Content-Aware Adversarial Network with Gradient-Enhanced Dose Rectification for Radiotherapy Dose Prediction

Generative adversarial networks (GANs) have accelerated radiotherapy plan-making by automatically predicting the dose distribution maps. Despite preserving more high-frequency details, GAN-based methods confront (1) a compromised dose accuracy for their sharper but distorted predictions and (2) ignore the content (dose) differences in the dose map. To solve these issues, we propose DRGAN, a content-aware GAN with gradient-enhanced dose rectification, to predict the dose distribution with both high accuracy and sharpness. Specifically, gaining a coarse prediction from GAN, a gradient-enhanced dose rectification module (GDRM) is devised to rectify it in a dose rectification branch by incorporating the explicitly captured gradient information from a gradient enhancement branch, thus ensuring the dose accuracy while preserving sharpness details. Furthermore, a content-aware discriminator with two expertized classifiers is designed to separately distinguish the fake and real dose distribution inner and outer the planning target volume, equipping the model with essential awareness of their content difference. Experiments on a clinical rectum cancer dataset demonstrate the superiority of our method.