Room: Track 2
Purpose: The aim of this study was to develop a predictive model for radiation pneumonitis (RP) after lung cancer stereotactic body radiation therapy (SBRT) using radiomic features within lung regions of interest (ROIs) segmented by dosimetric information on pretreatment planning computed tomography (CT) images.
Methods: A total of 135 non-small cell lung cancer patients, who received SBRT, were selected for modeling the RP prediction. Twenty-two patients had grade 2 or higher RP. A total of 486 radiomic features including wavelet-based radiomic features were calculated for two ROIs of whole lung volumes (WLV) and LV20 that was defined as normal lung volumes irradiated more than 20 Gy. The top four significant features with the highest frequency was selected as a signature using a least absolute shrinkage and selection operator (LASSO) logistic regression for each ROI. Predictive models with the signature were built via a linear support vector machine (SVM) for the prediction of RP grade =2. To address the issue of imbalanced data with RP-positive and RP-negative cases on SVM models, five balanced subsets were prepared by composing them of the same 22 RP-positive and randomly selected 22 or 23 RP-negative cases. The models were evaluated based on a leave-one-out cross validation by using the area under receiver operating characteristic curves (AUCs), sensitivity, and specificity for each subset.
Results: The means of AUC, sensitivity, and specificity were 0.744 (95% CI: 0.697-0.791), 0.645 (0.555-0.736), and 0.689 (0.625-0.753) for WLV, and 0.818 (0.765-0.871), 0.718 (0.691-0.745), and 0.773 (0.700-0.845) for LV20, respectively.
Conclusion: The radiomic signature calculated from pretreatment planning CT images can be utilized as a predictive imaging biomarker for RP in SBRT treatment planning for lung cancer.
Lung, Radiation Risk, Texture Analysis