ESTRO 2025 - Abstract Book

S1010

Clinical – Head & neck

ESTRO 2025

2019

Mini-Oral Integrating CT image features improves clinical prediction models for outcome in nasopharyngeal carcinoma patients treated with (chemo)radiation Guanzhi Zhou 1,2 , Baoqiang Ma 1 , Yan Li 1 , Pei Yang 2 , Yingrui Shi 2 , Arjen van der Schaaf 1 , Lisanne V. van Dijk 1 , Johannes A. Langendijk 1 , Nanna M. Sijtsema 1 1 Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands. 2 Department of Radiation Oncology, Hunan Cancer Hospital, Changsha, China Purpose/Objective: Predicting outcome for NPC patients using only clinical information is challenging due to tumor heterogeneity. CT imaging captures detailed anatomical structures, providing valuable insights into tumor characteristics and spatial relationships with surrounding tissues. This study aimed to investigate whether integrating CT image features could improve outcome prediction in terms of overall survival (OS), progression-free survival (PFS), local control (LC), and distant control (DC) for NPC patients. Material/Methods: The study population consisted of 1,355 patients with locoregionally advanced NPC treated with (chemo)IMRT at Hunan Cancer Hospital from January 2013 to December 2017. Eleven clinical variables were included and contrast enhanced CT scans were available for all patients. Gross Tumor Volumes (GTVs) of the primary tumor (GTVp) and pathologic lymph nodes (GTVn) were used as inputs for both radiomics and deep learning models. PyRadiomics was used to extract 107 radiomics features, including shape, first-order intensity, and texture features. To address overfitting and multicollinearity, feature pre-selection was performed based on Pearson correlation. Selected features were then refined using bootstrapped forward selection based on the Bayesian Information Criterion (BIC). Multivariable Cox regression models were built on the final selected features. For the deep learning approach, either a 3D ResNet18 or 3D DenseNet121 was used to effectively capture spatial features. This model was trained using CT images with GTVp, GTVn, or GTVp + GTVn as additional input channels, alongside clinical variables. Model performance was evaluated using the concordance index (c-index) and different models were compared using the log-likelihood test. Results: Incorporating CT data improved prediction performance for some endpoints compared to the clinical model. For LC, the radiomics model that incorporated clinical feature and GTVp texture features showed a significant improvement over the clinical model (c-index: 0.60 vs. 0.51, p<0.05), with Neighbor Grey Tone Differences Matrix Coarseness providing complementary information to tumor volume (see Figure). For DC, the 3D Resner18 model, which included clinical information and CT image features of GTVp + GTVn demonstrated an improvement, with the c-index increasing from 0.64 (95% CI: 0.57, 0.69) to 0.68 (95% CI: 0.59, 0.76). For OS and PFS, however, the improvements were not significant.