IMRT Automatic Planning Study Based On Three-dimensional Predicted Dose

The planning for Intensity Modulated Radiotherapy(IMRT)is time-consuming and the plan quality is easily affected by many factors.Automatic planning is attracting more attention from end users because it can improve not only the efficiency of planning but also the consistency.However,existing studies of IMRT automatic planning involving dose prediction are based on fixed and equally spaced beams mostly that are difficult for varying beam configurations.This project aims to verify the feasibility of developing an IMRT automatic planning method that is based on predicted three dimensional dose and that is suitable for multi-beam configurations.To achieve this aim,two tasks are completed:(1)Accurately predict the three-dimensional dose distributions on the dataset with multi-beam configurations.Firstly,the beam’s paths,as well as voxel to source and voxel to beam central-axis distances are calculated according to CT images and DICOM RT structure files;then train a neural network named 3D U-Net_ABS to export three-dimensional dose distributions.(2)Develop an inverse planning method based on the predicted dose.Take DVH curves of the predicted dose as targets,then optimize each field’s intensity map as well as aperture’s shape and weights,generating plans automatically.Seventy five IMRT cases of the cervical cancer with 11 beam configurations from the First Affiliated Hospital of Zhengzhou University are considered.After randomly deranging the data,60 cases are taken as the train-validation datasets and 15 cases as the test datasets.Differences between the predicted dose and clinical plan on the test dataset in terms of dose distributions and DVH parameters are analyzed.Results from 3D U-Net_ABS and 3D U-Net_CRT model,and dosimetric indexes from automatic plans and clinical plans are compared.Results show that:(1)3D U-Net_ABS model in this study is equivalent to or better than the 3D U-Net_CRT model in terms of most indexes.Taking the mean value of absolute deviation for all important DVH indexes of OARs,it is found that 3D U-Net_ABS yields 6.3317.06%,outperforming 3D U-Net_CRT model which yields 9.84±11.04%As for the mean dose of PTV and OARs,it is found that 3D U-Net_ABS yields 1.36±1.14%,still outperforming 3D U-Net_CRT model which yields 1.86±1.52%.The mean MAE value of the 3D U-Net ABS model is found to be 2.82±1.45Gy,while that for the 3D U-Net_CRT model is found to be 2.36±1.13Gy.The prediction accuracy of femoral heads are significantly improved.Compared with 3D U-Net_CRT,the mean absolute deviation of V2O decreases around 10%,and the mean absolute deviation of Dmean decreases around 1.6%in the volume of femoral heads for the 3D U-Net_ABS model.The Dice value of the 3D U-Net_ABS is in the range of[0.86,0.95].The 3D U-Net_ABS model improves the Dice scores associated with 30%-55%and 80%-100%isodose surfaces.(2)The DVH curves of automatic plans and clinical plans,the dose distributions of axial,coronal and sagittal are basically consistent.The mean value of some dosimetric indexes are better than those of clinical plans.As for V93%and D98 of PTV volumes,automatic plans are slightly higher than those of clinical plans(t=4.89,P=0.00;t=3.08,P=0.01).As for V40 of the rectum,automatic plans are lower than those of clinical plans(t=-3.63,P=0.00).There is no statistically significate difference for remaining dosimetric parameters(P>0.05).This study proves the feasibility of automatic planning in IMRT cases involving multi-beam configurations,and enhances the universality and practicability of automatic planning in the radiotherapy planning workflow.