One-Dimensional Convolutional Network for Dosimetry Evaluation at Organs-at-Risk in Esophageal Radiation Treatment Planning

Dose volume histogram (DVH) is an important dosimetry evaluation metric and it plays an important role in guiding the development of esophageal ra-diotherapy treatment plans. Automatic DVH prediction is therefore very use-ful to achieve high-quality esophageal treatment planning. This paper studied stacked denoise auto-encoder (SDAE) to compute correlation between DVH and distance to target histogram (DTH) based on the fact that the geometric information between PTV and OAR is closely related to DVH, this study aims to establish a multi-OAR geometry-dosimetry model through deep learning to achieve DVH prediction. Distance to target histogram (DTH) is chosen to measure the geometrical relationship between PTV and OARs. In the proposed method, stacked denoise auto-encoder (SDAE) is used to reduce the dimension of the extracted DTH and DVH features, and then one-dimensional convolutional network (one-DCN) is used for the correlation modeling. This model can predict the DVH of multiple OARs based on the individual patient’s geometry without manual removal of radiation plans with outliers. The average prediction error of the measurement focusing on the left lung, right lung, heart, spinal cord was less than 5%. The predicted DVHs could thus provide accurate optimization parameters, which could be a useful reference for physicists to reduce planning time.