A Planning Evaluation Method for Esophageal VMAT Based on Machine Learning

To reduce the complexity associated with VMAT planning, we developed a model which can predict the dose volume histograms (DVHs) of organ-at-risk using the prior knowledge of the high quality esophageal VMAT plans and the distance to target histograms (DTHs). We extracted the anatomical information and dose information of patients from DICOM-RT files. With these information, the DTH and DVH curves were calculated. Principal component analysis was used to identify the main features of DTH and DVH curves. Then, least absolute shrinkage and selection operator regression was used to establish the functional relationship between the main features of DTH and DVH curves. In this study, the training dataset consists of 35 esophageal VMAT plans and the trained model was validated by 12 cases outside the training dataset. The experimental results demonstrated that the DTH/DVH curves can be effectively expressed by one or two principal components, the accuracy of the model in prediction is about 75%. These promising results suggest that this method can predict does distribution in the esophageal VMAT plans and assist the physicist to make plans by giving objective function and orientation, which can improve the efficiency and quality of plan making.