Tumor radiotherapy reaction prediction system and method based on rule learning

Abstract

The invention discloses a tumor radiotherapy reaction prediction system and method based on rule learning. The prediction method comprises the steps of data and image acquisition, tumor risk area layering, medical feature extraction, rule learning prediction model construction and the like. Firstly, a random forest is constructed according to feature data extracted from FDG PET / CT images of patients in a training set, and rules are generated; secondly, performing rule pruning by adopting a grey wolf optimizer (GWO) algorithm, and selecting a better rule; and finally, performing rule fitting, and training a linear model to predict the radiotherapy reaction of the patient in the test set. The constructed prediction model can calculate the support degree and importance of each rule in the model, so that the clinical interpretability of the model is provided; the model provided by the invention has relatively good prediction performance and interpretable radiotherapy reaction prediction, so that doctors can be better assisted in radiotherapy clinical decision support for lung cancer patients.

Description

Technical Field

[0001] The present invention relates to the field of medical rule learning, and more specifically, to a tumor radiotherapy response prediction system and method based on rule learning. Background Art

[0002] Radiotherapy, as a key component of current cancer treatment, is less invasive and has fewer side effects than traditional surgery and chemotherapy, and can maximize tumor control when used in combination with other treatments. However, current standard treatment options for cancer patients have problems such as poor prognosis and limited tumor control. Early response assessment using fluorodeoxyglucose (FDG) positron emission tomography (PET) / computed tomography (CT) during radiotherapy has been shown to play an important role in the treatment of non-small cell lung cancer. PET / CT can quantify changes in tumors before and after treatment before morphological changes, thereby tailoring the best treatment strategy for patients. The tumor standard uptake value (SUV) has important prognostic value in the treatment of cancer patients and can provide a reference for the selection of clinical treatment options and prognosis judgment.

[0003] Scientific research has shown that feature data can be extracted from FDG PET / CT to more accurately evaluate and predict the outcomes of different cancer patients at different stages of treatment. Features including tumor volume and SUV of the entire tumor volume have been shown to be significantly correlated with treatment efficacy. Related studies have also shown that FDG PET / CT parameters of tumors, combined with tumor volume and metabolic activity of tumors, can have good predictive performance. Radiomic features extracted by radiomics-based image analysis techniques have been used to provide accurate and personalized treatment strategies. Radiomic features include FDG PET / CT image features of tissues surrounding tumors, such as shape and texture features, which have been used to predict the effect of radiotherapy for locally advanced non-small cell lung cancer. Current research work is dedicated to extracting and evaluating radiomics features to predict the clinical prognosis of patients.

[0004] Machine learning models, such as support vector machines (SVM) and decision trees (DT), have been widely used to predict the clinical prognosis of patients. Ensemble models such as random forests (RF) and gradient boosting decision trees (GBDT) are able to process complex feature data sets and improve prediction accuracy. However, the complexity of the model makes it a black box that lacks clinical interpretability, which poses a major challenge to the translation of scientific research into clinical applications. Random forest is an ensemble learning model based on bagging that improves the decision-making ability of the model by integrating multiple decision trees. Random forests are composed of decision trees, where each decision tree model stores a set of information for each node in the form of an array. Based on this information, the RuleFit algorithm extracts rules from each decision tree and constructs an interpretable machine learning model for prediction through rule fitting. The quality of the selected rules will affect the prediction accuracy of the constructed model, and a method for efficiently screening high-quality rules has not yet been developed. On the other hand, existing models rarely predict the SUV change value of tumor radiotherapy response, and a model with high accuracy in predicting SUV change value has not yet been constructed based on rules.

[0005] Therefore, developing a rule-based tumor radiotherapy response prediction method that has both predictive accuracy and clinical interpretability is a technical problem that needs to be solved urgently. Summary of the invention

[0006] Due to the above-mentioned defects in the prior art, the present invention provides a method for predicting tumor radiotherapy response based on rule learning. First, the present invention collects tumor image data and performs tumor risk stratification. Secondly, the present invention develops a rule-based machine learning model, extracts rules from the FDG PET / CT radiomics of the entire tumor / tumor cluster sub-region through random forest, uses a heuristic algorithm to perform rule pruning, and builds a model based on this rule set to predict radiotherapy response. Finally, the present invention compares and summarizes the results of each model in predicting SUV change values ​​and whether the treatment is effective, proving that the method used can achieve good radiotherapy response prediction results while maintaining the interpretability of the model.

[0007] To achieve the above object, on the one hand, the present invention provides a rule-based method for predicting tumor radiotherapy response, comprising the following steps:

[0008] Step S1, image data acquisition and tumor risk stratification: FDG PET / CT image data of patients in the pre-treatment (PETpre) and mid-treatment (PETmid) stages in the original data set are acquired; the FDG PET / CT image data include but are not limited to the 3D spatial coordinates of tumor voxels and the standard uptake values ​​of tumor voxels; the K-means clustering algorithm is used to cluster the patient's tumor voxels and divide the tumor risk areas.

[0009] Step S2, radiomic feature extraction: extracting feature data from the FDG PET / CT image data of the above-mentioned patient's PETpre and PETmid, including multiple radiomic features, to characterize the tumor metabolic activity in response to the therapeutic radiation dose; using the change in the patient's tumor standard uptake value before and after radiotherapy to measure the radiotherapy response, including the FDG PET / CT radiotherapy mid-term binary response class for judging whether the radiotherapy is effective and the FDG PET / CT radiotherapy mid-term response change measurement value (ΔSUVmean) for measuring the effect of radiotherapy.

[0010] Step S3, constructing a random forest model and generating rules: constructing a random forest based on the training set, and extracting an initial rule set from the random forest; the random forest is composed of decision trees, where each decision tree model stores a set of information for each node in the form of an array, and based on this information, rules are extracted and stored from each decision tree.

[0011] Step S4: Use heuristic algorithm to perform rule pruning: Since the initial rule set contains a large number of rules, the Grey Wolf Optimizer (GWO) is used to perform rule pruning to screen out better rules for prediction.

[0012] Step S5, rule fitting, predicting radiotherapy response: Use the RuleFit algorithm to fit the pruned rule set. Each rule is regarded as a basic learner, and an integrated learning model is built based on the rule set; in this model, each input sample is converted into a 0-1 classification result based on whether it matches the rule; the 0-1 matrix composed of the 0-1 classification result of each sample corresponding to each rule and the original feature variable can be used as the new feature variable of the training set input to the model.

[0013] According to different prediction requirements, the least absolute shrinkage and selection operator (LASSO) is used to linearly combine the basic learners transformed from a single rule. The LASSO logistic regression model is used for classification tasks, and the LASSO regression model is used for regression tasks. In classification prediction, the binary response class of FDG PET / CT radiotherapy mid-term is used as the output of the rule, i.e., the prediction target; in regression prediction, ΔSUVmean is used as the prediction target; thus, an interpretable machine learning model is constructed for radiotherapy response prediction.

[0014] Furthermore, in step S1, a K-means clustering algorithm is used to generate clusters of pre-treatment and mid-treatment image time points for each patient, and n voxels are divided into k clusters through iteration to minimize the loss function corresponding to the clustering results; the tumor voxels of each patient are divided into 2 or 3 cluster sub-regions based on FDG PET / CT SUV and the 3D coordinates of the voxels; the 2 cluster sub-regions are defined as MTV high risk areas and MTV low risk areas; the sub-regions with intermediate risk levels among the 3 cluster sub-regions are removed, and the sub-regions with higher risk levels and the sub-regions with lower risk levels are re-clustered and defined as MTV high* risk areas and MTV low* risk areas.

[0015] Furthermore, in step S2, radiomic features are extracted by the MATLAB toolbox PET Oncology Radiomics Test Suite (PORTS), including 6 intensity histogram features, 19 texture features of Gray Tone Spatial Dependent Matrix (GTSDM) class, 5 texture features of Neighborhood Gray Tone Difference Matrix (NGTDM), and 11 texture features of Gray Level Zone Size Matrix (GLZSM) class.

[0016] Furthermore, in step S2, the calculation formula of the FDG PET / CT radiotherapy mid-term response change measurement value (ΔSUVmean) used to measure the radiotherapy effect is:

[0017]

[0018] Among them, PreSUVmean represents the average standard uptake value of all pixels before tumor radiotherapy; MidSUVmean represents the average standard uptake value of all pixels in the middle stage of tumor radiotherapy. The basis for judging whether radiotherapy is effective is whether the average standard uptake value of all pixels before tumor radiotherapy and the average standard uptake value of all pixels in the middle stage of radiotherapy satisfy ΔSUVmean>=0.2; if so, it is judged that the tumor radiotherapy for the pixel is effective; if not, it is judged that the tumor radiotherapy for the pixel is ineffective.

[0019] Furthermore, in step S4, in order to filter out a suitable number of rules from the selected rule set, an evaluation function is used to evaluate the selected rule set:

[0020]

[0021] For each rule in each rule set, cc represents the number of samples in the training sample that meet the rule, and ic represents the number of samples in the training sample that do not meet the rule. cc Indicates the sum of cc of each rule in the rule set, rule ic Indicates the sum of the ICs of each rule in the rule set; rule count Represents the sum of the number of times each rule is included in the initial rule set (multiple tree classifiers may use the same rule), tree num Represents the total number of tree classifiers in the random forest; m and n are set constants, m is used to keep the denominator non-zero, usually 4, and n is used to control the weights of the two parts, usually 10.

[0022] Furthermore, in step S4, based on the evaluation function, the GWO algorithm is used to iteratively select better rules, thereby completing the pruning of the rule set; the algorithm pseudo code is shown in Table 1, which shows the algorithm flow.

[0023] Table 1 GWO algorithm pseudo code

[0024]

[0025]

[0026] Furthermore, in the GWO algorithm, the position of the gray wolf is used to represent the selected rule set, and the prey is used to represent the optimal rule set, so as to simulate the hunting behavior of the gray wolf and screen out high-quality rules. For the initial rule set containing k rules, the GWO algorithm randomly generates p uniformly distributed weight vectors to represent the initial wolf pack of p gray wolves, that is, p selected rule sets. Each vector (x1, x2, x3, ..., x q) contains q elements, i.e., the positions of individual gray wolves, and represents a rule set containing q rules. Each element x in each vector i is a random number between 0 and 1, used to represent the |kxth i | rules.

[0027] Further, the iteration part of the GWO algorithm mainly includes the following steps. First, the gray wolf (per selected rule set) gradually approaches the prey and surrounds it. Where t is the current iteration number, A and C are the coordination coefficient vectors. A and C are calculated according to the formula:

[0028] A=2a×r1-a

[0029] C=2r2

[0030] A changes linearly from 2 to 0 during the iteration process, and r1 and r2 are random vectors between 0 and 1. X(t) represents the current position vector of the gray wolf, and X(t+1) represents the updated position vector of the gray wolf. The gray wolf uses the guidance of α wolf, β wolf, and δ wolf to find prey and approach it, and updates the position vector according to the formula.

[0031] D α =C1×X α -X(t),X1=X α -A1×D α

[0032] D β =C2×X β -X(t),X2=X β -A2×D β

[0033] D δ =C3×X δ -X(t),X3=X δ -A3×D δ

[0034] X(t+1)=(X1+X2+X3) / 3

[0035] Where X α , X β , X δ Represents the position of α wolf, β wolf, and δ wolf. α , D β , D δRespectively represent the distance between the current wolf and the best three wolves. The gray wolf updates its position and moves towards the prey, thereby obtaining a better set of rules. Subsequently, when |A|<1, the gray wolf attacks the prey and updates its position; when |A|>1, the gray wolf performs a random search. As described above, after 100 iterations, the optimal gray wolf individual can be obtained, that is, the rule pruning is completed, and the optimal rule group is obtained.

[0036] Furthermore, the number of rule sets obtained by the pruning is preferably 15-20. If the number of rules is too small, it is difficult to achieve high prediction accuracy, and if the number of rules is too large, the model will be too complex and difficult to build.

[0037] Furthermore, in step S5, the constructed model is interpretable; the constructed model can evaluate the rules that constitute the model, and explain how these rules and features work in the integrated model by calculating the support and importance of each rule and the importance of the features contained therein.

[0038] A second aspect of the present invention provides a rule-learning tumor radiotherapy response prediction system, comprising:

[0039] The image data preprocessing module is used to process the collected FDG PET / CT image data of patients before treatment (PETpre) and during treatment (PETmid), and cluster the patient's tumor voxels using the K-means clustering algorithm to divide the tumor risk area; extract feature data from the FDG PET / CT image data of the above patients' PETpre and PETmid, including 41 radiomics features. The change of the patient's tumor standard uptake value before and after radiotherapy is calculated to measure the radiotherapy response, including the FDG PET / CT radiotherapy mid-term binary response class used to determine whether the radiotherapy is effective and the FDG PET / CT radiotherapy mid-term response change measurement value (ΔSUVmean) used to measure the effect of radiotherapy; finally, the obtained data set is divided into a training set and a test set.

[0040] The model building module includes three parts: rule generation, rule pruning, and rule fitting. The first part generates an initial rule set based on the feature data set obtained by image data preprocessing. The second part uses the GWO algorithm to perform rule pruning to obtain a better rule set. The third part uses the RuleFit algorithm to perform rule fitting to build an interpretable machine learning model.

[0041] A training module is used to train the machine learning model using training set data.

[0042] The prediction module includes a trained machine learning model and rules for building the model, and is used to input the data to be predicted, i.e., the test set data, into the trained machine learning model to generate a prediction result.

[0043] The model analysis module evaluates the rules that make up the model based on the interpretability of the model; by calculating the support and importance of each rule and the importance of the features contained therein, it explains how these rules and features work in the integrated model.

[0044] Compared with the prior art, the above invention has the following advantages or beneficial effects:

[0045] (1) The present invention extracts features from image data and generates rules based on the feature data combined with random forests, thereby providing interpretable prediction of radiotherapy response;

[0046] (2) The present invention ingeniously uses the GWO algorithm for rule pruning, efficiently screening out high-quality rule sets for prediction, and improving the accuracy of prediction;

[0047] (3) The present invention uses the improved RuleFit algorithm to construct a rule-learning-based radiotherapy response prediction model, which has both explainability and high prediction accuracy. It can intelligently assist doctors in tumor radiotherapy diagnosis and promote clinical decision support in patient risk stratification, treatment selection and biologically driven adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention and its features and advantages will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following accompanying drawings.

[0049] Figure 1 A flowchart of the steps of a method for predicting tumor radiotherapy response in one embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the structure of a tumor radiotherapy response prediction system in one embodiment of the present invention;

[0051] Figure 3 A flowchart of the steps of constructing a prediction model in a method for predicting tumor radiotherapy response in one embodiment of the present invention;

[0052] Figure 4 FIG. 4 is a flow chart of a GWO algorithm in one embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0054] The terms "comprises", "including" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products or apparatuses.

[0055] In the following detailed description, many specific details are set forth to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the well-known control algorithms are not shown in detail to avoid blurring the main idea of ​​the present invention; and the K-means algorithm, feature extraction, rule generation, GWO algorithm, RuleFit algorithm and other technologies involved in the following effect embodiments are prior arts that can be retrieved.

[0056] Example 1

[0057] like Figure 1 and 3 As shown, this embodiment provides a tumor radiotherapy response prediction method based on rule learning, which is used to predict the FDG PET / CT radiotherapy mid-term binary response class that measures whether the radiotherapy is effective, and includes the following steps:

[0058] Step S1, image data acquisition and tumor risk stratification: FDG PET / CT image data of patients in the pre-treatment (PETpre) and mid-treatment (PETmid) stages in the original data set are acquired; the FDG PET / CT image data include the 3D spatial coordinates of tumor voxels, the standard uptake values ​​of tumor voxels, etc.

[0059] The K-means clustering algorithm was used to generate clusters of pre-treatment and mid-treatment image time points for each patient, and n voxels were iteratively divided into k clusters to minimize the loss function corresponding to the clustering results; the tumor voxels of each patient were divided into 2 or 3 cluster sub-regions based on FDG PRTSUV and the 3D coordinates of the voxels; the 2 cluster sub-regions were defined as MTV high risk area and MTV low risk area; the sub-regions with intermediate risk levels among the 3 cluster sub-regions were removed, and the sub-regions with higher risk levels and lower risk levels were re-clustered and defined as MTV high* risk area and MTV low* risk area.

[0060] Step S2, radiomics feature extraction: feature data are extracted from the FDG PET / CT image data of the above-mentioned patients' PETpre and PETmid, including 41 radiomics features; among them, 41 radiomics features are extracted through the MATLAB toolbox PET Oncology Radiomics Test Suite (PORTS), including 6 intensity histogram features, 19 texture features of Gray Tone SpatialDependent Matrix (GTSDM) class, 5 texture features of Neighborhood Gray Tone DifferenceMatrix (NGTDM), and 11 texture features of Gray Level Zone Size Matrix (GLZSM) class.

[0061] The calculation formula for the mid-term response change measure (ΔSUVmean) of FDG PET / CT radiotherapy used to measure the effect of radiotherapy is:

[0062]

[0063] Among them, PreSUVmean represents the average standard uptake value of all pixels before tumor radiotherapy; MidSUVmean represents the average standard uptake value of all pixels in the middle stage of tumor radiotherapy. The basis for judging whether radiotherapy is effective is whether the average standard uptake value of all pixels before tumor radiotherapy and the average standard uptake value of all pixels in the middle stage of radiotherapy satisfy ΔSUVmean>=0.2; if so, it is judged that the tumor radiotherapy for the pixel is effective; if not, it is judged that the tumor radiotherapy for the pixel is ineffective.

[0064] Step S3, constructing a random forest model and generating rules: constructing a random forest based on the training set, and extracting an initial rule set from the random forest; each decision tree model in the random forest stores a set of information for each node in the form of an array, and based on this information, extracting rules from each decision tree and storing them.

[0065] Step S4: Use heuristic algorithm to perform rule pruning: Since the initial rule set contains a large number of rules, the GWO algorithm is used to perform rule pruning to screen out better rules for prediction (such as Figure 4 As shown); the selected rule set is evaluated using the evaluation function:

[0066]

[0067] For each rule in each rule set, cc represents the number of samples in the training sample that meet the rule, and ic represents the number of samples in the training sample that do not meet the rule. ccIndicates the sum of cc of each rule in the rule set, rule ic Indicates the sum of the ICs of each rule in the rule set; rule count Represents the sum of the number of times each rule is included in the initial rule set (multiple tree classifiers may use the same rule), tree num Represents the total number of tree classifiers in the random forest; m and n are set constants, m is used to keep the denominator non-zero, usually 4, and n is used to control the weights of the two parts, usually 10.

[0068] Step S5, rule fitting, predicting radiotherapy response: Use the RuleFit algorithm to fit the pruned rule set, each rule is regarded as a basic learner, and an integrated learning model is built based on the rule set; each input sample is converted into a 0-1 classification result based on whether it matches the rule; the 0-1 matrix composed of the 0-1 classification result of each sample corresponding to each rule and the original feature variable can be used as the new feature variable of the training set input for the model. The basic learners transformed from a single rule are linearly combined, and the LASSO logistic regression model is used to predict whether radiotherapy is effective. Whether radiotherapy is effective is used as the output of the rule, that is, the prediction target, and classification prediction is performed.

[0069] like Figure 2 As shown, this embodiment uses a tumor radiotherapy response prediction system based on rule learning to implement the above prediction method. The prediction system includes:

[0070] The image preprocessing module is used to process the FDG PET / CT image data collected from patients in the pre-treatment period (PETpre) and mid-treatment period (PETmid), and cluster the patient's tumor voxels using the K-means clustering algorithm to divide the tumor risk area.

[0071] Feature data, including 41 radiomic features, were extracted from the FDG PET / CT image data of the above patients' PETpre and PETmid. The changes in the standard uptake values ​​of the tumors before and after radiotherapy were calculated to measure the radiotherapy response, that is, the FDG PET / CT radiotherapy mid-term binary response class used to determine whether radiotherapy is effective. Finally, the obtained data set was divided into a training set and a test set.

[0072] The model building module includes three parts: rule generation, rule pruning, and rule fitting. The first part generates an initial rule set based on the feature data set obtained by image data preprocessing. The second part uses the GWO algorithm to perform rule pruning to obtain a better rule set. The third part uses the RuleFit algorithm to perform rule fitting to build an interpretable machine learning model.

[0073] A training module, using training set data to train the machine learning model;

[0074] The prediction module includes a trained machine learning model and rules for building the model, and is used to input the data to be predicted, i.e., the test set data, into the trained machine learning model to generate a prediction result.

[0075] The model analysis module is used to evaluate the rules that make up the model based on the interpretability of the model. By calculating the support and importance of each rule and the importance of the features contained in it, it explains how these rules and features work in the integrated model.

[0076] The prediction system predicts the binary response class of FDG PET / CT radiotherapy mid-term after processing the collected tumor image data by the prediction method, that is, whether the radiotherapy is effective. The area under the curve (AUC) is selected as the evaluation index of classification prediction. AUC is the area under the receiver operating characteristic (ROC) curve, which is often used as an evaluation criterion for the model, corresponding to the model with a larger AUC and a better classification effect.

[0077] The model was validated by the leave-one-out cross-validation method. The validation idea and process of the leave-one-out cross-validation method are as follows: retain the data of one patient for testing each time, and use the data of the remaining patients for training, and use the trained model to predict the test sample. At the same time, common machine learning methods were used as control methods, including the least absolute shrinkage and selection operator (LASSO), random forest (RF), and gradient boosted decision tree (GBDT). According to the above-mentioned image data acquisition and tumor risk stratification method, the patient's tumor was divided into risk areas, and data sets of five risk areas, MTV, MTV high, MTV low, MTV high*, and MTV low*, were obtained for prediction. Table 2 shows the prediction results of the FDG PET / CT radiotherapy mid-term binary response class by the rule learning method of the present invention and the common machine learning method. It can be seen from the table that the rule learning method can achieve a stable and better AUC, and an AUC of 0.86 can be achieved in the MTVlow* risk area.

[0078] Table 2 Prediction results of binary response in the middle term of radiotherapy by FDG PET / CT

[0079]

[0080] According to the interpretability of the rule learning method, the present invention can evaluate the rules that constitute the model. Taking the model trained on the entire tumor as an example, the model consists of 15 rules, and X[i] represents the value of feature i. As shown in Table 3, the support and importance of each rule explain how these rules work in the integrated model. The importance reflects the impact of each rule in the model on the decision result, and the support is the proportion of data in the data set that conforms to the rule. The rule learning method can further analyze the rule set and calculate the importance of the features contained therein. Table 4 shows the importance of the features, among which feature 4 (KURT) and feature 22 (CS) have high importance in the model (1.419 and 4.132). These evaluations and analyses explain the role of each rule and feature in the model, which contributes to more accurate medical diagnosis.

[0081] Table 3 Rule set included in the rule learning model

[0082]

[0083]

[0084] Table 4 Features included in the rule set

[0085]

[0086] Example 2

[0087] This embodiment uses a prediction method similar to that of Embodiment 1 to predict radiotherapy response. Unlike Embodiment 1, this embodiment predicts the FDG PET / CT radiotherapy mid-term response change metric (ΔSUVmean) used to measure the radiotherapy effect, so a rule learning regression prediction method is used, and the steps are as follows:

[0088] Step S1, image data acquisition and tumor risk stratification: FDG PET / CT image data of patients in the pre-treatment (PETpre) and mid-treatment (PETmid) stages in the original data set are acquired; the FDG PET / CT image data include the 3D spatial coordinates of tumor voxels, the standard uptake values ​​of tumor voxels, etc.; the K-means clustering algorithm is used to cluster the patient's tumor voxels and divide the tumor risk areas.

[0089] Step S2, radiomics feature extraction: extract feature data from the FDG PET / CT image data of the above patient's PETpre and PETmid, including 41 radiomics features; use the change in the patient's tumor standard uptake value before and after radiotherapy to measure the radiotherapy response, that is, the FDG PET / CT radiotherapy mid-term response change measurement value (ΔSUVmean) used to measure the radiotherapy effect; the calculation formula of ΔSUVmean is:

[0090]

[0091] Among them, PreSUVmean represents the average standard uptake value of all pixels before tumor radiotherapy; MidSUVmean represents the average standard uptake value of all pixels in the middle stage of tumor radiotherapy.

[0092] Step S3, constructing a random forest model and generating rules: constructing a random forest based on the training set, and extracting an initial rule set from the random forest; each decision tree model in the random forest stores a set of information for each node in the form of an array, and based on this information, extracting rules from each decision tree and storing them.

[0093] Step S4, using a heuristic algorithm to perform rule pruning: Since the initial rule set contains a large number of rules, the GWO algorithm is used to perform rule pruning as in Example 1 to screen out better rules for prediction.

[0094] Step S5, rule fitting, predicting radiotherapy response: Use the RuleFit algorithm to fit the pruned rule set. Each rule is regarded as a basic learner, and an integrated learning model is built based on the rule set; each input sample is converted into a 0-1 classification result according to whether it matches the rule; the 0-1 matrix composed of the 0-1 classification result of each sample corresponding to each rule and the original feature variable can be used as a new feature variable of the training set for model input. The basic learners transformed from a single rule are linearly combined, and the LASSO regression model is used to predict whether radiotherapy is effective. The FDG PET / CT radiotherapy mid-term response change measure (ΔSUVmean) is used as the output of the rule, that is, the prediction target for regression prediction.

[0095] This embodiment uses a tumor radiotherapy response prediction system based on rule learning to implement the above prediction method. The prediction system includes:

[0096] The image preprocessing module is used to process the FDG PET / CT image data collected from patients in the pre-treatment period (PETpre) and mid-treatment period (PETmid), and cluster the patient's tumor voxels using the K-means clustering algorithm to divide the tumor risk area.

[0097] Feature data, including 41 radiomics features, were extracted from the FDG PET / CT image data of the patients' PETpre and PETmid. The changes in the standard uptake value of the tumor before and after radiotherapy were calculated to measure the radiotherapy response, namely the FDG PET / CT radiotherapy mid-term response change measure (ΔSUVmean). Finally, the obtained data set was divided into a training set and a test set.

[0098] The model building module includes three parts: rule generation, rule pruning, and rule fitting. The first part generates an initial rule set based on the feature data set obtained by image data preprocessing. The second part uses the GWO algorithm to perform rule pruning to obtain a better rule set. The third part uses the RuleFit algorithm to perform rule fitting to build an interpretable machine learning model.

[0099] A training module, using training set data to train the machine learning model;

[0100] The prediction module includes a trained machine learning model and rules for building the model, and is used to input the data to be predicted, i.e., the test set data, into the trained machine learning model to generate a prediction result.

[0101] The model analysis module is used to evaluate the rules that make up the model based on the interpretability of the model. By calculating the support and importance of each rule and the importance of the features contained in it, it explains how these rules and features work in the integrated model.

[0102] The prediction system predicts the FDG PET / CT radiotherapy mid-term response change measure (ΔSUVmean) after processing the collected tumor image data through the prediction method. The root mean square error (RMSE) is selected as the evaluation index of regression prediction, that is, the sum of the squares of the deviations between the predicted value and the true value divided by the square root of the number of observed values, which is often used to measure the deviation between the predicted value and the true value.

[0103] The model was validated by the leave-one-out cross-validation method. The validation idea and process of the leave-one-out cross-validation method are as follows: retain the data of one patient for testing each time, and use the data of the remaining patients for training, and use the trained model to predict the test sample. At the same time, common machine learning methods were used as control methods, including the least absolute shrinkage and selection operator (LASSO), random forest (RF), and gradient boosted decision tree (GBDT). According to the above-mentioned image data acquisition and tumor risk stratification method, the patient's tumor was divided into risk areas, and data sets of five risk areas, MTV, MTV high, MTV low, MTV high*, and MTV low*, were obtained for prediction. Table 5 shows the prediction results of the FDG PET / CT radiotherapy mid-term response change measure (ΔSUVmean) of the rule learning method of the present invention and the common machine learning method. It can be seen from the table that the rule learning method can achieve a stable and low RMSE, and an RMSE of 0.189 can be achieved in the MTV high* risk area.

[0104] Table 5 Prediction results of the mid-term response change measure (ΔSUVmean) of FDG PET / CT radiotherapy

[0105]

[0106] Those skilled in the art should understand that those skilled in the art can implement variations by combining the prior art and the above embodiments, which will not be described in detail here. Such variations do not affect the essential content of the present invention, and will not be described in detail here.

[0107] The above describes the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the above-mentioned specific embodiments, and the devices and structures that are not described in detail should be understood to be implemented in a common manner in the art; any technician familiar with the art can use the above-disclosed methods and technical contents to make many possible changes and modifications to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, or modify them into equivalent embodiments of equivalent changes, which does not affect the essential content of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention are still within the scope of protection of the technical solutions of the present invention.

Claims

1. A method for predicting tumor radiotherapy response based on rule learning, characterized in that: The following steps are involved: Step S1, image data acquisition and tumor risk stratification: collecting FDG PET / CT image data of the patient's pre-treatment PETpre and mid-treatment PETmid in the original data set; the FDG PET / CT image data includes but is not limited to the 3D spatial coordinates of the tumor voxels and the standard uptake value of the tumor voxels; clustering the patient's tumor voxels using the K-means clustering algorithm to divide the tumor risk area; Step S2, radiomics feature extraction: extracting feature data from the FDG PET / CT image data of the patient's PETpre and PETmid, including multiple radiomics features, to characterize the tumor metabolic activity in response to the therapeutic radiation dose; using the change in the tumor standard uptake value SUV of the patient before and after radiotherapy to measure the radiotherapy response, including the FDG PET / CT radiotherapy mid-term binary response class used to determine whether the radiotherapy is effective and the FDG PET / CT radiotherapy mid-term response change measurement value ΔSUVmean used to measure the effect of radiotherapy; Step S3, constructing a random forest model and generating rules: constructing a random forest based on the training set, and extracting an initial rule set from the random forest; Step S4: Use heuristic algorithm to perform rule pruning: Use GWO to prune the initial rule set and select the best rules for prediction; Step S5, rule fitting, predicting radiotherapy response: use the RuleFit algorithm to fit the pruned rule set; each rule is regarded as a basic learner, and an integrated learning model is constructed based on the rule set; in this model, each input sample is converted into a 0-1 classification result according to whether it matches the rule; a 0-1 matrix consisting of the 0-1 classification result of each sample corresponding to each rule and the original feature variable is used as a new feature variable of the training set input to the model; according to different prediction needs, the LASSO linear model is used to linearly combine the basic learners transformed from a single rule; the LASSO logistic regression model is used for classification tasks, and the LASSO regression model is used for regression tasks; in classification prediction, the FDG PET / CT radiotherapy mid-term binary response class is used as the output of the rule, that is, the prediction target; in regression prediction, ΔSUVmean is used as the prediction target; thereby constructing an interpretable machine learning model for radiotherapy response prediction.

2. The method for predicting tumor radiotherapy response based on rule learning according to claim 1, characterized in that: In step S1, a K-means clustering algorithm is used to generate clusters of pre-treatment and mid-treatment image time points for each patient; voxels are divided into 2 or 3 cluster sub-regions based on FDG PET / CT SUV and 3D coordinates of voxels; the 2 cluster sub-regions are defined as MTV high risk region and MTV low risk region; sub-regions with intermediate risk levels among the 3 cluster sub-regions are removed, and sub-regions with higher risk levels and sub-regions with lower risk levels are re-clustered and defined as MTV high* risk region and MTV low* risk region.

3. The method for predicting tumor radiotherapy response based on rule learning according to claim 1, characterized in that: In step S2, radiomic features are extracted through the MATLAB toolbox PET oncology radiomics test suite PORTS, including 6 intensity histogram features, 19 texture features of GTSDM class, 5 texture features of NGTDM class, and 11 texture features of GLZSM class.

4. The method for predicting tumor radiotherapy response based on rule learning according to claim 1, characterized in that: In step S2, the calculation formula of the FDG PET / CT radiotherapy mid-term response change measurement value ΔSUVmean used to measure the radiotherapy effect is: Among them, PreSUVmean represents the average standard uptake value of all pixels before tumor radiotherapy; MidSUVmean represents the average standard uptake value of all pixels in the middle stage of tumor radiotherapy; the basis for judging whether radiotherapy is effective is whether the average standard uptake value of all pixels before tumor radiotherapy and the average standard uptake value of all pixels in the middle stage of radiotherapy satisfy ΔSUVmean>=0.2; if so, it is judged that the tumor radiotherapy for the pixel is effective; if not, it is judged that the tumor radiotherapy for the pixel is ineffective.

5. The method for predicting tumor radiotherapy response based on rule learning according to claim 1, characterized in that: In step S4, in order to filter out a suitable number of rules from the selected rule set, an evaluation function is used to evaluate the selected rule set: For each rule in each rule set, cc represents the number of samples in the training sample that meet the rule, and ic represents the number of samples in the training sample that do not meet the rule. cc Indicates the sum of cc of each rule in the rule set, rule ic Indicates the sum of the ICs of each rule in the rule set; rule count Represents the sum of the number of times each rule is included in the initial rule set (multiple tree classifiers may use the same rule), tree num Represents the total number of tree classifiers in the random forest; m and n are set constants, m is used to keep the denominator non-zero, usually 4, and n is used to control the weights of the two parts, usually 10.

6. The method for predicting tumor radiotherapy response based on rule learning according to claim 5, characterized in that: Based on the evaluation function, the GWO Grey Wolf Optimization algorithm is used to iteratively select better rules, thereby completing the pruning of the rule set.

7. A method for predicting tumor radiotherapy response based on rule learning according to claim 1 or 6, characterized in that: The number of rule sets obtained by the pruning is 15-20.

8. The method for predicting tumor radiotherapy response based on rule learning according to claim 1, characterized in that: In step S5, the constructed model is interpretable; the constructed model can evaluate the rules that constitute the model, and explain how these rules and features work in the integrated model by calculating the support and importance of each rule and the importance of the features contained therein.

9. A tumor radiotherapy response prediction system based on rule learning, characterized in that: include: The image data preprocessing module is used to process the collected FDG PET / CT image data of the patient's pre-treatment PETpre and mid-treatment PETmid, and cluster the patient's tumor voxels using the K-means clustering algorithm to divide the tumor risk area; extract feature data from the patient's PETpre and PETmid FDG PET / CT image data, including multiple radiomics features; The change in the standard uptake value of the tumor before and after radiotherapy was calculated to measure the radiotherapy response, including the binary response class of the FDG PET / CT radiotherapy mid-term response used to determine whether the radiotherapy is effective and the FDG PET / CT radiotherapy mid-term response change measurement value ΔSUVmean used to measure the effect of radiotherapy; finally, the obtained data set was divided into a training set and a test set; The model building module includes three parts: rule generation, rule pruning, and rule fitting. The first part generates an initial rule set based on the feature data set obtained by image data preprocessing. The second part uses the GWO algorithm to perform rule pruning to obtain a better rule set. The third part uses the RuleFit algorithm to perform rule fitting to build an interpretable machine learning model. A training module, which trains the machine learning model using training set data; A prediction module, including a trained machine learning model and rules for building the model, is used to input the data to be predicted, i.e., the test set data, into the trained machine learning model to generate a prediction result; The model analysis module evaluates the rules that make up the model based on the interpretability of the model; by calculating the support and importance of each rule and the importance of the features contained therein, it explains how these rules and features work in the integrated model.

Images