Incorporating dosimetric features into the prediction of 3D VMAT dose distributions using deep convolutional neural network

An accurate prediction of achievable dose distribution on a patient specific basis would greatly improve IMRT/VMAT planning in both efficiency and quality. Recently machine learning techniques have been proposed for IMRT dose prediction based on patient's contour information from planning CT. In these existing prediction models geometric/anatomic features were learned for building the dose prediction models and few features that characterize the dosimetric properties of the patients were utilized. In this study we propose a method to incorporate the dosimetric features in the construction of a more reliable dose prediction model based on the deep convolutional neural network (CNN). in addition to the contour information, the dose distribution from a PTV-only plan (i.e. the plan with the best PTV coverage by sacrificing the OARs sparing) is also employed as the model input to build a deep learning based dose prediction model. A database of 60 volumetric modulated arc therapy (VMAT) plans for the prostate cancer patients was used for training. The trained prediction model was then tested on a cohort of ten cases. Dose difference maps, DVHs, dosimetric endpoints and statistical analysis of the sum of absolute residuals (SARs) were used to evaluate the proposed method. Our results showed that the mean SARs for the PTV, bladder and rectum using our method were 0.007 +/- 0.003, 0.035 +/- 0.032 and 0.067 +/- 0.037 respectively, lower than the SARs obtained with the contours-based method, indicating the potential of the proposed approach in accurately predicting dose distribution.