Predicting Gamma Passing Rates Of Gliomas IMRT Plans Based On Machine Learning

To predict the gamma passing rates(GPR)of fields in Gliomas intensity-modulated radiation therapy(IMRT)plans by developing a convolution neural network(CNN)model based on deep learning and traditional machine learning models based on radiomics methods.Forty-eight pretreatment verification plans from Gliomas patients treated using IMRT were extracted from Eclipse treatment planning system.All plans were delivered on Varian 23 EX Linac and portal dosimetry system was used for dosimetry and gamma evaluation with2(global)%/2mm criterion.By using the dose distribution map of each field as input,a convolution neural network(CNN)model was developed to learn the correlation between the dose distribution map and the GPR.The data set was divided into training set with 208 fields,validation set and test set with 26 fields respectively,and mean absolute error(MAE)was used to evaluate the prediction effect of the model.In order to obtain the radiomics characteristics of dose distribution map,the dose distribution map was manually segmented one by one according to the shape of fields to obtain the region of interest(ROI)in itk-snap.A python program was written and the pyradiomics software package was used to extract the shape features,intensity features,texture features and high-order features of each ROI area.Feature engineering was carried out on the extracted parameters to select the features that are more relevant to the GPR and eliminate the interference of unstable features to prevent the model from over fitting.The data set was divided into validation set and test set according to the ratio of 8:2.The traditional machine learning algorithms such as linear regression,Lasso regression and random forest regression were used to train the training set.Root mean square error(RMSE),mean absolute error(MAE)and decision coefficient Score(R~2)were used to evaluate the prediction effect of those model.In our CNN model,the results show that 96%of predictions between±3%error from the measured value in the validation set and test set.The maximum prediction error are 3.09%and 3.54%,respectively,and mean absolute error(MAE)are 0.99%and 1.17%,respectively.The Pearson correlation coefficients between predicted and measured GPR in validation set and test set are 0.96 and 0.90,respectively.By using pyradiomics software package,we extracted a total of 843 radiomic features from original image and derived image of the dose distribution map,and retained 24 most important features after feature engineering.Among the machine learning models based on 24 radiomic features,random forest regression model has the best prediction performance witch the MAE is 1.71%,RMSE and R2 scores were 2.20%and 0.82%respectively.Compared with deep learning CNN model,the prediction effect of random forest regression model seems to be worse,but we can get the feature importance order of the model with a good interpretability..The results show that the dose distribution map of filed in IMRT plans is a useful data for GPR prediction.The CNN model and random forest regression model developed in this study can accurately predict the GPR of fields in IMRT plans,and it could inform the physicist which fields may not pass QA measurement in advance and effectively promote the QA work of clinical radiotherapy.