S655
Clinical – Head & neck
ESTRO 2026
Purpose/Objective: The absence of standardized protocols and
and 2.8 (DARTBOARD, n=50) years after completing radiotherapy. The 3- and 5-year risks of solitary elective nodal recurrence were both 0% (n=75/43 followed for a minimum of 3/5 years). The 3-year cumulative incidence of local recurrence, regional recurrence and distant metastasis were 9.5%, 4.3% and 11% respectively. The 3-year and 5-year overall survival probabilities were 89% and 87% respectively, while 3- and 5-year progression-free survival probabilities were 78% and 74% respectively. There was no significant decline in composite MD Anderson Dysphagia Index scores after treatment, with a mean of 84.9 at 12 months. Standard ENI planning target volumes received mean V56Gy, V40Gy, and V20Gy of 35.9%, 49.4%, and 67.9%, respectively. Conclusion: These long-term results suggest that omission of ENI is safe, with minimal risk of solitary elective nodal recurrence. Randomized evidence is needed to support this promising paradigm before implementation in a non-trial setting. References: 1. Sher DJ, Moon DH, Vo D, Wang J, Chen L, Dohopolski M, et al. Efficacy and Quality-of-Life Following Involved Nodal Radiotherapy for Head and Neck Squamous Cell Carcinoma: The INRT-AIR Phase II Clinical Trial. Clinical Cancer Research. 2023 Sep 1;29(17):3284–91.2. Sher DJ, Avkshtol V, Lin MH, Wang J, Chen L, Liao CY, et al. Impact of daily adaptive head and neck radiotherapy on toxicity and quality-of-life: results of the DARTBOARD phase II randomized trial. JNCI: Journal of the National Cancer Institute. 2025 Aug 20;00(00),1-7. Keywords: INRT, ENI omission, AI, radiomics Development & validation of a prospective radiomics-clinical signature for locoregional recurrence in locally advanced head and neck cancer Balu Krishna Sasidharan 1 , Amal Joseph Varghese 1 , Hasan Shaikh 1 , Rajendra Benny Kuchipudi 1 , Ezhil Parvath Sindhanai 1 , Julia Priyadarshini Rao 1 , Vijayshree Chandru 1 , Jino Wilson Victor 1 , Simon P Pavamani 1 , Devadhas Devakumar 2 , Rajesh Isiah 1 , Leonard Wee 3 , Digital Poster Highlight 4315 Andre Dekker 3 , Hannah Mary T Thomas 1,4 1 Quantitative Imaging Research and Artificial Intelligence Lab, Department of Radiation Oncology, Unit 2, Chriistian Medical College, Vellore, India. 2 Department of Nuclear Medicine, Chriistian Medical College, Vellore, India. 3 Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre, Maastricht, Netherlands. 4 Biomedical Informatics Unit, Chriistian Medical College, Vellore, India
prospective data is often reported as a limitation in radiomics. This study aimed to address that to develop and validate a CT-based radiomics signature for prediction of loco-regional tumour recurrence (LRR) for locally advanced head and neck squamous cell carcinoma (LA-HNSCC) based on prospective, standardized planning CT and clinical data from a large tertiary care center in India. Material/Methods: We prospectively enrolled 1,467 LA-HNSCC patients treated between 2020-2024 and included patients who received primary radiochemotherapy, minimum follow-up of one year following end of radiotherapy. All patients had contrast-imaging planning CT using a standard imaging acquisition protocol and harmonized across two CT simulators (Siemens Somatom and GE Discovery). LRR is defined as local or regional disease as determined on CT imaging and clinical examination at follow-up.From clinician defined primary gross tumor volume (GTV).107 quantitative imaging features were extracted using PyRadiomics First, a radiomics model was trained using CT imaging features following which demographic and clinical parameters were included (Table 1). A final signature was derived using clinical and radiomics using repeated 5-fold cross-validation on the discovery cohort. Seven feature selection (i.e LASSO, SelectKBest, Particle Swarm Optimization, Whale Optimization Algorithm, Grey Wolf Optimizer, Genetic Algorithm, Simulated Annealing) and five machine learning classifiers (i.e. Logistic Regression, Naive Bayes, Support Vector Machine, Decision Tree, Random Forest) were evaluated. All models used 80/20 train-test separation while maintaining the LRR event-rate across the training and hold-out datasets. The performance was evaluated on an internal hold- out dataset using area under receiver operating curve over 1,000 bootstrap iterations with 95% confidence
interval. Results:
Made with FlippingBook - Share PDF online