Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancers, which remain the leading cause of cancer mortality worldwide. Even after complete surgical resection and standardized adjuvant therapy, 20-25% of patients develop recurrence or distant metastases, highlighting the inadequacy of current staging systems for guiding individualized treatment decisions.
Traditional prognostic tools such as TNM staging and histologic subtype fail to capture the full complexity of tumor biology, particularly the heterogeneity that exists within and between tumors. Patients with comparable clinical profiles can have dramatically different outcomes, suggesting that important biological differences are being missed.
Radiomics - the extraction of quantitative features from medical images using mathematical algorithms - has shown promise for capturing tumor heterogeneity non-invasively. However, most existing radiomic models have relied on statistical correlations without demonstrating what biological processes their features actually reflect, limiting confidence in their clinical use.
This study takes a different approach, integrating CT radiomic data with genomic, immune, and clinical information to build a comprehensive framework linking imaging features to tumor molecular characteristics, immune microenvironment composition, and clinical phenotypes.
The study enrolled 238 NSCLC patients who underwent radical surgical resection at two Chinese cancer centers between 2018 and 2020. Patients were split 70:30 into training and internal validation sets. An external validation cohort of 96 patients was drawn from the publicly available NSCLC Radiogenomics dataset in The Cancer Imaging Archive (TCIA).
All patients had preoperative contrast-enhanced chest CT scans acquired within one month of surgery, which served as the imaging source for radiomic feature extraction. Patients who received prior chemotherapy, radiotherapy, or other anti-tumor treatments before surgery were excluded to ensure clean imaging data reflecting untreated tumor biology.
The primary outcome was disease-free survival (DFS), defined as time from surgery to first recurrence, metastasis, or death. Overall survival (OS) was also analyzed. Follow-up was structured - every 3 months for the first 2 years, every 6 months from years 2 to 5, then annually - with a median follow-up of 50 months across the full cohort.
The external TCIA cohort was valuable because it included RNA sequencing data alongside CT images, enabling the genomic analyses (gene enrichment and immune cell profiling) that form the biological validation core of this study.
A total of 851 radiomic features were extracted from CT images for each patient using PyRadiomics, including shape, first-order statistics, texture matrices (GLCM, GLRLM, GLSZM, NGTDM, GLDM), and 744 wavelet-transform features. Tumor boundaries were manually delineated by an experienced radiation oncologist, with blood vessels and bronchi excluded.
Feature reproducibility was rigorously tested through both intra-observer (same observer, 1-week interval) and inter-observer (two observers) reliability assessments in 50 randomly selected patients. Only features with intraclass correlation coefficient above 0.90 were retained, reducing 851 to 452 stable features.
Three-step feature selection was applied: first, maximum relevance minimum redundancy (MRMR) and random survival forest algorithms identified the top 20 most informative features; then LASSO-Cox regression with 10-fold cross-validation further reduced this to 4 key radiomic features that composed the final Rad-score.
The resulting Rad-score was applied with fixed coefficients and thresholds to internal and external validation cohorts, ensuring that no information from validation patients influenced model development - a critical requirement for unbiased performance assessment.
The Rad-score demonstrated significant prognostic stratification in both training and external validation cohorts. High-risk patients had substantially shorter DFS and OS compared to low-risk patients, as confirmed by Kaplan-Meier survival curves and log-rank testing. The Rad-score AUC for DFS prediction was 0.643 in training and was validated in the independent TCIA cohort.
In multivariate Cox analysis, the Rad-score emerged as an independent predictor of both DFS (HR approximately 2.9, p = 0.004) and OS (HR approximately 3.1, p = 0.001) after adjusting for established clinical factors including T stage, N stage, tumor differentiation, and inflammatory-nutritional markers.
When combined with clinical parameters, the Rad-score significantly improved model performance. The combined model achieved C-index values of 0.704 (DFS) and 0.748 (OS) in training - exceeding the clinical-only model's values of 0.655 and 0.684. A similar improvement was observed in the internal validation cohort.
A prognostic nomogram was built incorporating the Rad-score alongside N stage, T stage, platelet count, and prognostic nutritional index. Calibration curves confirmed strong agreement between the nomogram's predicted probabilities and actual observed survival rates at 1, 3, and 5 years.
Gene set enrichment analysis (GSEA) was performed on 96 TCIA patients stratified by Rad-score and linked to RNA sequencing data. The high-risk group showed significant enrichment of pathways associated with aggressive tumor behavior, including epithelial-mesenchymal transition (EMT), hypoxia, TNFA-NF-kB signaling, inflammatory response, angiogenesis, and KRAS signaling.
Epithelial-mesenchymal transition enables cancer cells to detach from the primary tumor, migrate through tissue, and seed distant metastases - its enrichment in the high-risk group provides a molecular explanation for the worse outcomes observed in these patients.
Conversely, tumor-suppressive pathways were significantly downregulated in the high-risk group, including p53 signaling, interferon-alpha response, reactive oxygen species pathways, and oxidative phosphorylation. This combination of activated pro-invasive pathways and suppressed anti-tumor defenses creates a permissive environment for unchecked tumor growth.
Differential gene expression analysis identified 71 upregulated and 116 downregulated genes in the high-risk group, with pathway enrichment pointing to alterations in cell cycle regulation, DNA repair, apoptosis, and notably platinum-based drug resistance - a clinically significant finding for patients likely to receive platinum chemotherapy adjuvantly.
ESTIMATE algorithm analysis showed that while tumor purity was comparable between risk groups, the immune score and stromal score were significantly lower in the high-risk group. Reduced immune infiltration suggests a more immunologically cold tumor environment with impaired anti-tumor immune activity.
Immune phenotype scoring revealed a significantly lower MHC (major histocompatibility complex) score in high-risk patients, reflecting reduced antigen presentation capacity. Without adequate antigen presentation, the immune system cannot efficiently recognize and target tumor cells, facilitating immune escape.
Detailed immune cell composition analysis using CIBERSORTx identified a significantly lower proportion of naive B cells (p = 0.043) and a significantly higher proportion of activated mast cells (p = 0.02) in the high-risk group. B cells support anti-tumor immunity by producing antibodies and coordinating immune responses, while elevated mast cells may promote inflammatory and pro-tumorigenic signaling.
Together, these immune findings characterize an immunosuppressive tumor microenvironment in high Rad-score tumors: reduced immune surveillance, impaired antigen presentation, fewer B cells, and a pro-inflammatory mast cell-rich infiltrate - all features that could explain resistance to immune checkpoint therapies and conventional treatments.
At the clinical level, the high-risk group showed significantly elevated inflammatory markers including white blood cell count, neutrophils, monocytes, neutrophil-to-lymphocyte ratio (NLR), and systemic immune-inflammation index (SII). These blood-based inflammatory indices reflect systemic immune dysregulation that mirrors the local tumor microenvironment findings.
Nutritional status was also compromised in the high-risk group, with significantly lower albumin levels. Malnutrition impairs immune cell function and reduces treatment tolerance - creating a cycle where poor nutrition, enhanced inflammation, and immune suppression synergistically promote tumor progression.
Pathologically, high Rad-score patients had more advanced T and N stages, larger tumor diameters, more positive lymph nodes, poorer tumor differentiation, and higher Ki67 proliferation index - all established markers of more aggressive disease and worse prognosis.
This convergence of findings across molecular (gene pathways), immunological (immune cell composition), and clinical (inflammatory markers, staging) dimensions provides a coherent, multi-scale explanation for why the Rad-score stratifies prognosis - the imaging features appear to genuinely reflect underlying tumor biology rather than being arbitrary statistical correlations.
The study's key contribution is providing a biological rationale for CT-based radiomic prognostication in NSCLC. Rather than treating radiomic features as black-box predictors, this work demonstrates that they reflect real molecular and immune characteristics of tumors - making them more credible candidates for clinical adoption.
The Rad-score identifies high-risk patients who may be missed by conventional TNM staging, potentially enabling earlier and more aggressive adjuvant therapy decisions. The association with platinum drug resistance pathways in particular suggests potential utility in chemotherapy selection decisions.
Important limitations include the relatively small sample size (238 primary patients and 96 external validation patients), which limits statistical power and generalizability. The external TCIA cohort also has different CT scanner equipment and acquisition parameters, which may affect feature stability and introduce variability.
The biological associations are observational, not causal. Whether the gene pathway differences and immune cell changes directly drive the imaging phenotypes seen in the Rad-score features requires experimental validation. Larger, prospective multicenter studies with standardized imaging protocols and matched genomic data will be needed before clinical implementation.