ESTRO 2024 - Abstract Book

S4351

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2024

2486

Digital Poster

Impact of patient specific deep learning OAR segmentation on accumulated online adapted MRgRT dose

Yuqing Xiong 1 , Maria Kawula 1 , Marvin Ribeiro 1 , Sebastian Marschner 1 , Stefanie Corradini 1 , Claus Belka 1,2,3 , Guillaume Landry 1 , Moritz Rabe 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, Munich, Germany. 3 Bavarian Cancer Research Center, (BZKF), Munich, Germany

Purpose/Objective:

Online adaptation is crucial in magnetic resonance imaging guided radiotherapy (MRgRT) for lung cancer. However, the time-consuming re-contouring of organs at risks (OARs) on daily MR images suffers from inter- and intra-observer variabilities and prolongs treatment. This underscores the need for a more time efficient solution such as deep learning-based auto-segmentation. While networks trained on a large population may offer generic segmentation results, fine tuning patient specific (PS) models by exploiting available expert-delineated planning MR images (pMRI) has been shown to provide better results for fraction images [1].

In this study, we evaluate the impact of directly using PS segmentation models for online plan adaptation of lung cancer patient by evaluating accumulated doses.

Material/Methods:

Pre-trained population baseline models (BM, 3D U-Nets) for the left lung, right lung, heart, aorta, spinal canal, and esophagus [2] served as starting point. Thus far, for three patients who received 3-5 treatment fractions, we performed PS training by refining the BM using the pMRI and corresponding clinical segmentations for 500 epochs. Subsequently, OAR contours for each fraction generated by the BM and PS models were imported into an MRgRT treatment planning system (VR-SingleBox, Viewray, USA), along with the clinically used manual contours. For target structures, the clinical contours were used in all cases. The online adaptive doses were re-optimized using both models’ contours and the same clinical objective functions. The three sets of fraction doses (clinical, BM- and PS derived) were accumulated on the pMRI using a previously validated workflow [3]. For this comparison, we evaluated PTV D 95% , D 50% and GTV D 98% , D 50% , lung V 20Gy , as well as D 0.1cc for the OARs in proximity of the target. The clinical contours on pMRI were used to calculate all DVHs.

Results:

The BM encountered challenges in accurately contouring the lung when the tumor was at the thorax wall. In contrast, the PS model was able to overcome this issue (Figure 1).

Preliminary findings indicated that dose re-optimization using PS model generated contours showed smaller overall differences than BM contours compared to the clinical plan (differences smaller than 1.5 Gy and 1 ccm) (Table 1) and