S563
Clinical – Head & neck
ESTRO 2026
models. Material/Methods: A total of 1,105 NPC patients treated at Hunan Cancer Hospital (Changsha, China) were retrospectively included and assigned to a training set (n = 766) and a testing set (n = 339). Primary tumour (GTVp) contours were transferred from planning CT to MRI by rigid registration. The involved lymph nodes (GTVn) were automatically segmented using the nnU-Net algorithm trained on manual delineations from a subset of 100 patients in this cohort. The combined volume (GTVtot) was created by merging GTVp and GTVn. Radiomics features were extracted using PyRadiomics from GTVp, GTVn and GTVtot on T1-weighted, T2-weighted, and contrast-enhanced T1-weighted MRI. Highly correlated features were first removed using Pearson correlation (|r| >= 0.8), after which bootstrapped forward selection was used before multivariable Cox modelling. Model performance was evaluated with the concordance index (C-index). For each endpoint, the best-performing MRI radiomics model was compared with clinical-only and CT-based radiomics models, using the same modelling pipeline. Significant differences in predictive performance were tested using likelihood-ratio tests. Results: In the testing set, best-performing MRI-based models demonstrated higher C-indices across all endpoints than both clinical-only and CT-based radiomics models (all p < 0.05; see Table 1 for model covariates and C- index comparisons). The largest gains were for LC and LRC (improvement in C-index = 0.12 for both). The models also demonstrated strong performance in risk stratification (Figure 1). Regarding OS, PFS and DC the models stratified patients into distinct high- and low- risk groups of predicted risk (all p < 0.01). This stratification was particularly robust for the radiomics- only DC model, which yielded a hazard ratio of 4.70 for 5-year distant control between high- and low-risk groups.
The median FU of patients treated by U-ENI is 42 months and all patients have a follow-up >12 months at the moment of this analysis. Of all patients received U-ENI (n=82), only 2 patients developed CRF (2.4%), 5 and 24 months after radiotherapy and received radiotherapy alone or neck dissection and postoperative radiotherapy to the CL neck, combined with cisplatin, respectively. Both patients are still alive and disease-free. Patients with unilateral treatment showed significantly lower incidence of acute G ≥ 2 dysphagia, G3 dysphagia (feeding-tube), late G ≥ 2 dysphagia and xerostomia (37%, 11%, 12%, and 11%, respectively), compared to a historical cohort, composed by propensity score matching (83%, 48%, 37%, and 48%, respectively) (p<0.05 for all items). Conclusion: In this study, first of its type, SPECT/CT-based SNP to guide U-ENI in HNSCC patients (who have an indication for B-ENI) seems safe and effective. It resulted in reducing the proportion of patients treated to both sides of the neck from 100% to only 9%, and significantly reduced the incidence of acute and late toxicity, compared to a historical cohort, without increasing the rate of CRF. Keywords: SPECT- and SNP-guided unilateral elective nodal RT Improved outcome prediction in nasopharyngeal carcinoma using pre-treatment MRI radiomics Guanzhi Zhou 1,2 , Baoqiang Ma 1,3 , Mark L. Frederiks 1 , Jie Dai 2 , Suzanne P. M. de Vette 1 , Lisanne V. van Dijk 1 , Yingrui Shi 2 , Pei Yang 2 , Johannes A. Langendijk 1 , Nanna M. Sijtsema 1 1 Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands. 2 Department of Head and Neck Radiation Oncology, Hunan Cancer Hospital, Changsha, China. 3 Image Sciences Institute, University Medical Center Utrecht, Utrecht, Netherlands Purpose/Objective: Nasopharyngeal carcinoma (NPC) patients with the same TNM stage often experience different outcomes despite standard treatment reflecting tumour heterogeneity. MRI, with its superior soft-tissue contrast, is particularly effective in detecting extranodal extension and the extent of skull-base invasion, which are not fully reflected by stage-based risk stratification. This retrospective study aimed to develop pre-treatment MRI-based radiomics models to predict overall survival (OS), progression-free survival (PFS), local control (LC), locoregional control (LRC), and distant control (DC), and to compare their performence with clinical-only and CT-based radiomics Poster Discussion 1253
Made with FlippingBook - Share PDF online