Several Key Application Researches Of Machine Learning In Radiotherapy

Radiation therapy is one of the most commonly used methods to treat the malignant tumors.According to the statistics,more than 70%of malignant tumor patients need radiation therapy.The main goal of radiation therapy is to reduce the exposure of the surrounding normal tissues and the organ at risk as much as possible while killing the malignant tumor cells within the target area.Nowadays,radiotherapy technology becomes more and more mature.However,there are still many problems needs to be resolved:First of all,overall survival(OS)prediction is very important for doctors to formulate the radiotherapy plans.Doctors’ predictions of the OS are often inaccurate due to the inadequate medical technology.Therefore,a prognostic model which can predict the OS accurately is crucial for the radiation therapy.This is also one of the key research contents of this thesis.Secondly,in the actual treatment process,the target area(tumor area)changes dramatically due to the respiratory motion.In order to reduce the side effect of the respiratory motion,the doctors always add a margin to the tumor area.However,this may increase unnecessary exposure to the normal tissues and organs.In addition,we need to know the position of the tumor target moving with respiration in advance as for the dose delivery system needs time to synchronize with the position of the target.Hence,accurate prediction of respiratory signal is also one of the key research contents of this thesis.Finally,the evaluation of radiation therapy treatment effect is also important.This is because that after the surgery,the tumor may disappear from the medical images and the tumor markers may also recover to be normal again.In contrast,tumor markers will change gradually after radiation therapy.Therefore,we can continue to monitor these changes and combine these qualitative changes with the clinical experience to provide accurate evaluation of radiation therapy outcome.This is also one of the key research contents of this thesis.In recent years,the machine learning(ML)technology is changing the world rapidly along with the data sets expansion and sharing,and the improvement of the computing power in many fields.A lot of studies show that the ML technology has great potential in the field of radiation therapy.Therefore,in order to solve the above mentioned problems,this thesis uses ML methods to implement four applied studies in three aspects:respiratory prediction,OS prediction of non-small cell lung cancer(NSCLC)patients and evaluation of concurrent stereotactic radiosurgery(SRS)and bevacizumab(BVZ)treatment of recurrent malignant gliomas.The first part is about the OS prediction.The second and third part are about the respiratory prediction.The fourth part is about the treatment evaluation:1.Effect of ML methods on predicting NSCLC overall survival time based on radiomics analysis.We investigated the effect of ML(regression)methods on predicting the OS for non-small cell lung cancer based on radiomics features analysis.We extracted 339 radiomic features from the segmented tumor volumes of pretreatment CT images.The tumor phenotypic characteristics of the medical images were quantified through these radiomic features using the tumor shape and size,the intensity statistics,and the textures.We investigated the performance of 8 ML methods and 5 feature selection methods for the OS prediction.The concordance index(CI)between the real and predicted OS for the NSCLC patients was used to evaluate the predicted performance.The gradient boosting linear models based on Cox’s partial likelihood method(GB-Cox)using the CI feature selection method get the best performance.The results showed that appropriate combination of ML and radiomics methods could predict the OS accurately for the NSCLC patients.2.Respiratory signal prediction based on adaptive boosting and multi-layer perceptron neural network.We improved the respiratory signals prediction accuracy using the multi-layer perceptron neural network and the adaptive boosting methods(ADMLP-NN)for gated treatment of moving target in radiation therapy.In this study,the respiratory signals obtained from the Real-time Position Management(RPM)device from 138 previous 4D-CT scans were used,retrospectively.The ADMLP-NN was composed of several artificial neural networks(ANNs).These ANNs were utilized as the weaker predictors to constitute a stronger predictor.The recorded respiratory signals were firstly smoothed by a Savitzky-Golay finite impulse response smoothing filter(S-G filter).Then,the multi-layer perceptron neural networks(MLP-NNs)were established to predict the future respiratory signals according to its previous positions.Finally,the weights for each MLP-NN based on the prediction error of each MLP-NN were determined using the adaptive boosting(Adaboost)decision algorithm.The MLP-NN and ADMLP-NN(MLP-NN plus adaptive boosting)methods were evaluated by root-mean-square-error(RMSE),correlation coefficient(CC)and maximum error(ME)between predicted and real signals.For 500-ms ahead time,the average CC were improved from 0.83(MLP-NN method)to 0.89(ADMLP-NN method).Besides,the averages of the RMSE(relative unit)and the ME were reduced by 27.9%and 22.2%,respectively.The experiment results demonstrated that the ADMLP-NN method could improve the performance of respiratory signal prediction compared to the MLP-NN method.Besides,several variants of Adaboost methods showed great potential for the regression prediction problem recently.Hence,we investigated the prediction performance of eight popular adaptive boosting method based on the MLP-NN for the respiratory prediction problem in this study.The Adaboost.RT method(power factor is 1)get the best RMSE and CC while the Adaboost.BCC obtained the best ME.The experiment results demonstrated that different Adaboost methods were suited for different radiotherapy application scenarios.Hence,in order to treat patients with radiation therapy more accurately,we should select the ADMLP-NN methods according to the specific application scenarios in the radiation therapy.3.Adaptive respiratory signal prediction using dual multi-layer perceptron neural networks(MLP-NNs).We improved the prediction accuracy of respiratory signals by adapting the predictor to changing signals using dual multi-layer perceptron neural networks(MLP-NNs).When one MLP-NN was performing prediction of the respiratory signals,another one was being re-trained using the updated data and vice versa.The effects of adding an additional network,re-training parameter,and respiratory signal irregularity on the performance of new predictor were investigated based on 4 different network configurations:a single MLP-NN,high-computation dual MLP-NNs(U1-adaptation),two different combinations of high-and low-computations dual MLP-NNs(U2-adaptation and U3-adaptation).In the U2-adaptation method,an HC-MLP-NN was used to fully re-train the prediction model and an LC-MLP-NN was utilized to update(fine tune)the prediction model in real time.In the U3-adaptation method(also called the sliding window approach method),an HC-MLP-NN was used to fully re-train the prediction model and then an LC-MLP-NN would be used to predict all the respiratory signals after the weights and biases of the trained model established by the HC-MLP-NN were first transferred to the LC-MLP-NN.The RMSEs between the predicted signals and the real signals were used to evaluate the predicted performance.204 patients’ real respiratory signals were tested in this study.The RMSEs using the U1-adaptation method were reduced by 34%,19%,and 10%compared to those using the MLP-NN,U2-adaptation and U3-adaptation methods,respectively.The experiment demonstrated that continuous re-training of MLP-NN based on a dual-network configuration using updated respiratory signals improved prediction accuracy compared to one-time training of MLP-NN using fixed signals.4.Assessment of concurrent stereotactic radiosurgery and bevacizumab treatment of recurrent malignant gliomas using multi-modality MRI imaging and radiomics analysis.We assessed the response of recurrent malignant gliomas(MG)patients treated with concurrent BVZ/SRS using multi-modality MRI imaging and radiomics analysis.Besides,we selected the radiomics features related to the OS and used these features to predict the OS of the patients with recurrent malignant gliomas.Various patients with recurrent MG were tested in this study.For the DCE results,Ktrans and VB illustrated significant decrease 2 months after SRS and FB demonstrated significant decrease as early as 1 week(p=0.017)after SRS.No functional parameters reflected statistically significant treatment response 1 week after SRS.A total of 888 radiomics features were extracted.31/126 features showed significant changes 1 week/2 months after SRS,respectively.Five features(highest correlation with the OS)and the SVRC model were used for the OS prediction.The CI was 0.68.The results of this work demonstrated the potential of combined radiomics analysis and functional MR imaging in quantitatively identifying early treatment response of concurrent SRS/BVZ.Hence,this study provides the basis for further treatment.In this thesis,the applications of several ML methods,such as neural networks,adaptive boosting,random survival forest and support vector regression for deleted data,in radiation therapy are investigated.The main innovation points of this thesis are as follows:1.The effects of various ML models which could dealt with continuous time-to-event data for NSCLC based on radiomics were first investigated.Most of this type of studies transformed outcome of interest into a dichotomized endpoint.This may lead to the bias of prediction accuracy problem.This study was an important supplementary reference for the use of prognostic models based on radiomics analysis.In addition,this study was helpful to select the optimal ML method based on radiomics analysis for the OS prediction of NSCLC patients and has considerable application value in the clinic.2.A new algorithm(ADMLP-NN)was proposed to predict the respiratory signal accurately by introducing the Adaboost optimization method into the MLP-NN method.The Adaboost method improved the respiratory signal prediction accuracy by reducing the risk of local minimum and overfitting for the MLP-NN algorithm.Hence,the proposed method had important application value in clinic.In addition,the ADMLP-NN method was more robust than the MLP-NN method(lower ME value and fewer predicted outliers).This was also very important for the radiation therapy because large deviations(outliers)of the instantaneous motion may lead to large exposure to the highly sensitive organs which were near the target area(tumor).3.The effect of several popular Adaboost algorithm on the respiratory signal prediction is investigated in this study.We found that we should select the appropriate Adaboost algorithm according to the application scenarios in radiation therapy.This novel founding has high clinical application value.4.The respiratory signals often changed with time based on emotional and physical conditions in clinical situations.In order to solve this crucial problem,we proposed two new methods(Ul-and U2-adaptation methods)using dual MLP-NNs to accurately predict respiratory signals.The MLP-NN models in the proposed methods were continue to be trained and updated use the real signals collected on-line.The proposed methods improved the prediction performance by overcoming the underfitting and prediction errors accumulating problems for the traditional on-line MLP-NN prediction method.Hence,the methods proposed in this thesis could potentially be utilized as a valuable tool to predict respiratory motion for dynamic tracking of moving targets during radiation therapy.5.In order to improve the evaluation accuracy of the treatment,we proposed a novel method to evaluate the treatment effect of the concurrent BVZ with SRS by using multiple ROIs of multi-model MRI within multiple time periods together with the radiomics analysis.For the standard analysis of the medical imaging,we found that the FB could be used as an early evaluation index to predict whether it is suitable for a possible adaptive therapy.For the radiomics analysis of the medical imaging,we found that the radiomics features extracted from the anatomic MRI had the potential value in capturing the therapeutic effects.In addition,as the radiomics features is highly correlated with the OS,it could be used to accurately predict the OS of the glioma patients.The method proposed in this paper provides a new idea for evaluating the therapeutic response of the concurrent SRS with BVZ in the treatment of malignant glioma,and hence had potential prosperous application value in clinic.The experimental results show that the performance of these above mentioned innovative methods proposed in this thesis are positive in the clinical application of radiotherapy.These methods have a wide application prospect in the field of radiation therapy.