: In the model training module, training and generating an improved support vector machine model specifically includes: Training sample set D = {(x1, y1), (x2, y2), ..., (x m ,y m )},y i ∈{-1, +1}; m is the number of features; Introducing "soft margin" to allow some samples not to meet the constraints y i (ω T x i +b)≥1 The optimization objective can be written as s.t.y i (ω T x i +b)≥1-ε i e i ≥0, i=1,2,…,m. That is the improved support vector machine model; Gaussian kernel function in the model: Among them, ω and b are model parameters, C is called the penalty coefficient, γ is the width coefficient, ε i is the slack variable, ε i ≥0.