A radiomic approach to non-invasively decode tumor immune microenvironmental features in surgically resected non-small cell lung cancer

Introduction: This study aimed to develop radiomics-based multivariate classifiers able to non-invasively decipher the tumor immune microenvironment (TIME) in resected Non-small cell lung cancer (NSCLC) patients. Methods: Tumor-infiltrating lymphocytes (TILs) subpopulations and PD-L1 TPS were immunohistochemically assessed. Radiomic features (RFs) were extracted from preoperative CT scans using the PyRadiomics platform. Six machine learning-based models were built to predict distinct TIME-related endpoints: PD-L1 (M1), CD8/CD3 ratio (M2), PD1/CD8 ratio (M3), CD3+ TILs (M4), CD4+ TILs (M5), CD8+ TILs (M6). Except for PD-L1 (classified as positive vs negative), patients were dichotomized into high versus low groups based on median values. Model development relied on Random Forest classifiers with sequential feature selection and internally validated through Leave-One-Out Cross-Validation splits Results: A total of 98 patients with stage I–IIIA resected NSCLC were included. From 1702 initially extracted RFs per scan, 231 non-redundant features (116 tumor and 115 peritumoral) were retained after correlation filtering. Final models incorporated between 3 and 7 RFs derived from both tumor and peritumoral regions. Model performance showed accuracy, precision, sensitivity, receiver operating curve (ROC) Area under the Curve (AUC) and precision-recall curve (PRC) AUC values ranging from 70% to 80%, with slightly lower performance for M2. Specificity ranged from 55% to 77%, with the highest value observed in the CD8+ TILs model (M6). Feature selection varied across models, although one peritumoral feature (HLL-wavelet GLCM Correlation) was consistently selected in four models (M1, M2, M4, M5). Conclusions: Tumoral and peritumoral CT-based radiomics might effectively intercept TIME features, thus suggesting a potential non-invasive approach in resected NSCLC patients.