Improving PD-L1 Expression Prediction in non-Small Cell Lung Cancer Using Radiomic Analysis and Ensemble Machine Learning Models on Whole-Body vs 18F-FDG Lung PET/CT Data

Lung cancer remains the leading cause of cancer-related mortality globally. Non-small cell lung cancer (NSCLC) account for 85% of lung cancer cases. Programmed Death-Ligand 1 (PD-L1) is a critical biomarker in NSCLC for guiding immunotherapy, yet conventional detection methods, such as immunohistochemistry, are invasive and insufficiently capture tumour heterogeneity. This study explores the potential of radiomics-based machine learning (ML) models to predict PD-L1 expression using whole-body and lung [18F]-FDG PET/CT imaging data. Radiomic features were extracted from 134,590 whole-body and 43,142 lung slices obtained from 308 NSCLC patients. The study evaluated the performance of ensemble models including random forest, and XGBoost at feature- and patient-level analyses. Models trained on whole-body data consistently outperformed those using lung data, highlighting the importance of systemic metabolic information. At the feature level, the random forest model achieved 99.1% accuracy and an ROC-AUC of 0.998. Similarly, at the patient-level, the random forest model also achieved 98% accuracy and an ROC-AUC of 0.995. The findings demonstrate the superiority of whole-body radiomic analysis over localised imaging for predicting PD-L1 expression. This approach underscores its potential as a non-invasive diagnostic tool, providing an advanced framework for precision oncology in NSCLC management.