3715 - Prediction of Radiation Pneumonitis Using Dose and Computed Tomography Data: First Insights from the Prospective RTOG 0617 and REQUITE Trials
Presenter(s)
L. M. Reuter1,2, K. M. Kraus1,3, D. Pletzer1,2, S. Fischer1,2, D. Bernhardt1,3, S. E. Combs1,4, J. Schnabel2,5, and J. C. Peeken1,3; 1Department of Radiation Oncology, School of Medicine and Health and Klinikum rechts der Isar, Technical University of Munich (TUM), Munich, Germany, 2Institute of Machine Learning in Biomedical Imaging (IML), Helmholtz Zentrum München (HMGU) GmbH, German Research Center for Environmental Health, Neuherberg, Germany, 3Institute of Radiation Medicine (IRM), Helmholtz Zentrum München (HMGU) GmbH, German Research Center for Environmental Health, Neuherberg, Germany, 4Institute of Radiation Medicine (IRM), Department of Radiation Sciences (DRS), Helmholtz Center Munich, Munich, Germany, 5School of Computation, Information and Technology, Technical University of Munich (TUM), Munich, Germany
Purpose/Objective(s): Radiation pneumonitis (RP) is a common and relevant side effect of thoracic radiotherapy. The primary objective of this study is to predict symptomatic RP grade =2 using radiomics and dosiomics modeling with subsequent external validation.
Materials/Methods: Prospective data of 698 multicenter lung cancer patients were utilized from the NRG Oncology/RTOG 0617 trial (n=441) as the training cohort and from the REQUITE study (n=257) as an external test set. The training cohort (CTCAE v3, RP =2: 15%) consisted exclusively of stage III patients treated with 60 or 74 Gy. The test cohort (CTCAE v4, RP =2: 7%) comprised stages I-III with doses between 45-70 Gy. Radiomics and dosiomics features were extracted from pre-treatment CT images, 3D dose volumes, and 3D EQD2 volumes according to IBSI guidelines, using label maps for the GTV, PTV, PTV+2 cm, and lungs-GTV. To mitigate covariate shifts, batch harmonization of the study arms was performed via ComBat. Combinations of radiomics, dosiomics, DVH, and clinical features were analyzed, with the minimum Redundancy - Maximum Relevance algorithm selecting top features across 1000 bootstrap iterations for each feature set. Classification employed Random Forest and XGBoost models, using three iterations of five-fold nested cross-validation with SMOTE-Tomek to address class imbalance. For each feature set, the optimal combination of classifier and feature number was selected and further tuned in a separate repeated five-fold cross-validation. Final models were validated on the external test set.
Results: The model based on radiomics achieved a test AUC of 0.62 (95%-CI: 0.48 – 0.74). The best-performing dosiomics model in cross-validation, based on EQD2 dosiomics, achieved a test AUC of 0.56 (95%-CI: 0.43 – 0.69). A subgroup analysis showed a higher test AUC for 3D-CRT compared to IMRT (best model: EQD2 dosiomics: 0.71 (95%-CI: 0.50 – 0.84) vs. radiomics + physical dosiomics: 0.63 (95%-CI: 0.50 – 0.75)). Models based on clinical data and DVH parameters achieved test AUCs of 0.47 (95%-CI: 0.31 – 0.62) and 0.49 (95%-CI: 0.32 – 0.65).
Conclusion: The results show that machine learning has the potential to predict symptomatic RP. The EQD2 dosiomics model outperformed the clinical and DVH models, but the overall prediction quality was low. The limited generalizability may be attributed to differences in patient characteristics, treatment regimens, and event rates between the training and test cohorts. Due to wide confidence intervals, no significant model superiority could be determined. The radiotherapy technique proved to be a decisive factor for the dosiomics analysis and should be considered in future studies.