S2495
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
1 Radiotherapy Department, Instituto Nacional de Câncer - INCA, Rio de Janeiro, Brazil. 2 Department of Medical Physics, Instituto Nacional de Câncer - INCA, Rio de Janeiro, Brazil Purpose/Objective: Radiation pneumonitis (RP) is a dose-limiting toxicity in thoracic radiotherapy. While V20 and mean lung dose (MLD) are established predictors in lung cancer, their applicability to esophageal cancer remains unclear due to anatomical differences. This study characterized the dosimetric pattern of esophageal cancer radiotherapy, identified predictors of RP, and developed a machine learning-based personalized risk prediction model. Material/Methods: Retrospective cohort of 53 patients with stage II-III esophageal carcinoma treated with neoadjuvant chemoradiotherapy (50.4 Gy/28 fractions, carboplatin AUC 2 + paclitaxel 50mg/m ² weekly) at a national cancer institute (2018-2023). Clinical variables: age, sex, BMI, FEV1, smoking, histology, tumor location (gastroesophageal junction [GEJ] vs. other). Dosimetric parameters: V5, V10, V20, V30, MLD, spared volume <5 Gy (VS5), V5/V20 ratio. Primary endpoint: RP grade ≥ 1 (CTCAE v5.0) within 12 months by serial CT scans. Statistical analyses: univariate tests, multivariate logistic regression, ROC curves, Kaplan-Meier analysis. Machine learning: logistic regression, gradient boosting, random forest, neural network with 5-fold cross-validation. Results: RP incidence: 52.8% (28/53); median time 124 days; grade distribution: 1 (60.7%), 2 (28.6%), 3 (7.1%), 4 (3.6%). Dosimetric pattern: V5 median 52.8% (IQR 45.6- 62.2%), V10 37.4% (31.2-44.8%), V20 16.8% (13.7- 19.8%), V30 8.5% (6.8-11.2%), MLD 9.8 Gy (8.5-11.3 Gy), demonstrating low-dose large-volume exposure. Dosimetric paradox: patients WITH RP had LOWER doses (V20: 15.8% vs. 18.0%, p=0.56; MLD: 9.6 vs. 10.1 Gy, p=0.77). All dosimetric parameters showed poor discrimination (AUC ≤ 0.56). GEJ location was the only significant predictor (68.4% vs. 42.1%, p=0.04). Multivariate analysis: GEJ OR 1.97 (95% CI 1.05-3.71, p=0.04); V20 OR 0.98 (0.85-1.13, p=0.78). Machine learning: logistic regression achieved best performance (AUC 0.587±0.216). Feature importance: age (14.9%), V20 (14.0%), FEV1 (13.0%), BMI (10.1%), GEJ (3.5%), confirming multifactorial phenomenon. Excel-based risk calculator stratifies patients into low/moderate/high risk categories.
Low-dose large-volume pattern (V5=52.8%, V10=37.4%) and dosimetric paradox in esophageal cancer.
Machine learning feature importance and Excel-based personalized risk calculator for radiation pneumonitis prediction. Conclusion: Esophageal cancer radiotherapy exhibits a distinct low-dose large-volume pattern (V5=52.8%, V10=37.4%) contrasting with lung cancer. Anatomical location supersedes volumetric dosimetry as predictor. Total volumetric metrics fail to capture spatial dose distribution. The personalized risk calculator enables tailored surveillance. These findings challenge extrapolation of lung cancer predictors to esophageal cancer and advocate a paradigm shift from "how much dose" to "where is the dose."
Made with FlippingBook - Share PDF online