Author: David J. Carlson, Ming Chao, Tian Liu, Yong Hum Na, Kenneth E Rosenzweig, Robert Samstein, Lewis Tomalin 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, Yale University School of Medicine, Department of Therapeutic Radiology, Yale University School of Medicine 🌍
Purpose: To investigate the potential of regional radiomic features extracted from five lung sub-lobes on pre-treatment CT as biomarkers for predicting radiation pneumonitis (RP) using machine learning.
Methods: Bilateral lungs from planning CT images from 485 lung cancer patients enrolled in the NRG Oncology RTOG 0617 trial were segmented into five sub-lobes, and 107 radiomic features were calculated for each sub-lobe using PyRadiomics. Patients were categorized into Group1 (n=419, grades 0-1 pneumonitis) and Group2 (n=66, grade≥2 pneumonitis). The Synthetic Minority Oversampling Technique (SMOTE) and Class Weighting Technique (CWT) were applied independently to create balanced datasets and minimize biases from imbalanced data. A variance threshold technique was applied to reduce collinearity in the radiomic features. The random forest (RF) classifier was employed for RP prediction, and model performance was assessed using the area under the receiver operating characteristic curve (AUROC). Bootstrapping with 1,000 iterations was performed to assess the uncertainties in AUROC and accuracy. Highly important features were analyzed with the RF classifier and SHapley Additive exPlanations (SHAP) independently to explore image-based biomarkers for RP prediction.
Results: Regional features from the right-middle lobe (RML) showed a correlation with RP, with AUROCs of 0.635±0.080(SMOTE) and 0.636±0.080(CWT). The left-lower lobe (LLL) also showed some correlation with RP, but its corresponding AUROCs were lower. Prominent features from RML emphasize structural and volumetric changes, capturing the global extent of RP in this lobe, while important features from LLL focus on texture and intensity-based markers, reflecting subtle heterogeneities caused by RP in a perfusion-dominant lobe. No significant predictive biomarkers were identified from other sub-lobes.
Conclusion: This study indicates that pre-treatment CT contains features that can serve as biomarkers for predicting radiation pneumonitis, which could help guide RT planning for lung cancer to reduce toxicity. Further investigations with larger patient datasets are needed to confirm these findings.