ESTRO 2024 - Abstract Book

S4088

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

[1] “PreciseART AS A DAILY TREATMENT PLAN QUALITY ASSURACE TOOL”, Accuray white paper, Froedtert & Medical College of Madison, 2017.

[2] Brock et al. “Use of image registration and fusion algorithms and techniques in radiotherapy: Report of the AAPM Radiation Therapy Committee Task Group No. 132”. Med Phys. 2017,44:e43-e76. https://doi.org/10.1002/mp.12256

2541

Proffered Paper

Personalized auto-segmentation models for adaptive fractionated MRgRT

Maria Kawula 1 , Sebastian Marschner 1 , Marvin Fernando Ribeiro 1 , Stefanie Corradini 1 , Claus Belka 1,2,3 , Guillaume Landry 1 , Christopher Kurz 1 1 LMU University Hospital, LMU Munich, Department of Radiation Oncology, Munich, Germany. 2 German Cancer Consortium (DKTK), partner site Munich, a partnership between DKFZ and LMU University Hospital, Munich, Germany. 3 Bavarian Cancer Research Center, (BZKF), Munich, Germany

Purpose/Objective:

The benefits of adaptive magnetic resonance guided radiation therapy (MRgRT) at MR linear accelerators (MR-Linacs) entail labor-intensive workflows, mostly due to the necessity of manual segmentation. This is particularly pronounced in fractionated treatments, where daily contour adjustments are required. State-of-the-art deep learning (DL) auto segmentation networks (called baseline models (BM) in this work) learn anatomical features shared among many patients; therefore, they produce generic delineations. This might be sub-optimal for patients undergoing fractionated irradiation, since manually segmented planning and prior fraction MR images (MRIs) exist but are not integrated into BMs. Moreover, the BM segmentations might differ from the institutional segmentation guidelines. Previous studies suggest that including patient-specific knowledge in DL models might be advantageous in segmentation tasks [1,2,3]. In this work, we explore methods leveraging expert-segmented initial planning as well as prior fraction MRIs to improve auto-segmentation on consecutive fraction days. The dataset consisted of 151 abdominal tumor patients treated at a 0.35T MR-Linac (MRIdian, ViewRay inc, USA). For each patient, there was one segmented planning and up to 5 segmented fraction MRIs (in total 151 planning and 215 fraction images). Our focus was on key abdominal organs-at-risk (OARs) including the aorta, bowel, duodenum, kidneys, liver, spinal canal, and stomach. For benchmarking and as a BM for the subsequent patient-specific (PS) models, state-of-the-art 3D U-Nets were trained on 107 planning images for each organ separately. The first type of PS models (PS BM F0 ) was created by fine-tuning the BMs with a patient's planning ("fraction 0") image. The second PS method explored the potential benefits of progressively fine-tuning PS BM F0 models with segmented fraction images (PS BM F0-4 ). We tested the presumed upper limit of this approach for patients undergoing five fractions, which is a common fractionation scheme at MR-Linacs. We fine-tuned the BMs with the planning and the first four fraction MRIs, which should lead to the most accurate segmentation for the fifth fraction. In the next two PS approaches, instead of Material/Methods: