BLUE RIBBON POSTER MULTI-DISCIPLINARY: Evaluation of a Radiomics-Based Predictive Model for Radiation Pneumonitis Using Lung Ventilation Imaging
Abstract
Purpose
Grade ≥2 radiation pneumonitis (RP) occurs in approximately 30% of patients undergoing radiotherapy for lung cancer. Therefore, a predictive model to estimate RP risk before radiotherapy is required. Radiomics and dosiomics have shown potential for RP prediction. Recent studies suggest that radiomics features derived from lung ventilation images may enable more accurate prediction of RP. This study aimed to develop a Grade ≥2 RP prediction model incorporating radiomics features from four-dimensional computed tomography (4DCT) - based ventilation imaging.
Methods
77 Patients with locally advanced non-small cell lung cancer, included 24 patients with Grade ≥2 RP, were included. Radiomics and dosiomics features were extracted from the whole lung using CT, ventilation imaging, and dose distributions. Feature selection was using leave-one-out cross-validation, LASSO regression, and pearson correlation analysis. Clinical, radiomics, and dosiomics models were developed using the random forest method. Six prediction models (clinical, CT, ventilation image, dose distribution, CT/ventilation image/dose distribution, and clinical/ CT/ventilation image/dose distribution) were created to predict RP occurrence, and their performance was assessed using the area under the curve (AUC) of the receiver operating characteristic (ROC). The ROC curves were compared using DeLong’s method (95% confidence intervals).
Results
For the optimal classification model, 8 features were selected for the clinical/ CT/ventilation image/dose distribution model, including two clinical, one CT, two ventilation image, and three dose distribution. The AUCs (95% confidence intervals) for the clinical, CT, ventilation image, dose distribution, CT/ventilation image/dose distribution, and clinical/ CT/ventilation image/dose distribution was 0.71 (0.58–0.83), 0.69 (0.56–0.82), 0.73 (0.60–0.85), 0.67 (0.53–0.80), 0.79 (0.67–0.91), and 0.84 (0.74–0.93). The best model was significantly more accurate than the clinical model (p = 0.02).
Conclusion
Incorporating features derived from ventilation images, in addition to conventional metrics, was suggested to improve predictive performance.