Prediction of inter-fractional radiotherapy dose plans with domain translation in spatiotemporal embeddings

External beam radiation therapy fractions have become extremely complex and tedious procedures to plan, due to stringent requirements of delivering the highest radiation dose to the tumor while maximally avoiding organs at risk. However, due to anatomic and/or biological changes between fractions, dose re-optimization may be needed. Re-optimization is a time-consuming task which is typically triggered based on subjective visual assessment by an experienced physician. To address limitations in this process, we introduce a predictive framework which learns the evolution of tumor anatomy as well as inter-fractional dose delivery variations for head and neck cancers. First, joint low-dimensional discriminant embeddings maximizing the separation between responsive and non-responsive groups to external beam radiotherapy plans are constructed from deep neural networks in order to capture patient-specific dose modulations with respect to anatomical variations. Then, latent representations are fed to a domain-level adversarial network to translate observed anatomical changes into dosimetric variations, which aims to enforce local semantic consistency in the overall translation. Dose distribution trajectories are represented in a group-average piecewise-geodesic setting to handle anatomical variations during therapy, using a quadratic optimization to perform curve regression. At test time, an annotated baseline CT is projected onto the latent space and translated to dose domain, from which a spatiotemporal regression model is constructed using parallel transport trajectories defined from closest samples. This allows to predict dosimetry changes during the course of treatment. The model was trained on 337 cases and tested on 50 separate patients using sequential CT and associated dosimetry data, with the probabilistic framework yielding a Dice score of 92% and an overall dose difference of 1.2 Gy in organs at risk and tumor volume over the course of multi-day treatment course, with a 5% reduction in delivered fraction segments.