ESTRO 2025 - Abstract Book

S2443

Physics - Autosegmentation

ESTRO 2025

We retrospectively collected 3,500 computed tomography (CT) scans from 2,972 patients across four hospitals. A comprehensive set of 43 OARs categories was designed to accommodate all clinical scenarios in HNC radiotherapy. Our FM was trained on 2,874 CT scans and validated on an internal test cohort comprising 500 CT scans (250 non contrast CT (ncCT) and 250 contrast-enhanced CT (ceCT)). The accuracy of model was compared against three previously established algorithms. To further demonstrate its clinical utility, we evaluated the model's transfer learning capability on the SegRap dataset for the gross tumor volume (GTV) segmentation task. We evaluated the accuracy of each HNC OARs by calculating Dice Similarity Coefficient (DSC), Normalized Surface Dice (NSD), and Added Path Length (APL). Results: In the internal testing cohort, 81.8% (9/11) of anchor OARs achieved a mean DSC (mDSC) ranging from 85.09% to 96.56%, and 64.3% (9/14) of mid-level OARs achieved a mDSC ranging from 80.14% to 92.41%. Among small and hard (S&H) OARs, 66.7% (12/18) exhibited a mDSC ranging from 75.03% to 82.37%, with all S&H OARs achieving a mean APL (mAPL) below 100 voxels. Consistent performance was observed in ncCT and ceCT scans, with 83.7% (n=36) and 79.1% (n=34) of 43 OARs achieving a mDSC above 75.04%, respectively, and 95.3% (41/43) of OARs demonstrating a mAPL below 5000 voxels. Moreover, when the pretrained model was transferred to downstream gross tumor volume (GTV) delineation tasks, our method achieved a mDSC of 72.58% for primary GTV (GTVp) and 56.07% for node GTV (GTVnd), closely approaching the fully supervised method, which achieved mDSCs of 74.80% and 57.38%, respectively. Across two testing cohorts, our method outperformed four comparative algorithms, demonstrating significantly higher mDSC and lower mAPL. Conclusion: The proposed FM demonstrated accurate and robust segmentation of 43 OARs for HNC radiotherapy. Furthermore, the model effectively transferred to downstream GTV segmentation tasks, showcasing its versatility. This promising work holds significant clinical applicability, contributing to ongoing efforts to enhance radiotherapy planning and streamline clinical workflows. Digital Poster Deep Learning-Based Automatic Delineation of Nasopharyngeal Carcinoma Targets Using Prior Information from Original Plan Guanqun Zhou 1,2 , Ziquan Wei 3 , Yuxian Yang 1,2 , Yuxi Xiong 1,2 , Hua Li 3 , Lecheng Jia 3 , Guangyu Wang 1,2 , Xiaobo Jiang 1,2 , Feng Chi 1,2 , Ying Sun 1,2 1 Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, China. 2 Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, State Key Laboratory of Oncology in South China, Guangzhou, China. 3 Research cooperation department, Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China Purpose/Objective: The precision of target delineation for nasopharyngeal carcinoma (NPC) has significantly improved due to the application of deep learning mehtod. However, few studies have investigated the automatic delineation of cervical lymph nodes (GTVn) and clinical target volumes (CTVs) for NPC in adaptive radiotherapy (ART). This study proposes a deep learning-based automatic delineation algorithm that leverages prior information from the original plan to improve the accuracy during online ART. Keywords: Head and neck cancer; OAR; Foundation model 1763

Material/Methods: