Logo: to the web site of Uppsala University

In radiotherapy, treatments are frequently distributed over multiple weeks, and the radiation dose delivered across several sessions. A significant hurdle in this approach is the anatomical changes that occur between the planning stage and subsequent treatment sessions, leading to uncertainties in the treatment. The MR-Linac system, which combines a linear accelerator with an MRI scanner, addresses this issue by allowing for daily adjustments to the treatment plan based on the patient's current anatomy. However, the process for making these adjustments, involving image fusion, re-contouring, and plan re-optimization, can be quite elaborate and time-consuming. This project aimed to identify opportunities within the daily treatment routine where machine learning and deep learning could streamline the process, thereby enhancing efficiency, with a focus on prostate cancer treatments due to their frequent occurrence at our facility. We leveraged deep learning to train patient-specific models for segmenting anatomical structures in daily MRI scans, matching the accuracy of existing deformable image registration techniques. Furthermore, we extended this concept to segmenting structures and predicting radiation dose distributions, offering a swift assessment of potential dose distribution before engaging in the more complex manual workflow. This could aid in selecting the most suitable adaptation method more quickly. Additionally, we developed motion models for intrafractional motion and for segmenting images at lower resolutions to facilitate a target tracking process. Throughout this project, we showed how machine learning and deep learning techniques could contribute to optimizing the daily MR-Linac workflow.