Development of daily automated treatment adaptation for MRI-guided radiotherapy of head and neck cancer patients

Abstract

Adaptive radiotherapy (ART) using an MR-Linac enables treatment plans to be adapted daily to account for any observed anatomical changes. To adapt these plans effectively, the regions of interest (ROIs) must be re-delineated on the daily scans, which requires fast organ delineation. Manual delineation, particularly for head and neck cancer (HNC) patients, is subjective and time-intensive task. Following re-delineation, a new treatment plan must be optimised to accommodate the observed anatomical changes. Manual treatment planning requires manually selecting and adjusting the optimisation parameters, a process that is both time-consuming and heavily dependent on the treatment planner's experience. These manual interventions frequently involve trial-and-error iterations, potentially prolonging the planning process and introducing variability in plan quality. Automated contouring and planning methodologies offer a promising solution to these challenges. This thesis aimed to develop and clinically validate an automated approach for treatment adaptation. To accomplish this, a widely adopted network architecture incorporating established segmentation techniques was employed. Additionally, an automated treatment planning tool called Erasmus-iCycle, developed by Erasmus MC in Rotterdam, The Netherlands, was utilised to automatically generate treatment plans for HNC patients undergoing MR-Linac treatment. For the evaluation and validation of the automated structures, first an evaluation on purely geometric features was performed followed by a blind test conducted by a radiation oncologist to assess the clinical acceptability. Subsequently, the dosimetric impact was evaluated. The automated plans were assessed against dosimetric criteria, and the dose-volume histograms achieved by the automated plans were compared to those achieved by manual plans. The auto-segmentation model demonstrated high clinical accuracy and led to minimal dosimetric differences for most structures, all within a segmentation time of under 10 seconds. Additionally, the automated treatment planning yielded good results, producing dose distributions comparable to those of manual plans. Notably, the auto-plans exhibited significant sparing of normal tissue, especially for the brainstem. The integration of auto-segmentation and auto-planning led to the development of clinically acceptable treatment plans, reducing reliance on expert experience, and almost completely eliminating the necessity for manual adjustments and physical presence. This thesis demonstrates the potential of automating contouring and planning for HNC patients on MR-Linacs, showing effective automated contouring, especially for OARs, and comparable plan quality to manual methods. Further research is needed to address limitations such as broader patient population inclusion and clinical deliverability of automated plans, aiming to create an efficient workflow for clinical implementation. Overall, this research provides promising results for automating HNC treatment planning on MR-Linacs, with potential benefits for efficiency and consistency. Continued refinement and integration of automated tools hold promise for advancing radiation oncology and personalised therapy delivery.