Impact of Dose- and Function-Based Lung Regions on Radiomics-Based Prediction of Radiation Pneumonitis
Abstract
Purpose
Radiation pneumonitis (RP) remains a major toxicity in thoracic radiotherapy. While lung radiomics has shown promise for RP prediction, the optimal region for feature extraction remains unclear. This study investigated whether restricting the extraction region based on dose and lung function improves RP prediction.
Methods
Seventy-two patients with lung cancer treated with radiotherapy were retrospectively analyzed, including 22 patients who developed grade ≥2 RP. Radiomics features were extracted from CT images, dose distributions, and CT-based ventilation images using the following four lung regions: (1) whole-lung, (2) lung receiving ≥20 Gy, (3) the top 85% ventilated lung, and (4) lung receiving ≥20 Gy and the top 85% ventilated lung. Feature selection was performed using LASSO regression and Pearson correlation analysis. Prediction models were created using random forest to predict grade ≥2 RP, including radiomics-based models and a clinical-feature-only model incorporating dose-volume metrics. Model performance was evaluated using the area under the curve (AUC) obtained from leave-one-out cross-validation, and AUCs were compared using DeLong’s method with 95% confidence intervals (95%CIs).
Results
The highest predictive performance was achieved when radiomics features were extracted from the top 85% ventilated lung (AUC: 0.84, 95%CI: 0.72–0.95). The extraction region to lung receiving ≥20 Gy yielded an AUC of 0.80 (95%CI: 0.69–0.91). Whole-lung radiomics achieved an AUC of 0.74 (95%CI: 0.60–0.87), while the clinical-feature-only model yielded an AUC of 0.70 (95%CI: 0.56–0.84). The combined dose- and function-based region showed an AUC of 0.77 (95%CI: 0.64–0.89). Pairwise comparisons of AUCs did not reach statistical significance by DeLong’s method.
Conclusion
Restricting radiomics feature extraction to functionally defined lung regions was associated with improved RP prediction compared with whole-lung or clinical feature only approaches. These findings suggest that incorporating CT-based lung function into region selection may provide additional value for radiomics-based RP risk assessment in lung cancer radiotherapy.