ESTRO 2025 - Abstract Book

S3379

Physics - Machine learning models and clinical applications

ESTRO 2025

2. Milewski A, Li G. Stability and Reliability of Enhanced External-Internal Motion Correlation via Dynamic Phase-Shift Corrections Over 30-min Timeframe for Respiratory-Gated Radiotherapy. Technol Cancer Res Treat. 2022;21:15330338221111592.

1416

Digital Poster Clinical application of multiple target volumes auto-segmentation models which can accelerate the process of All-in-One radiotherapy Luqi Wang 1 , Ran Peng 1 , Xuemin Li 1 , Mingqing Wang 1 , Siyi Lu 1 , Yunsong Ji 2 , Wei Zhang 3 , Lecheng Jia 4 , Mengying Yang 5 , Hao Wang 1,6 1 Radiation Oncology, Peking University 3rd Hospital, Beijing, China. 2 Radiation Oncology, Chengyang District Hospital, Qingdao, China. 3 Radiotherapy Business Unit, Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China. 4 Innovative Medical Equipment, United Imaging Research Institute, Shenzhen, China. 5 Intelligent Imaging, United Imaging Research Institute, Beijing, China. 6 Cancer Center, Peking University 3rd Hospital, Beijing, China Purpose/Objective: To assess the clinical applicability of computed tomography (CT)-based multiple target volumes auto-segmentation models for patients with rectal cancer. Material/Methods: This study involved 282 patients diagnosed with T 3-4 stage rectal cancer, treated at Peking University Third Hospital between June 2012 and April 2024. The CT images from these patients were manually segmented to encompass the clinical target volume 1 (CTV45), clinical target volume 2 (CTV50), and the gross tumor volume (GTV). Utilizing the nnU-Net framework, which is based on deep learning for medical image segmentation, a comprehensive training regimen consisting of 300 iterations was conducted on a dataset comprising 272 patient cases (training set). Subsequently, the trained models were employed to assess auto-segmentation on an external test set consisting of 10 patients. The segmentation outcomes were quantitatively assessed using several metrics, including the time taken, the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), 95% Hausdorff Distance (HD95), Average Surface Distance (ASD), and Average Symmetric Surface Distance (ASSD). Results: There was a strong correlation between the auto-segmentation outcomes and the manual segmentation, as evidenced by the following metrics: the DSC of CTV45, CTV50, and GTV were 0.90, 0.86, and 0.71; the HD were 17.08, 25.48, and 79.59 mm; the HD95 were 4.89, 7.33, and 56.49 mm; the ASD were 1.24, 1.58, and 11.61 mm; and the ASSD were 1.23, 1.90, and 6.69 mm. Two clinicians manually segmented the target volumes for these ten patients, and the time required for the automatic segmentation process, including fine-tuning, was significantly less than that needed for manual segmentation (Doctor A: (164.0 + 13.4) seconds vs. 1514.7 seconds; Doctor B: (620 + 12.1) seconds vs. 1942.9 seconds). This efficiency translates to a time savings of 67.5% to 88.3%, thereby substantially alleviating the workload associated with target volumes segmentation for the clinicians.