Prior knowledge-aware deep learning auto-segmentation for MRI-guided radiotherapy

Rädler, Maria (2025): Prior knowledge-aware deep learning auto-segmentation for MRI-guided radiotherapy. Dissertation, LMU München: Medizinische Fakultät

Abstract

Thanks to combined magnetic resonance (MR) linear accelerators (LINACs) called MR- LINACs, the integration of image guidance and online plan adaptation into radiotherapy became feasible. Dose-free magnetic resonance imaging (MRI), daily plan adjustments, and gating were shown to improve treatment outcomes and reduce side effects. However, treatment adaptation entails time-consuming segmentation of organs at risk (OARs) and target volumes, not only on the pre-treatment planning MRI but also on all daily fraction images. Quick segmentation of fraction images is crucial to minimize patient waiting time for irradiation. To improve and speed up fraction segmentation compared to conventional auto-segmentation models, expert knowledge from the planning phase could be leveraged. The aim of this work was to investigate deep learning (DL) auto-segmentation methods for cancer patients receiving conventional radiotherapy or magnetic resonance-guided radiation therapy (MRgRT). The main focus was on personalized models to enhance the segmentation in fractionated online adaptive treatments. The first paper analyzed the impact of DL contours on dose optimization in conventional radiotherapy of prostate cancer patients. It investigated the possible correlation between contour quality and the quality of treatment plans optimized using these contours. The study concluded that networks achieving state-of-the-art segmentation performance predict contours that lead to satisfactory dose distribution in most investigated cases. No strong correlations were found between the geometric and dosimetric metrics. The second study explored personalized models generated by fine-tuning conventional DL population models with the manually delineated planning MRI of a patient. The target group were prostate cancer patients undergoing MRgRT. Personalized models effectively learned organ shapes as defined on the planning image for each patient. They were particularly beneficial for target volume segmentation and for patients with unusual anatomies. The third study investigated networks for combined image registration and segmentation as an alternative to personalized models from the second paper. These networks were trained to register the planning and fraction MRIs and propagate the planning expert contours to the daily anatomy. The registration-based networks were successful in prostate clinical target volume (CTV) segmentation. The latter is difficult to segment with population models due to the individual shape but does not change significantly throughout the weeks. Personalized models performed better than registration networks for OARs undergoing larger changes. The last study explored further options for personalized training. It analyzed the impact of population models on patient-specific segmentation networks for abdominal OARs. The study investigated the adjustment of personalized models to the patient’s anatomy from the planning day and optionally from additional fractions. It also explored whether training from scratch (i.e., without the population model) using only the segmented planning MRI is sufficient to create a personalized model. The study showed that by fine-tuning population models with expert delineations of a given patient (planning or planning plus previous fractions), the models predict clinically usable contours with little to no corrections needed. Using single patient data was insufficient to develop robust personalized models. Regardless, all personalized models improved with updates to the prior fraction anatomies. Summarizing, personalized models created by fine-tuning population models with expert-segmented images of a given patient performed best among all investigated alternatives. Training times of personalized models were short enough for clinical implementation. By reducing the necessity of manual contour corrections, personalized models have the potential to shorten treatment adaptation and reduce inter and intra-observer segmentation variability in MRgRT at MR-LINACs.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Themengebiete: 600 Technik, Medizin, angewandte Wissenschaften ; 600 Technik, Medizin, angewandte Wissenschaften > 610 Medizin und Gesundheit
• Fakultäten: Medizinische Fakultät
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 19. Dezember 2025
• 1. Bericht­erstatter:in: Kurz, Christopher