Towards efficient online adaptive MRI-guided prostate radiotherapy workflows

Abstract

This thesis shaped the landscape of efficient online adaptive MRI-guided prostate radiotherapy workflows on the 1.5 T MR-linac. The first part of the thesis introduced DeepDose, a physics-based deep learning model for forward dose calculations of IMRT segments. Trained on various abdominal sites and randomized beam angles, it showed high agreement with the ground truth doses and successfully modeled the electron return effect. Next, DeepDose was extended to VMAT dose predictions using a broader set of tumor anatomies. It performed well across in- and out-of-distribution cases, but large input sizes slowed pre-processing and inference. As such, its current use is limited to offline settings or secondary checks. The second part of the thesis explored intrafraction motion management for prostate radiotherapy. First, a commercial comprehensive motion management algorithm was evaluated for detecting 3D prostate drifts using 2D cine bTFE MRI. Due to low image contrast and tracking quality, the development of a T2-weighted 2D cine sequence was recommended for improved motion monitoring. In addition,a clinical sub-fractionated workflow with rigid plan adaptations was retrospectively evaluated. The dose-volume parameter evaluation using deformable contour registration showed good dosimetric agreement, confirming effective motion handling. Finally, rigid and deformable intrafraction dose accumulation methods using time-resolved 3D cine MRI were compared. While both approaches showed similar CTV and OAR doses, deformable dose accumulation achieved higher image similarity in cases with large deformations. Therefore, it should be preferred for accurate voxel-wise dose reconstruction and future real-time plan adaptation.