Voxel-Level Radiotherapy Dose Prediction Using Densely Connected Network with Dilated Convolutions

Radiotherapy treatment planning often demands substantial manual adjustments to achieve maximal dose delivery at the planning target volumes (PTVs) and protecting surrounding organs at risk (OARs). Automatic dose prediction can reduce manual adjustments by providing close to optimal radiotherapy planning parameters, which is studied in this work. We developed a voxel-level dose prediction framework based on an end-to-end trainable densely-connected network. We designed a four-channel map to record the geometric features of PTVs, OARs, and the prescription dose of each patient. The densely connected block was modified with dilated convolutions to catch multi-scale features, which can result in accurate dense prediction. 90 esophageal radiation treatment plans from 90 patients were used in this work (72 plans used for training and the remaining 18 plans for testing). Average value of mean absolute error of dose volume histogram (DVH) and voxel-based mean absolute error were used to evaluate the prediction accuracy, with [0.9%, 1.9%] at PGTV, [1.1%, 2.8%] at PTV, [2.8%, 4.4%] at Lung, [3.5%, 6.9%] at Heart, [4.2%, 5.6%] at Spinal Cord, and [1.7%, 4.8%] at Body. These encouraging results demonstrated that the proposed framework could provide accurate dose prediction, which could be very useful to guide radiotherapy treatment planning.