S537
Clinical – Head & neck
ESTRO 2026
Purpose/Objective: Accurate target delineation in radiotherapy for nasopharyngeal carcinoma (NPC) is critical for ensuring tumor control and organ-at-risk protection. However, manual contouring is time-consuming and varies between observers. Deep-learning models trained on limited, physician-specific datasets are prone to unstable learning or complete tumor- recognition loss, known as the blackout phenomenon. This study introduces a delay-stopping training framework to overcome this problem, elucidates the underlying trade-off phenomenon, and validates the approach using a semi-supervised EfficientNetV2 model to improve the clinical reliability and efficiency of AI-assisted contouring. Material/Methods: Retrospective CT datasets of NPC patients were collected from Chiayi Christian Hospital (IRB No. 2023043), including 109 cases delineated by two board-certified radiation oncologists. Gross Tumor Volume (GTV) and Clinical Target Volume (CTV) were contoured according to institutional and international guidelines.The proposed framework was implemented across three architectures (U-Net, DeepLabV3+, UNETR) and further evaluated using a semi-supervised EfficientNetV2 network that leveraged both labeled and unlabeled data to enhance CTV learning stability.Training dynamics were analyzed by decomposing the loss into foreground (tumor) and background components, revealing three characteristic learning phases: (1) initial learning, (2) trade-off, and (3) convergence. Optimization was performed with Adafactor and Adafactor with Gradient Centralization (AdafactorGC). A delay-stopping mechanism dynamically monitored class-specific loss, ensuring training continued until complete tumor- feature acquisition. Results: Loss-curve analysis confirmed that the trade-off phenomenon universally occurred across architectures and datasets in Figure1. Conventional early-stopping prematurely terminated training during this phase, leading to blackout (DSC^fore ≈ 0) in Figure2. The delay-stopping framework successfully guided models through unstable regions, maintaining tumor recognition even under extreme imbalance.Quantitatively, the framework achieved DSC 0.754 for GTV and DSC 0.872 for CTV, reducing HD95 by 17%. The semi-supervised EfficientNetV2 model demonstrated enhanced generalization, decreased inter-observer variability, and shortened manual correction time by 42%. Clinically, AI contours generated under this framework required minimal editing, supporting integration into adaptive or physician-specific planning workflows.
follow up of 10 years, the 5 year local and regional failure were favorable for pediatric vs adult group (1.8% vs 8.7% [p=0,06] and 0% vs 8.7% [p=-.03], respectively). However, distant metastasis was the predominant pattern of failure, and the 5-year distant metastasis rate for pediatric vs adult NPC was comparable (11.4% vs 12.4%, p=0.84), which resulted in no statistically significant differences in EFS (85.2% vs 84.9%, p=0.37) or OS (90.7% vs 89.6% vs 90.7%, p=0.38) between pediatric and adult NPC patients. On multivariable analysis, WHO histology and ECOG performance status, but not age group, independently predicted outcomes. Conclusion: Pediatrics with NPC achieved favorable locoregional control than adults, yet both groups remained equally vulnerable to distant relapse, leading to similar OS. These findings reinforce the need to further study systemic treatment intensification to reduce distant failure and further improve long-term outcomes. References: 3. Petit C, Lee AWM, Ma J, et al. Role of chemotherapy in patients with nasopharynx carcinoma treated with radiotherapy (MAC-NPC): an updated individual patient data network meta- analysis. Lancet Oncol. 2023;24(5):611-623. doi:10.1016/S1470-2045(23)00163-8.4. Tsai M-H, Wu S-Y, Lu H-H, et al. Improved overall survival is associated with adjuvant chemotherapy after definitive concurrent chemoradiotherapy for N3 nasopharyngeal cancer. Sci Rep. 2022;12(1):13390. doi:10.1038/s41598-022-16422-w5. Chen L, Hu C-S, Chen X-Z, et al. Adjuvant chemotherapy in patients with locoregionally advanced nasopharyngeal carcinoma: long-term results of a phase 3 multicentre randomised controlled trial. Eur J Cancer. 2017;75:150–158. doi:10.1016/j.ejca.2017.01. Keywords: Adult, Pediatric Nasopharyngeal cancer Digital Poster 391 A New Reliable Training Framework for AI Contouring: Enhancing Model Stability in Limited- Data Nasopharyngeal Carcinoma Radiotherapy HsiaoJu Fu 1,2 , ChihChia Chang 1,3 , Wei-Min Liu 4 1 Department of Radiation Oncology, Ditmanson Medical Foundation Chia-Yi Christian Hospital, Chiayi, Taiwan. 2 Department of Mechanical Engineering and Advanced Institute of Manufacturing with High-Tech Innovations, National Chung Cheng University, Chiayi, Taiwan. 3 Department of Nursing, Chung Jen Junior College of Nursing, Health Science and Management, Chiayi, Taiwan. 4 Department of Computer Science and Information Engineering and Advanced Institute of Manufacturing with High-tech Innovations, National Chung Cheng University, Chiayi, Taiwan
Made with FlippingBook - Share PDF online