Dose prediction for protontherapy using neural networks

Proton therapy is an emerging modality of radiotherapy, which is one of the main cancer treatments. Compared to conventional radiotherapy, it has the promising potential sparing more healthy tissues while ensuring that the tumor gets the prescribed dose. Both modalities require a treatment plan, which is nowadays done in a semi-automatic manner, but still requires several hours of manual works of physicians (for the contouring and prescription) and by the dosimetrist (to generate the plan). Ongoing research aims to speed up the process and tends toward a complete automatic workflow to allow the generation of a plan along the treatment. This Master thesis contributes to this aim by proposing a model for dose prediction for proton therapy using a UNet architecture. The dosimetric quality of the model was evaluated by comparing predictions with dose delivered in plans generated by a dosimetrist. The method proposed in this study is accurate (less than 4% of mean error between the clinical metrics of the prediction and ground truth, expressed in percent of the highest prescription dose), fast and achievable through dose mimicking.