Delopoulos Nikolaos, Marschner Sebastian, Lombardo Elia, Ribeiro Marvin F, Rogowski Paul, Losert Christoph, Winderl Tobias, Albarqouni Shadi, Belka Claus, Corradini Stefanie, Kurz Christopher, Landry Guillaume
Department of Radiation Oncology, LMU University Hospital, LMU Munich, Munich, 81377, Germany.
German Cancer Consortium (DKTK), partner site Munich, a partnership between DKFZ and LMU University Hospital Munich, Munich, 81377, Germany.
Phys Imaging Radiat Oncol. 2025 Aug 5;35:100819. doi: 10.1016/j.phro.2025.100819. eCollection 2025 Jul.
Magnetic resonance imaging-guided radiotherapy (MRgRT) facilitates high accuracy, small margins treatments at the cost of time-consuming and labor-intensive manual delineation of organs-at-risk (OARs). Auto-segmentation models show promise in streamlining this workflow. This study investigates the clinical applicability of a set of thoracic OAR segmentation models for baseline treatment planning in lung tumor patients. We investigate the use of the models for treatment at a 0.35 T MR-linac, assess their potential to reduce physician workload in terms of time savings and quantify the extent of required manual corrections, providing insights into the value of their integration into clinical practice.
Deep-learning based auto-segmentation models for 9 thoracic OARs were integrated into the MRgRT workflow. Two groups of 11 lung cancer cases each were prospectively considered. For Group 1 auto-segmentation contours were corrected by physicians, for Group 2 manual contouring according to standard clinical workflows was performed. Contouring times were recorded for both. Time savings between the groups as well as correlations of the extent of corrections to correction times for Group 1 patients were analyzed.
The model performed consistently well across all Group 1 cases. Median contouring times were reduced for six out of nine OARs leading to a reduction of 50.3 % or 12.6 min in median total contouring time.
Feasibility of auto-segmentation for baseline treatment planning at the 0.35 T MR-linac was shown with significant time savings demonstrated. Time saving potential could not be estimated from model geometric performance metrics.
磁共振成像引导放疗（MRgRT）有助于实现高精度、小边界的治疗，但代价是需要耗时且费力地手动勾画危及器官（OARs）。自动分割模型有望简化这一工作流程。本研究调查了一组用于肺癌患者基线治疗计划的胸部OAR分割模型的临床适用性。我们研究了这些模型在0.35 T MR直线加速器上用于治疗的情况，评估它们在节省时间方面减少医生工作量的潜力，并量化所需手动校正的程度，从而深入了解将其整合到临床实践中的价值。
将基于深度学习的9种胸部OAR自动分割模型整合到MRgRT工作流程中。前瞻性地考虑了两组，每组11例肺癌病例。对于第1组，医生对自动分割轮廓进行校正；对于第2组，按照标准临床工作流程进行手动轮廓勾画。记录两组的轮廓勾画时间。分析两组之间的时间节省情况以及第1组患者校正程度与校正时间的相关性。
该模型在所有第1组病例中表现一致良好。9个OAR中有6个的中位轮廓勾画时间减少，导致中位总轮廓勾画时间减少50.3%或12.6分钟。
结果表明，在0.35 T MR直线加速器上进行基线治疗计划时，自动分割是可行的，且显著节省了时间。无法从模型几何性能指标估计节省时间的潜力。
