NSCLC and the CCRT Challenge. Non-small cell lung cancer (NSCLC) accounts for the majority of lung malignancies, with approximately 60% of cases identified at locally advanced or metastatic stages. For those ineligible for surgery, concurrent chemoradiotherapy (CCRT) is the primary treatment, yet the 5-year survival rate remains only 15%, partly due to locoregional failure in roughly 73% of patients.
Intra-Tumoral Heterogeneity as a Barrier. Intra-tumoral heterogeneity (ITH) refers to the coexistence of distinct cellular subpopulations with different genetic, phenotypic, and functional characteristics within a single tumor. This spatial diversity leads to variable and often unpredictable responses to CCRT, as different subregions may respond differently to radiation and chemotherapy.
Limitations of Current Biomarker Approaches. While molecular biomarkers identified from tumor biopsies help guide treatment, single biopsy specimens sample only a limited fraction of the tumor, failing to capture the full spectrum of spatial heterogeneity. This sampling bias means biopsy-based biomarkers may not reflect the true biological diversity driving treatment resistance.
Radiomics as a Solution. CT-based radiomics offers a noninvasive alternative by extracting quantitative texture and shape features from the entire tumor volume on pretreatment scans. This approach overcomes sampling bias and has shown promising results for predicting treatment response in lung cancer. However, most prior radiomics studies analyze whole-tumor features and overlook the spatial heterogeneity within different tumor subregions.
Patient Population. This retrospective multicenter study enrolled 164 NSCLC patients who underwent CCRT at six medical centers in Guizhou, China between January 2019 and December 2023. The cohort was split into a training set (132 patients) and a validation set (32 patients). Treatment response was assessed at 3 months post-CCRT using RECIST 1.1 criteria, with complete or partial remission classified as objective response.
CCRT Regimen. All patients received standard conventional-dose radiotherapy at 60 to 66 Gy in 30 to 33 fractions. Concurrent chemotherapy followed histology-specific regimens: paclitaxel plus carboplatin weekly for squamous cell carcinoma, and pemetrexed plus cisplatin every 3 weeks for non-squamous subtypes. Only patients without prior specialized cancer treatments were eligible.
Tumor Subregion Clustering. To capture intra-tumoral heterogeneity, the Gaussian Mixture Model (GMM) was applied to the tumor region of interest with the Bayesian Information Criterion determining the optimal number of clusters. Three tumor subregions were identified, each with distinct imaging profiles representing different biological populations within the tumor.
Feature Extraction Volume. From each of the three subregions, 1,835 radiomic features were extracted using PyRadiomics, yielding 5,505 intra-tumoral ecological diversity features per patient. An additional 1,835 conventional whole-tumor radiomic features were extracted separately for comparison. Features covered seven categories including first-order statistics, shape, and multiple texture matrices.
Rigorous Feature Reduction. Features with high inter-feature correlation (Pearson index greater than 0.9) were removed to eliminate redundancy. For the ecological diversity features, a Student's t-test first filtered features with P less than 0.05, followed by LASSO regression with five-fold cross-validation to reduce dimensionality further. All feature selection steps were performed exclusively on training data.
Final Feature Sets. After selection, 13 intra-tumoral ecological diversity features and 7 conventional radiomics features (both with P less than 0.05) were retained for model development. Two clinical variables were identified through univariate and multivariate logistic regression as independent predictors: the number of chemotherapy cycles during CCRT and the neuron-specific enolase (NSE) level.
Seven Models Compared. Using the SVM algorithm, seven predictive models were built covering all combinations of clinical, conventional radiomics, and ITH features: (1) clinical only, (2) C-radiomics only, (3) ITH only, (4) clinical plus radiomics, (5) clinical plus ITH, (6) radiomics plus ITH, and (7) the combined model using all three feature sets.
Algorithm Selection Process. Before finalizing SVM, 11 machine learning algorithms were systematically compared including logistic regression, Naive Bayes, K-nearest neighbors, Random Forest, ExtraTrees, XGBoost, LightGBM, Gradient Boosting, AdaBoost, and multilayer perceptron. SVM demonstrated the highest and most consistent AUC on both training and validation sets, avoiding the severe overfitting seen with Random Forest and XGBoost.
ITH Model Outperforms Conventional Radiomics. The ITH model using only intra-tumoral ecological diversity features achieved AUCs of 0.78 (95% CI: 0.70 to 0.86) in training and 0.77 (95% CI: 0.59 to 0.94) in validation. This substantially exceeded the conventional whole-tumor radiomics model, which achieved AUCs of only 0.64 and 0.69, and the clinical-only model at 0.62 and 0.64.
Combined Model Best Performance. The combined model integrating all three feature types (clinical, conventional radiomics, and ITH features) achieved the highest performance: AUC of 0.92 (95% CI: 0.86 to 0.98) in training and 0.87 (95% CI: 0.72 to 1.00) in validation. This represented a substantial improvement over any individual component model.
Clinical Predictors Identified. Multivariate analysis confirmed three independent predictors of CCRT response: receiving fewer than 3 chemotherapy cycles during CCRT (OR 0.32, P=0.006), elevated NSE level (OR 0.41, P=0.03), and high binary ITH index (OR 2.42, P=0.02). High ITH was independently associated with more than double the odds of a favorable response.
Decision Curve Analysis. Decision curve analysis confirmed that the combined model provided the highest net clinical benefit across most threshold ranges, outperforming all simpler models including the clinical-only and radiomics-only approaches. This demonstrates practical superiority of the combined approach for real-world clinical decision support.
High vs. Low Risk by ITH Score. Using the optimal Youden index cutoff of 0.65 derived from the ITH model, patients were dichotomized into high-risk and low-risk groups. Kaplan-Meier analysis revealed significantly worse progression-free survival (PFS) and overall survival (OS) in the high-risk group in both training (P less than 0.001 for PFS, P=0.047 for OS) and validation cohorts (P=0.04 for PFS, P=0.007 for OS).
Three-Tier Stratification with the Combined Model. The combined model established two cutoff thresholds (0.53 and 0.86), classifying patients into low-risk, intermediate-risk, and high-risk groups. Kaplan-Meier analysis showed significant differences in both PFS and OS across all three strata in training (P less than 0.001 for both) and validation (P less than 0.001 for PFS, P=0.002 for OS) cohorts.
Worst Outcomes in High-Risk Group. Patients classified as high-risk by the combined model experienced significantly shorter PFS and OS compared to intermediate and low-risk groups. This three-tier stratification provides finer granularity for clinical decision-making than binary classification alone, potentially allowing intermediate-risk patients to be monitored more closely.
Prognostic Implications. The ability of pretreatment CT-derived features to predict not only immediate treatment response but also long-term survival outcomes positions the ITH model as a genuinely prognostic tool. Identifying high-risk patients before CCRT begins could enable clinicians to explore alternative treatment intensification strategies or enroll these patients in clinical trials of novel agents.
Why Subregional Features Perform Better. Whole-tumor radiomics treats a tumor as a single homogeneous entity, averaging out the spatial variation between biologically distinct subpopulations. The subregional approach used here captures phenotypic heterogeneity by separately characterizing each subregion, providing richer information about the range of tumor cell behaviors that drive differential responses to treatment.
Ecological Diversity as a Conceptual Framework. The use of ecological diversity metrics to quantify tumor heterogeneity is a novel conceptual borrowing from ecology. Just as species diversity indices describe the complexity of an ecosystem, these metrics describe the diversity of imaging phenotypes within a tumor, offering a principled framework for heterogeneity quantification that goes beyond simple texture analysis.
Comparison with Prior Studies. Previous subregional radiomics work in other tumor types, including oropharyngeal cancers and breast cancer, achieved combined AUCs of 0.83 to 0.90 in multi-cohort validation. This study's combined model AUC of 0.87 is consistent with these benchmarks, supporting the broader applicability of subregional radiomics across cancer types and treatment contexts.
Clinical Value for Personalized CCRT. For thoracic oncologists, the combined ITH model could help identify patients unlikely to benefit from standard CCRT before treatment begins, enabling timely exploration of dose escalation, immunotherapy addition, or alternative regimens. This aligns with precision oncology goals of matching treatment intensity to predicted response at the individual patient level.
Key Findings. The CT-based ITH model, derived from ecological diversity features across three Gaussian-mixture-defined tumor subregions, predicted CCRT response in NSCLC with AUC of 0.77 in external validation. The combined model integrating ITH with clinical and conventional radiomics features further improved prediction to AUC 0.87 and successfully stratified patients by survival risk.
Study Limitations. The study is retrospective with a relatively small validation cohort of only 32 patients, limiting statistical power and raising overfitting risk. The cohort is predominantly advanced-stage disease from Chinese centers, potentially limiting generalizability to Western populations or earlier-stage patients. Manual tumor delineation introduces inter-reader variability that automated segmentation could mitigate.
Need for Prospective Validation. Before clinical deployment, external prospective validation in larger, more diverse cohorts is essential. The authors also identify automated tumor segmentation as a priority to improve reproducibility of radiomic features across institutions and reduce dependence on expert radiologist delineation.
Future Directions. Expanding the model to incorporate additional imaging modalities such as PET, including more clinical biomarkers, and developing automated segmentation pipelines are identified as key next steps. Ultimately, the goal is to integrate this pretreatment prediction tool into routine oncology workflows to support personalized treatment planning for NSCLC patients undergoing CCRT.