DEEP RADIOGENOMICS FOR PREDICTING IMMUNOTHERAPY RESPONSE IN NON-SMALL CELL LUNG CANCER: A CROSS-DOMAIN ANALYSIS OF IMAGING, TUMOR MICROENVIRONMENT, AND GENETIC BIOMARKERS

Abstract

Non-small cell lung cancer (NSCLC) presents a significant clinical challenge in the era of immunotherapy due to its heterogeneous response patterns and the limitations of current predictive biomarkers. This study proposes a comprehensive radiogenomics-based predictive framework integrating 18F-FDG PET/CT imaging, radiomic feature extraction, delta-radiomics, and machine learning to non-invasively predict immunotherapy response. Quantitative radiomic features including SUVmean, SUVmax, metabolic tumor volume (MTV), and total lesion glycolysis (TLG) were extracted from baseline and follow-up PET/CT scans across a cohort of NSCLC patients undergoing immune checkpoint inhibitor therapy. Delta-radiomics captured temporal changes in tumor characteristics, while machine learning models—particularly ensemble classifiers—were trained to classify responders versus non-responders with high accuracy. SHAP-based explainable AI was incorporated to identify key imaging biomarkers influencing model predictions. Results showed that delta-radiomic features provided superior predictive power compared to static imaging biomarkers, with models achieving high F1-scores and ROC-AUC values. Volumetric PET parameters also demonstrated significant correlations with PD-L1 expression and progression-free survival, underscoring their clinical relevance. The proposed non-invasive, multimodal approach represents a scalable and interpretable solution for enhancing patient stratification and optimizing immunotherapy outcomes in NSCLC. These findings support the clinical translation of AI-driven radiogenomics tools in personalized cancer care.