ESTRO 2025 - Abstract Book

S1046

Clinical – Head & neck

ESTRO 2025

Material/Methods: A total of 1,808 iLNs from 442 N+ HNC patients treated over the past 12 years were collected and split into training (70%) and testing sets (30%) based on time. The clinical and CT radiomics models of Zhai et al. were tested and updated to optimize its performance in current expanded dataset. Predictive PET radiomics features were selected using forward selection based on clinical variables with or without the shape feature Least-axis-length and the CT image feature Correlation-of-GLCM and cox-regression models were generated in 1,000 bootstrap samples. The most frequently selected features were added to the clinical model. All model performances were evaluated by c index [95% CI]. The significancy of model improvement was tested by Likelihood Ratio-Test. Results: In the training set, 65 nodal failures occurred among 1,276 lymph nodes (5.1%); in the testing set, 22 failures occurred among 532 nodes (4.1%). The clinical model by Zhai et al. performed poorly in current testing set ( Figure 1). An updated clinical model including AJCC 8 tumor stage, tumor location, and packyears improved the c-index to 0.73 [0.66-0.78] in the training set, and to 0.63 [0.51-0.75] in the testing set. The shape feature Least-axis-length significantly improved the c-index to 0.83 [0.79-0.88] and 0.81 [0.72-0.89], outperforming the CT feature Correlation of-GLCM . While the PET feature Dependence Entropy-GLDM enhanced performance over clinical variables alone, its c index was lower than that of Least-axis-length in testing dataset. Neither CT nor PET features significantly enhanced the clinical model with Least-axis-length . Larger Least-axis-length indicated higher failure risk ( Figure 2 ).