Deep learning to predict the optimal beam intensities for proton therapy treatments

Radiotherapy is one of the main treatments for cancer. When compared to conventional radiotherapy with photons, proton therapy holds the potential to spare more healthy tissues while ensuring tumor coverage. In both modalities, a treatment plan needs to be generated. This task is time-consuming, and significant efforts have been invested in automating it. In particular, deep learning has already been employed to automate certain steps in this process. However, while deep neural networks can be used to predict the full dose distribution, their application has not yet been extended to the subsequent step in the planning process of proton therapy plans. This step involves determining the machine parameters, specifically the weights of the spots, that will most accurately achieve the predicted dose distribution. This thesis aims to explore potential solutions to fill this gap. A fully automated pipeline containing two deep neural networks was implemented. A first network predicts the dose of each individual beam. The second network should take this information together with other inputs to provide a prediction of the spot weights. Both networks employed a U-Net architecture and were trained on a database containing plans with the same beam configuration and prescription for 39 patients with esophageal cancer. If the obtained results were not satisfactory for direct use without further manual optimization, this work still represented the initial step toward achieving a fully automated workflow.