Exploration of dose trade-offs for proton therapy treatments using deep learning models

Proton therapy is an emerging alternative to conventional radiotherapy which is the most common treatment modality for cancer. Proton therapy allows a better sparing of the healthy tissues surrounding the tumor. To plan a treatment, medical physicians and physicists must do a time-consuming work through a specialized software choosing the best trade-off between tumor coverage and sparing of healthy tissues for a given patient. In recent years, deep learning has been proposed to automate the treatment planning process. However, most existing models only predict one dose distribution which is not always the most optimal for a given patient. An emerging method is Pareto optimization. It consists in allowing the medical physicists to choose the best trade-off for a patient through changing the inputs in a deep learning model. This master thesis presents a deep learning model for the prediction of a dose distribution with a scaled dose in an organ. The organ to scale and the scaling can be set as an input of the model. The deep learning architecture used is the HD U-Net. The predicted dose distribution is then analyzed and made clinically deliverable through dose mimicking. This thesis focuses on head and neck cancer and the possible late effects due to a proton therapy treatment. The predictions are thus evaluated through the computation of the Normal Tissue Complication Probability of a common complication: xerostomia.