DoseNet: Dose-adaptive prediction of the parotid glands deformation for radiotherapy planning

Parotid glands (PGs) toxicity caused by radiation-induced anatomy deformation occurs among a significant amount of patients with nasopharyngeal carcinoma treated with radiotherapy. Early prediction of PGs deformation is critical, as it can facilitate the design of treatment plans to reduce radiation-induced anatomical change in an adaptive radiotherapy workflow. Previous studies used CT images to model anatomical variation in radiotherapy. However, they did not consider the radiation dose received by the PGs which is correlated to the PGs volumetric change and can influence the anatomical variation. To address this issue, we propose DoseNet, a dose-adaptive PGs deformation prediction deep neural network, which utilizes the radiation dose and CT images to generate different anatomy predictions accommodating to the changing dose. Specifically, we use parted dose input and multi-scale cross attention to reinforce the integration of PGs anatomy and the dose received by PGs, and present a novel data augmentation method to remedy the shortcoming of the skewed data distribution of the radiation dose. Besides, to help design improved treatment plans, a novel metric termed dose volume variation (DVV) curve is developed to visualize the predicted volumetric change in respect to the dose variation of the PGs. We verify the effectiveness of our method on a dataset collected from a collaborative hospital. The experiment results show the proposed DoseNet outperforms the state-of-the-arts on the dataset and attains a Dice coefficient of 82.2% and a relative volume difference of 12.2%. The code is available at https://github.com/mkdermo/DoseNet.