Multi-Scale Dose-Defined Lung Subregional Dosiomics for Radiation Pneumonitis Prediction In VMAT with Conventional Fractionation
Abstract
Purpose
This study aimed to develop and evaluate a multi-scale dose-defined lung subregional dosiomics approach for predicting radiation pneumonitis (RP) in lung cancer patients treated with conventionally fractionated VMAT.
Methods
This retrospective study included 118 patients treated with conventionally fractionated VMAT, split into a training cohort (n=99) and a held-out test cohort (n=19). RP grade ≥2 occurred in 40/118 (33.9%) patients (34/99 in training, 6/19 in test). Dose-defined lung subregions were generated using (1) cumulative dose thresholds (≥0, 5, ..., 40 Gy) and (2) interval dose bands (0–5, 5–10, ..., 40–max Gy), along with the whole-lung (excluding the internal target volume(ITV)) ROI. From each ROI, 3D dosiomics features were extracted (1409 features/ROI) to form cumulative, interval, and hybrid feature sets. A feature selection pipeline consisting of z-score normalization, t-test screening (p 0.9, and ℓ2,1-norm minimization with bootstrap-repeated stratified cross validation was used to select stable top-k features and train a logistic regression model. Model performance was evaluated using AUC and PR-AUC.
Results
In internal validation (10-fold cross-validation), the hybrid feature set achieved the highest performance (AUC 0.870; 95% CI: 0.807–0.937), outperforming whole-lung dosiomics (AUC 0.683; 0.629–0.729) and standard DVH metrics (V5, V10, V20, MLD; AUC 0.612; 0.573–0.658). The Hybrid model also demonstrated strong precision–recall performance (PR-AUC 0.823; 0.751-0.889), indicating improved discrimination under class imbalance. In the test cohort, the Hybrid model remained superior (AUC 0.846; PR-AUC 0.741) compared with whole-lung dosiomics (AUC 0.628; PR-AUC 0.568) and DVH metrics (AUC 0.593; PR-AUC 0.538).
Conclusion
The multi-scale hybrid dosiomics approach demonstrated superior predictive performance compared to single-scale and conventional DVH approaches. By capturing spatial dose heterogeneity that global metrics miss, this approach facilitates more accurate RP risk prediction for patients treated with VMAT.