Author: Katelyn M. Atkins, Indrin J. Chetty, Elizabeth M. McKenzie, Taman Upadhaya, Samuel C. Zhang 👨🔬
Affiliation: Department of Radiation Oncology,Cedars-Sinai Medical Center, Cedars-Sinai Medical Center 🌍
Purpose:
We explored a multi-regional and multi-omics approach to extract CT-based radiomics and 3D dosiomics features to predict radiation pneumonitis (RP) in patients with locally advanced Non-Small Cell Lung Cancer (LA-NSCLC).
Methods:
The NRG Oncology/Radiation Therapy Oncology Group (RTOG) 0617 data set was utilized for this study. A radiomics analysis pipeline was implemented to extract features from CT and 3D dose maps using multi-regional lung contours (ipsilateral, contralateral and both lungs-CTV) to predict for RP³grade 2. Receiver operating characteristic (ROC) and correlation analyses were applied to identify significant predictors of toxicity. Predictive models, LASSO, Random Forest (RF), XGBoost and Support Vector Machine (SVM (linear and RBF)) were trained on 315 patients using 10-trials, 10-fold nested cross-validation (CV). Independent predictors were ranked on importance score and stepwise-forward feature selection was used to select the subset of features minimizing the validation error. The final model (based on highest AUC) was applied to 136 independent, “unseen” test patients.
Results:
RP>=grade 2 occurred in ~15% (n=67). Models based solely on clinical features were not predictive. The highest performing model (LASSO) used a combination of radiomics, dosiomics, and filter features extracted from both lungs (contra-and ipsilateral-CTV); AUC was 0.82 (validation dataset) and 0.67 (test dataset). The second-highest AUCs were 0.83 (validation), and 0.66 (test), achieved (using the LASSO algorithm) by combining radiomics, dosiomics, and filter features from the contralateral lung.
Conclusion:
Results are suggestive that radiomic, dosiomic and filter features on planning CT’s and 3D dose distributions from both lungs could provide complimentary information to enhance model performance implying that such underlying features may serve as a signature for prediction of post-treatment RP. Although statistical methods were employed to minimize bias, the limited sample size prevents definitive conclusions, underscoring the need for further research in this space using larger sample sizes.