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

Lung cancer remains the leading cause of cancer-related deaths world-wide. Non-small cell lung cancer (NSCLC) account for 85% of lung cancer cases. Accurate assessment of Programmed Death-Ligand 1 (PD-L1) expression is essential for guiding immunotherapy decisions; however, conventional detection methods such as immunohistochemistry (IHC) are invasive and may not fully account for tumour heterogeneity. This study investigates the potential of deep learning-based models to predict PD-L1 expression using radiomic features extracted from full-body and lung-segmented 18F-FDG PET/CT scans. Radiomic features were extracted from 134,590 full-body and 43,142 lung slices obtained from 308 NSCLC patients. Fully Connected Neural Networks (FCNNs) were trained and optimized with different network depths, activation functions, dropout rates, and optimizers. Models trained on whole-body data consistently outperformed those using lung data, emphasizing the importance of systemic metabolic information. At the feature level, the best FCNN model achieved 99% accuracy and an ROC-AUC of 0.997, while at the patient level, it achieved 98% accuracy and an ROC-AUC of 0.995. The findings highlight the superiority of deep learning-based full-body radiomic analysis over localized imaging in predicting PD-L1 expression. This approach demonstrates the potential of FCNNs as a non-invasive diagnostic tool, providing an advanced framework for precision oncology and NSCLC immunotherapy decision-making.