Predictive Modelling for Radiation Pneumonitis Using Mid-Treatment Functional Information
Abstract
Purpose
Current clinical toxicity models for predicting radiation-induced pneumonitis are limited by lack of functional information; furthermore, current practice assumes a relative biological effectiveness (RBE) of 1.1 for proton therapy patients. We hypothesize that predictive models for RP will demonstrate improved accuracy when variable RBE and lung functional information are incorporated.
Methods
For 72 photon (n=12 ≥grade 2 pneumonitis) and 38 proton (n=8 ≥grade 2 pneumonitis) lung cancer patients, 4DCT-derived ventilation maps were generated using stress-based methods at planning and mid-treatment. Voxel-wise change in ventilation from planning to mid-treatment was calculated for each patient. Proton dose was recalculated with variable RBE using an in vivo, imaging-based endpoint. Three models were trained based on: i) planning CT and original dose map (CT+Dose), ii) planning CT and variable RBE-weighted dose map (CT+RBE), and iii) planning CT, variable RBE-weighted dose map, and ventilation change map (CT+RBE+VC). Radiomics features from the planning CT, dose maps, patient demographics, and ventilation change maps were evaluated for stability across multiple selection methods and reduced to 6 features. Each model was selected from the highest performance of 6 machine learning models. Models were assessed on 7/6-fold nested cross-validation and area under the curve (AUC) was calculated.
Results
Features demonstrating highest stability were and V70 (lung volumes receiving ≥40 and ≥70 Gy respectively), mean HU within lung, first-order entropy within lung CT, first-order mean absolute deviation within lung CT, and first-order entropy of ventilation change at mid-treatment (excluded from CT+Dose and CT+RBE models). All models demonstrated highest performance when trained using Random Forest. The AUC for CT+Dose, CT+RBE, and CT+RBE+VC models was 0.75±0.08, 0.76±0.08, and 0.78±0.10 respectively.
Conclusion
This study demonstrated the potential of functional information and variable RBE to improve prediction modeling for pneumonitis. Future work will incorporate lobe-wise ventilation information and tumor response into predictive models.