An exploratory study on whole-lung radiomics features from computed tomography for prognostic prediction in non-small cell lung cancer with concurrent chemoradiotherapy

Introduction

Background

Lung cancer is the second most commonly diagnosed malignancy and the leading cause of cancer-related mortality, accounting for nearly 156,000 deaths annually. Among these newly diagnosed cases, approximately 85% are classified as non-small cell lung cancer (NSCLC), with over 50% of these patients presenting at an advanced stage upon initial diagnosis (1,2). Despite concurrent chemoradiotherapy (CCRT) prolonging the survival of patients with advanced lung cancer (3), the 5-year overall survival (OS) rate of patients with NSCLC undergoing CCRT remains approximately 15% (4). This poor prognosis primarily stems from the aggressive biological behavior of the tumor and complex microenvironmental regulation. Also, intra-tumoral heterogeneity (ITH) and the tumor microenvironment (TME) are important causes of disease progression and relapse/treated failure (5-7).

Rationale and knowledge gap

Radiomics is a technique that extracts quantitative features from medical imaging data to determine the biological characteristics of the tumor (8-10), which can provide clinicians with valuable information for predicting patient prognosis. This approach identifies high-risk patients and guides personalized treatment strategies, ultimately improving therapeutic outcomes and quality of life (11-20). However, the accurate prognostic prediction of patients with advanced NSCLC undergoing CCRT remains challenging. Thus, exploring more precise prognostic tools is essential for optimizing treatment regimens, enhancing patient survival, and improving the quality of life. Moreover, several studies have suggested that the heterogeneity of lung tissue is associated with the prognosis of NSCLC (21-25); however, the prognostic value of ipsilateral lung tissue radiomics on predicting outcomes for cell lung cancer patients undergoing CCRT has not been fully elucidated.

Objective

This study aimed to enhance the prognostic accuracy for patients with advanced NSCLC undergoing CCRT by investigating the impact of radiomics features extracted from the entire ipsilateral lung. Specifically, we aim to compare these features using conventional tumor-centric two-dimensional (2D) and three-dimensional (3D) volumetric radiomics models, develop integrated models combining both approaches, and validate the predictive performance of the integrated models across multiple machine learning algorithms to identify the optimal prognostic model. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2586/rc).

Methods

Patients

We retrospectively obtained the pre-treatment computed tomography (CT) examinations of patients with NSCLC receiving CCRT at different centers from January 1 to December 31, 2019–2023. This paper’s dataset involves six centers performing CCRT on 164 patients: Guiyang Pulmonary Hospital, Qiandongnan People’s Hospital; Qiannan Prefecture TCM Hospital; The Second Affiliated Hospital, Guizhou Medical University, Guiyang First People’s Hospital; Qiannan People’s Hospital. A total of 131 cases were included in the training set, and 33 cases were included in the validation set. Selection criteria: (I) histologically confirmed NSCLC; (II) with CT-enhanced images of the chest during the month before starting therapy; (III) with routine irradiation of conventional split dose radiotherapy and CCRT ( no previous special cancer treatment before this study); (IV) Karnofsky Performance Status (KPS) ≥70. Exclusion criteria: (I) other malignant tumors; (II) radical operation or aim irradiation, chemotherapy before chemoradiation treatment; (III) insufficient image quality; and (IV) missing data after follow-up visit (Figure 1). The total CCRT was 60–66 Gy/30–33 fractions to gross tumor volume (GTV). The chemotherapy regimens of the study were squamous cancer chemotherapy regimens and non-squamous cancer chemotherapy regimens. Squamous cancer: paclitaxel 50 mg/m², d1 weekly, then carboplatin AUC2, d1 weekly, non-squamous cancer: pemetrexed 500 mg/m², d1, 3 weeks + cisplatin 75 mg/m², d1, 3 weeks. In this trial, we defined CCRT as 2–4 cycles of chemotherapy delivered concurrently with thoracic radiation therapy. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by The Second Afffliated Hospital, Guizhou Medical University Ethics Committee (approval No. 2020-LS-03) and individual consent for this retrospective analysis was waived. The other institutions were also informed and approved the study (Figure 1).

Figure 1 Patient inclusion and exclusion criteria. CT, computed tomography; KPS, Karnofsky Performance Status; NSCLC, non-small cell lung cancer.

Statistical analysis

In this study, we conducted a series of statistical analyses using SPSS 25.0 software. The Kaplan-Meier method was applied to plot the survival curves. We used the Cox proportional hazards regression model for both univariate and multivariate analyses of clinical features to screen valuable clinical indicators for the construction of a clinical model. For continuous variables, the Mann-Whitney U test was employed, while for count data, the Chi-squared test was used to compare the differences between different groups. The DeLong test was utilized to conduct a significance test on the diagnostic capabilities of different models. A P value less than 0.05 was considered to indicate a statistically significant difference.

Feature segmentation

The CT images of eligible patients were exported from the hospital Picture Archiving and Communication System (PACS) and imported into the ITK-SNAP (v3.8.0) and 3D-Slicer (v5.2.2) software platforms. A radiation oncologist with 2 years of experience in thoracic CT diagnosis manually delineated the 2D tumor contours and 3D GTVs. Subsequently, automated segmentation of the ipsilateral lung field was performed using the nnU-Net framework, followed by manual correction of missegmented regions in the bronchi, major vessels, bones, and mediastinum. Anatomical boundaries were clarified using contrast-enhanced CT images whenever the tumor-normal tissue interface was ambiguous. The same physician repeated the segmentation in all cases after a 2-week interval to assess intra-observer variability. Additionally, a board-certified radiologist with 10 years of thoracic CT experience independently segmented a randomly selected subset (20%) of cases (Figure 2A) to assess interobserver agreement.

Figure 2 Research design. (A) Acquisition of original images and ROI outlining; (B) feature extraction and feature selection; (C) model building and evaluation. 2D, two-dimensional; 3D, three-dimensional; AUC, area under the curve; CI, confidence interval; CRL, Clinical-Radiomics-Lung integrated; GBM, Gradient Boosting Machine; KNN, k-Nearest Neighbors; LR, Logistic Regression; MLP, multilayer perceptron; RF, Random Forest; ROI, region of interest; SVM, Support Vector Machine; XGBoost, Extreme Gradient Boosting.

All CT images used for radiomics analysis were acquired prior to the initiation of CCRT. Therefore, radiation-induced lung injury, including radiation pneumonitis or fibrosis, was not present at the time of imaging and did not influence feature extraction. For ipsilateral whole-lung segmentation, the primary tumor volume was explicitly excluded from the lung mask. Automated lung segmentation was performed using the nnU-Net framework, followed by manual refinement to exclude non-parenchymal structures, including major pulmonary vessels, central airways, mediastinum, and bony structures. This step was performed to ensure that radiomics features reflected lung parenchymal heterogeneity rather than high-density anatomical structures, which are known to bias texture-based radiomics metrics. Segmentation consistency was further assessed through intra-observer repeat delineation and inter-observer validation on a randomly selected subset of cases, confirming the reliability of the final lung parenchyma masks used for analysis.

Feature extraction

Handcrafted features can be further divided into three types: (I) geometry; (II) intensity; (III) texture. Geometric features are used to describe the 3D shape of the tumor. Intensity Features describe the 1st order statistical distribution of voxel intensities within the tumor. Texture features are descriptions of patterns or 2nd and higher order intensity spatial distribution. In this work, the texture feature extraction was mainly done by Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level Dependence Matrix (GLDM), Gray-Level Size Zone Matrix (GLSZM), Neighboring Gray Tone Difference Matrix (NGTDM). All features were independently derived from three image-processing domains: original image space, logarithm-transformed space, and wavelet-transformed space (Figure 2B).

Feature selection

We used the Mann-Whitney U test for statistical analysis and feature screening in all radiomics features. Only radiomics features (P<0.05) were kept, and we also got the correlation between those who have high repeatability with Spearman’s rank correlation coefficient and choose a feature with a greater than 0.9 correlation coefficient. We used a greedy recursive deletion strategy for the feature filter—We removed the feature which had the most redundancy from the current set of features until We had retained as much of the full set of features as possible.

Secondly, least absolute shrinkage and selection operator (LASSO) regression model is implemented on discovered data using the python scikit-learn package to get the signature. LASSO is to shrink each of the regression coefficients back by λ, and set some of the regression coefficients of irrelevant features to be exactly 0. So, we used 10-fold cross validation with the min criterion to find the proper lambda value, the one that gives the minimum cross-validation error. The features that were extracted that had non-zero values were used to go through the next steps in the regression model fit. These were then combined into a radiomics signature. Finally, for each patient, a radiomics score was generated as a linear combination of features that was retained with the model coefficients as the linear combination’s coefficients (Figure 2C, Figure S1). Variables included in the multivariate Logistic Regression (LR) analysis were preselected based on clinical relevance and statistical significance in univariate analysis (P<0.05). To avoid model overfitting given the limited sample size of the training cohort, only variables demonstrating independent prognostic potential were entered into the multivariate model. This parsimonious modeling strategy is consistent with standard recommendations for multivariable regression analyses in clinical studies.

Rad signature

Three independent radiomic models were constructed based on the 2D region of interest (ROI), 3D ROI, and ipsilateral whole-lung ROI. We input the final features for each ROI into machine learning models such as LR, Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), k-nearest neighbor (KNN), Extremely Randomized Trees (Extra Trees), Light Gradient Boosting Machine (LightGBM), and Multilayer Perceptron (MLP) for risk model construction. This was followed by 5-fold cross verification to obtain the final read signature.

Clinical signature

The clinical signature was constructed similarly to that of the Rad signature. First, the features used for building clinical signatures were selected using univariate and multivariate analyses, with a P<0.05. Next, the same machine-learning models used in the radio-signature building process were used to screen the features. Finally, 5-fold cross validation was performed and the test cohort was fixed for fair comparison.

Integrated signature

An integrated model was developed by combining features derived from the 3D ROI, ipsilateral whole-lung ROI, and clinical characteristics. The same set of machine learning models used for the radiomic and clinical signatures was employed here to ensure a fair comparison.

Model evaluation

We used different commonly used evaluation metrics like the area under curve of receiver operating characteristic (AUC), sensitivity, specificity, accuracy and so on to judge whether the model is reliable. We employed decision curve analysis (DCA) to evaluate the clinical usefulness of the predictive versions: In the radiomics part, the cut-off value distinguishing between low- and high-score cases was maximized in the Youden index using the best algorithm among several compared. progression-free survival (PFS) and OS were compared between the two groups, respectively: PFS was defined as the time from the start of CCRT until disease progression or death from any cause occurred, and OS was defined as the interval from the beginning of CCRT until death occurred. Kaplan-Meier was conducted for PFS and OS analysis. Kaplan-Meier curves are compared by a log-rank test (Figure 2C).

Results

Characteristics of patients

We assessed differences in clinical characteristics using appropriately selected statistical tests: independent samples t-tests for normally distributed continuous variables, Mann-Whitney U test for non-normally distributed continuous variables, and Chi-squared test for categorical variables. Table 1 is the baseline clinic characteristics of the patients in our cohort. The study cohort comprised 164 participants: 131 (79.88%) were in the training cohort, while 33 (20.12%) were in the validation cohort. The inter-group differences in sex and the numbers of mediastinal lymph nodes were statistically significant (P<0.05). However, no significant differences (P>0.05) were detected in ethnicity, age, smoking status, diabetes mellitus, pulmonary tuberculosis, KPS scores, pathological type, oligometastasis, T stage, N stage, M stage, chemotherapy regimens, or number of chemotherapy cycles administered during CCRT, carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), and cytokeratin 19 fragment (CYFRA 21-1).

The median follow-up time for the entire cohort was 16.3 months (interquartile range, 8.6–24.5 months). For patients enrolled in the later years of the study period [2022–2023], follow-up duration was shorter, particularly for OS assessment. Consequently, OS results should be interpreted with caution, and PFS may represent a more robust endpoint in this cohort.

To minimize bias arising from heterogeneous follow-up durations, survival analyses were conducted using time-to-event methods with appropriate censoring, and no patients were excluded based on follow-up length.

Clinical variables associated with prognosis in CCRT

In univariate analysis, clinical M stage [odds ratio (OR), 2.33; 95% confidence interval (CI): 1.20–4.53; P<0.05] and received chemotherapy cycles during CCRT (OR, 0.45; 95% CI: 0.24–0.86; P<0.05), CYFRA21-1 (OR, 2.55; 95% CI: 1.32–4.93; P<0.05) were associated with prognosis in the training dataset. In terms of multivariate analysis, the number of chemotherapy cycles in CCRT and CYFRA21-1 were both statistically significant and therefore are independent predictive factors for CCRT treatment effect (Table S1). Then I produced eight models where the number of chemotherapy cycles given during CCRT and CYFRA21-1. In which XGboost model was 0.683 (95% CI: 0.596–0.770) in the training groups and 0.627 (95% CI: 0.439–0.815) in the validating groups. Model’s AUC as in Figure S2A,S2B. The accuracy and the specificity and sensitivity of 8 models illustrated in Table S2.

Construction of predictive models and performance evaluation

The extraction of 1,051 features was carried out separately for 2D, 3D and ipsilateral whole-lung ROIs. Eight prediction models were constructed by screening variables based on feature screening. The XGBoost model performs well. In both cohorts, the XGBoost AUC of Lung-Rad model is 0.710 (95% CI: 0.530–0.888). This AUC is bigger than that of 2D-Rad and 3D-Rad models (AUCs: 0.678, 0.686 in validation cohort). In terms of all evaluation models, the combined model 3D-CRL formed by the 3D-ROI, ipsilateral whole-lung ROI, and clinical features has the best performance for evaluating all three aspects. In all models’ key parameters (AUC, sensitivity, and specificity), the 3D-CRL key parameters are always larger than other models. XGBoost AUC of 3D-CRL models were respectively 0.966 (95% CI: 0.941–0.992) and 0.816 (95% CI: 0.661–0.972) (Figure 3, Table 2). From Figure 4, the distribution of the prediction score of the XGBoost prediction model on the PFS sample distribution shows that the model is well done at the classification task. Each model’s 2D-Rad, 3D-Rad, Lung-Rad, 3D-CRL AUCs are illustrated in Figure S2C-S2J. All 3 models’ accuracy, specificity, and sensitivity can be found in Tables S3-S6. To further assess whether the differences in discriminative performance between models were statistically significant, pairwise comparisons of AUCs were performed using the DeLong test. The results showed that the integrated 3D-CRL model achieved a significantly higher AUC than the clinical model (C), 3D-Rad, and Lung-Rad models in the validation cohort (all P<0.05). In contrast, no statistically significant difference was observed between the 2D-Rad and 3D-Rad models nor between the 2D-Rad model and the 3D-CRL model (P>0.05). Although the AUC between the ensemble model and the 2D-Rad model was not statistically significant, the ensemble model demonstrated superior specificity and sensitivity. Moreover, while learning the features from the training data, the ensemble model could effectively prevent overfitting, thereby exhibiting better generalization ability on unseen data (Figure S7).

Figure 3 Performance of five developed models for predicting the prognosis of CCRT for NSCLC in all datasets. (A) Training dataset; (B) validation dataset. 2D, two-dimensional; 3D, three-dimensional; AUC, area under the receiver operating characteristic curve; C, clinical model; CCRT, concurrent chemoradiotherapy; NSCLC, non-small cell lung cancer.

Figure 4 Distribution of the prediction scores of the XGBoost model for PFS. (A) Training dataset; (B) validation dataset. Blue indicates failure to achieve progression-free survival, while red denotes progression-free survival attainment. PFS, progression-free survival; XGBoost, Extreme Gradient Boosting.

PFS and OS analysis of combined models

The optimal cut-off value of 0.57 was determined through receiver operating characteristic (ROC) curve analysis of the 3D-CRL model. This threshold, identified by maximizing the Youden index, stratified patients with NSCLC into two prognostic groups: a low-risk cohort (score <0.57) and a high-risk cohort (score ≥0.57). The Kaplan-Meier method was used to evaluate the correlation between the 3D-CRL combined model and the PFS and OS of patients with NSCLC. Survival curves showed significant differences in PFS and OS between the two groups (log-rank test, P<0.05) (Figure 5). DCA further confirmed that the combined model had a superior performance in most threshold ranges compared to the other models, indicating that this model had better clinical decision-making guidance value (Figure 6).

Figure 5 Survival curve analysis of progression-free survival (A,B) and overall survival (C,D) for patients in the high- and low-risk groups in the training and validation cohorts of 3D-CRL model. 3D, three-dimensional; CI, confidence interval; HR, hazard ratio.

Figure 6 DCA of patients in the training and validation cohorts in the 2D-Rad, 3D-Rad, Lung-Rad, 3D-CRL models. 2D, two-dimensional; 3D, three-dimensional; C, clinical model; DCA, decision curve analysis.

Discussion

Key findings

CCRT is a crucial treatment modality for patients with unresectable and advanced NSCLC and plays an important role in improving patients’ quality of life and prolonging their survival. However, the accurate prognostic prediction of patients with advanced NSCLC undergoing CCRT remains challenging. In this study, we demonstrated that tumor-based and lung-based radiomic features can predict prognosis of patients with NSCLC receiving CCRT. Specifically, we developed a novel survival prediction model that innovatively integrates 3D ROI, ipsilateral whole-lung ROI, and clinical characteristics. The validation results showed that this model demonstrates competitiveness in terms of AUC and prediction accuracy and has significant advantages over traditional prediction methods.

Strengths

In this study, we extracted the radiomics features of the ROI in the ipsilateral whole lung (rather than the whole lung) as the biological effects of CCRT are dominated by the lesion area. This is because tissue changes induced by radiation (such as inflammation and vascular injury) are confined to the irradiated area, and their features are directly related to treatment response (26,27). Additionally, vascular permeability increases after radiotherapy, in addition to an increase in the local concentration of chemotherapeutic drugs and permeability of the TME on the injured side (27,28). Furthermore, adding contralateral features introduces noise and blurs key treatment-specific signals. Consequently, this approach ensured the simplicity of the model while capturing the main biological determinants of prognosis.

Comparison with similar research

This study systematically compared the predictive performance of radiomic features in different dimensions and found that 3D ROI features based on whole tumor volume outperformed 2D ROI features in survival prediction. This advantage stems from the ability of the 3D ROI to comprehensively quantify intratumoral heterogeneity. Specifically, the 3D ROI captures the spatial distribution characteristics of substructures, such as the necrotic core and enhanced area, fundamentally avoiding the selection bias inherent in 2D single-slice analysis (29-32). Additionally, we found that the efficacy of the radiomics model based on ipsilateral whole-lung features for survival prediction was significantly higher than that of a radiomics model based on tumor features due to fundamental biological differences between the two feature types. The features of the tumor ROI directly map the malignant phenotypes of cancer cells, and their spatially heterogeneous distribution is closely related to prognosis (33-36). However, the model failed to account for the prognostic impact of TME dynamics and spatial heterogeneity in lung tissue architecture. In addition, the combined model integrated tumor imaging features focusing on the lesion area, anatomical features of the lung tissue that characterize the tumor microenvironment and pulmonary complications, and clinical features, which significantly improved its predictive performance (AUC =0.816) and confirmed the existence of systematic complementary information among cross-domain data sources.

Explanations of findings

Radiomics turns medical images into high-dimensional quantitative information and is a very useful tool for predicting the response to radiotherapy and the prognosis of the disease (14,37,38). But so far, most studies have only studied the radiomics of the whole tumor, but did not study the relationship between the radiomics of the whole ipsilateral lung outside the tumor and prognosis. As for this part, lots of recent investigations point out that lung tissue’s different kinds have something to do with the OS (21,39-41). So, we combined the ipsilateral whole lung stuff and what was inside the tumor, because we found that it could predict even better. It was also applying separate CCRT on patients with different risk of survival with different groups based upon combined-model thresholds. The high-risk group had much smaller PFS and OS rates than the low-risk group. These might allow for finding patients straight up who won’t probably profit much from CCRT, so their best ways of getting treated can be set right away.

Limitations and actions needed

Despite the multi-center study design, this study has some limitations. First, the retrospective nature of the study led to a lack of external prospective validation, and the size of the validation cohort was limited, which may have caused model overfitting. Second, the sample size was relatively small; therefore, validation in a large-scale multi-ethnic cohort is required to improve the generalizability of the model. Third, multimodal data such as genomics (e.g., driver gene mutations and tumor mutation burden) were not integrated into our model, which limited the completeness of the prediction dimensions of the model. Finally, while classical machine learning methods were employed in this study, contemporary deep learning architectures have demonstrated superior capability in capturing high-dimensional spatial dependencies within medical imaging data. Consequently, future investigations should implement end-to-end deep learning frameworks to completely exploit latent feature representations, thereby potentially augmenting the model predictive accuracy.

Statistical robustness and overfitting control

We acknowledge the apparent performance discrepancy between the training and validation cohorts observed in the tumor-based radiomic models (2D-Rad and 3D-Rad), where the AUC values decreased substantially from the training set (0.936 and 0.908, respectively) to the validation set (0.678 and 0.686). This phenomenon reflects the intrinsic risk of overfitting in high-dimensional radiomic analyses, particularly when the feature-to-sample ratio is large and the validation cohort is limited in size.

Several methodological considerations mitigate the concern that the proposed modeling framework lacks robustness. First, feature selection was performed using a strictly nested pipeline combining univariate filtering, redundancy reduction based on Spearman correlation, and LASSO regression with 10-fold cross-validation, which has been widely adopted to control model complexity and reduce overfitting in radiomics studies. Importantly, the observed performance degradation was most pronounced in tumor-centric models, whereas the ipsilateral whole-lung radiomic model demonstrated relatively stable discrimination across cohorts (training AUC =0.782; validation AUC =0.710), suggesting that lung parenchymal features capture more generalizable biological information beyond tumor-localized heterogeneity.

Second, although individual component models exhibited instability, the integrated 3D-CRL model consistently achieved superior and more stable performance (training AUC =0.966; validation AUC =0.816). This improvement indicates that complementary information from tumor morphology, lung microenvironmental heterogeneity, and clinical variables effectively compensates for the variance introduced by any single feature domain. Such cross-domain integration is a recognized strategy to enhance generalizability in radiomics-based prognostic modeling.

Third, we recognize that the validation cohort (n=33) is relatively small for a multicenter study and may be susceptible to statistical noise. However, despite this limitation, the combined model preserved clinically meaningful discrimination and demonstrated consistent risk stratification in PFS and OS analyses, supporting its potential clinical relevance. Nonetheless, the reported AUC values should be interpreted as preliminary estimates, and prospective validation in larger, independent cohorts is necessary to further confirm the robustness and reproducibility of the proposed model.

Cohort imbalance and potential selection bias

For the imbalance in baseline characteristics between the training and validation cohorts. As shown in Table 1, significant differences were observed in sex distribution (P=0.02) and the prevalence of mediastinal lymph node invasion (P=0.002), with the validation cohort exhibiting a markedly higher proportion of nodal invasion (93.94%) compared with the training cohort (67.18%).

Such imbalance may introduce potential selection bias and complicate direct performance comparison between cohorts. However, it is important to note that the validation cohort represents a clinically more advanced and homogeneous high-risk population, rather than a randomly sampled subset of the training distribution. From a methodological perspective, this cohort shift constitutes a more stringent and conservative test of model generalizability, as predictive models trained on relatively heterogeneous disease stages are evaluated on patients with more aggressive nodal involvement.

Importantly, despite this unfavorable imbalance, the integrated 3D-CRL model maintained stable discriminatory performance in the validation cohort (AUC =0.816) and preserved significant stratification of PFS and OS. This observation suggests that the proposed model is not solely driven by nodal stage or sex-related confounding, but rather captures complementary prognostic information from tumor radiomics, ipsilateral lung parenchymal heterogeneity, and clinical variables.

Nevertheless, we recognize that the imbalance limits causal inference and may affect absolute performance estimates. Accordingly, the present findings should be interpreted with caution, and future studies employing balanced sampling strategies, propensity score matching, or stratified external validation across nodal stages and sex subgroups are warranted to further confirm model robustness.

Biological interpretation of ipsilateral whole-lung radiomics

The prognostic value of ipsilateral whole-lung radiomics may be explained by a combination of oncological and non-oncological mechanisms that extend beyond tumor-localized heterogeneity (42,43).

From an oncological perspective, lung parenchymal radiomic features may capture spatially diffuse tumor-host interactions that are not confined to the primary tumor volume. Microscopic tumor infiltration, occult lymphovascular spread, and field cancerization effects can induce subtle but measurable alterations in the surrounding lung tissue, which are reflected as texture and intensity heterogeneity on CT imaging (44). Moreover, intrathoracic tumor burden and nodal involvement may influence regional ventilation-perfusion patterns and tissue density, indirectly shaping lung radiomic signatures associated with aggressive disease biology and unfavorable survival outcomes.

From a non-oncological perspective, ipsilateral lung radiomics may reflect baseline pulmonary vulnerability, including subclinical emphysema, fibrosis, or inflammatory remodeling, even in the absence of overt chronic lung disease. Such parenchymal conditions can reduce pulmonary reserve and increase susceptibility to treatment-related toxicity during CCRT, thereby affecting treatment tolerance, completion of chemotherapy cycles, and ultimately survival. In addition, lung radiomic heterogeneity may capture variations in vascular architecture, hypoxia-related changes, and immune-stromal interactions, which are increasingly recognized as important determinants of radiosensitivity and systemic treatment response (45,46).

Importantly, the observed stability of ipsilateral whole-lung radiomics across cohorts suggests that lung parenchymal features may encode more generalizable host-related information compared with tumor-centric radiomics alone. The improved performance of the integrated 3D-CRL model further supports the concept that combining tumor morphology, lung microenvironmental heterogeneity, and clinical variables enables a more comprehensive representation of disease aggressiveness and patient resilience, ultimately leading to improved survival prediction.

Conclusions

This study provides the first evidence that ipsilateral whole-lung radiomics has an independent prognostic value, beyond tumor-focused features, in patients with NSCLC undergoing CCRT. We developed and validated a 3D-CRL prediction model based on radiomics, which integrates tumor and pulmonary radiomic features, and clinical features, and can effectively predict the prognosis of patients with NSCLC receiving CCRT. Compared to the single-feature models, this integrated model demonstrated the best predictive performance and could accurately distinguish patient groups with different risk levels.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2586/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2586/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2586/prf

Funding: This work was supported by the Guizhou Medical University National Natural Science Foundation Cultivation Project (No. 25NSFCP35), Qian Dong Nan Science and Technology Program (No. qdnkhJz [2023] 14), Public Hospital High-Quality Development Research Public Welfare Project Fund (No. GL-A014), Cultivation of High-Level Innovative Talents in Guizhou Province (No. qian qian ceng ren cai[2024]202215), Spark Program (No. XHJH-0048), Chen Xiao-Ping Foundation for The Development of Science and Technology of Hubei Province (No. CXPJJH125009-05), Wu Jieping Medical Foundation (No. 320. 6750.2025-16-21), and the Scientific Research Project of Guizhou Provincial Health and Wellness Commission (No. gzwkj2025-608).

Conflicts of Interest: All the authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1-2586/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work and ensure that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the principles of the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of The Second Affiliated Hospital, Guizhou Medical University (approval No. 2020-LS-03), and individual consent for t his retrospective analysis was waived. The other institutions we re also informed and approved the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Miller KD, Jemal A. Cancer Statistics, 2017. CA Cancer J Clin 2017;67:7-30. [Crossref] [PubMed]
2. Chen VW, Ruiz BA, Hsieh MC, et al. Analysis of stage and clinical/prognostic factors for lung cancer from SEER registries: AJCC staging and collaborative stage data collection system. Cancer 2014;120:3781-92. [Crossref] [PubMed]
3. Curran WJ Jr, Paulus R, Langer CJ, et al. Sequential vs. concurrent chemoradiation for stage III non-small cell lung cancer: randomized phase III trial RTOG 9410. J Natl Cancer Inst 2011;103:1452-60. [Crossref] [PubMed]
4. Ma H, Peng F, Xu Y, et al. Five-year survival rate analysis: the combination of fortnightly-administration of endostar and concurrent chemoradiotherapy versus concurrent chemoradiotherapy in the treatment of inoperable locally advanced non-small cell lung cancer. Ann Palliat Med 2021;10:7560-70. [Crossref] [PubMed]
5. Ghaderi N, Jung JH, Odde DJ, et al. Clinically validated model predicts the effect of intratumoral heterogeneity on overall survival for non-small cell lung cancer (NSCLC) patients. Comput Methods Programs Biomed 2021;212:106455. [Crossref] [PubMed]
6. Tang Y, Wang T, Li Q, et al. A cuproptosis score model and prognostic score model can evaluate clinical characteristics and immune microenvironment in NSCLC. Cancer Cell Int 2024;24:68. [Crossref] [PubMed]
7. Ying Q, Fan R, Shen Y, et al. Small Cell Lung Cancer-An Update on Chemotherapy Resistance. Curr Treat Options Oncol 2024;25:1112-23. [Crossref] [PubMed]
8. Ferreira JR Junior, Koenigkam-Santos M, de Vita Graves C, et al. Quantifying intratumor heterogeneity of lung neoplasms with radiomics. Clin Imaging 2021;74:27-30. [Crossref] [PubMed]
9. Wang X, Xie T, Luo J, et al. Radiomics predicts the prognosis of patients with locally advanced breast cancer by reflecting the heterogeneity of tumor cells and the tumor microenvironment. Breast Cancer Res 2022;24:20. [Crossref] [PubMed]
10. Henry T, Sun R, Lerousseau M, et al. Investigation of radiomics based intra-patient inter-tumor heterogeneity and the impact of tumor subsampling strategies. Sci Rep 2022;12:17244. [Crossref] [PubMed]
11. Peng J, Kang S, Ning Z, et al. Residual convolutional neural network for predicting response of transarterial chemoembolization in hepatocellular carcinoma from CT imaging. Eur Radiol 2020;30:413-24. [Crossref] [PubMed]
12. Peng J, Zhang J, Zou D, et al. Deep learning to estimate durable clinical benefit and prognosis from patients with non-small cell lung cancer treated with PD-1/PD-L1 blockade. Front Immunol 2022;13:960459. [Crossref] [PubMed]
13. Perez-Johnston R, Araujo-Filho JA, Connolly JG, et al. CT-based Radiogenomic Analysis of Clinical Stage I Lung Adenocarcinoma with Histopathologic Features and Oncologic Outcomes. Radiology 2022;303:664-72. [Crossref] [PubMed]
14. Peng J, Zhang X, Hu Y, et al. Deep learning to estimate response of concurrent chemoradiotherapy in non-small-cell lung carcinoma. J Transl Med 2024;22:896. [Crossref] [PubMed]
15. Peng J, Huang J, Huang G, et al. Predicting the Initial Treatment Response to Transarterial Chemoembolization in Intermediate-Stage Hepatocellular Carcinoma by the Integration of Radiomics and Deep Learning. Front Oncol 2021;11:730282. [Crossref] [PubMed]
16. Aerts HJ, Velazquez ER, Leijenaar RT, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006. [Crossref] [PubMed]
17. Wang X, Cui H, Gong G, et al. Computational delineation and quantitative heterogeneity analysis of lung tumor on 18F-FDG PET for radiation dose-escalation. Sci Rep 2018;8:10649. [Crossref] [PubMed]
18. Shen C, Liu Z, Guan M, et al. 2D and 3D CT Radiomics Features Prognostic Performance Comparison in Non-Small Cell Lung Cancer. Transl Oncol 2017;10:886-94. [Crossref] [PubMed]
19. Huang Y, Liu Z, He L, et al. Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016;281:947-57. [Crossref] [PubMed]
20. Chen NB, Xiong M, Zhou R, et al. CT radiomics-based long-term survival prediction for locally advanced non-small cell lung cancer patients treated with concurrent chemoradiotherapy using features from tumor and tumor organismal environment. Radiat Oncol 2022;17:184. [Crossref] [PubMed]
21. Zhai R, Yu X, Shafer A, et al. The impact of coexisting COPD on survival of patients with early-stage non-small cell lung cancer undergoing surgical resection. Chest 2014;145:346-53. [Crossref] [PubMed]
22. Lim JU, Yeo CD, Rhee CK, et al. Overall survival of driver mutation-negative non-small cell lung cancer patients with COPD under chemotherapy compared to non-COPD non-small cell lung cancer patients. Int J Chron Obstruct Pulmon Dis 2018;13:2139-46. [Crossref] [PubMed]
23. Dong W, Zhu Y, Du Y, et al. Impact of severe-to-very severe chronic obstructive pulmonary disease on the prognosis of patients with non-small cell lung cancer who received chemotherapy. Clin Respir J 2020;14:345-52. [Crossref] [PubMed]
24. Ueda T, Aokage K, Mimaki S, et al. Characterization of the tumor immune-microenvironment of lung adenocarcinoma associated with usual interstitial pneumonia. Lung Cancer 2018;126:162-9. [Crossref] [PubMed]
25. Sato T, Watanabe A, Kondo H, et al. Long-term results and predictors of survival after surgical resection of patients with lung cancer and interstitial lung diseases. J Thorac Cardiovasc Surg 2015;149:64-9, 70.e1-2.
26. Wijerathne H, Langston JC, Yang Q, et al. Mechanisms of radiation-induced endothelium damage: Emerging models and technologies. Radiother Oncol 2021;158:21-32. [Crossref] [PubMed]
27. Huang RX, Zhou PK. DNA damage response signaling pathways and targets for radiotherapy sensitization in cancer. Signal Transduct Target Ther 2020;5:60. [Crossref] [PubMed]
28. Huber RM, Flentje M, Schmidt M, et al. Simultaneous chemoradiotherapy compared with radiotherapy alone after induction chemotherapy in inoperable stage IIIA or IIIB non-small-cell lung cancer: study CTRT99/97 by the Bronchial Carcinoma Therapy Group. J Clin Oncol 2006;24:4397-404. [Crossref] [PubMed]
29. Meng L, Dong D, Chen X, et al. 2D and 3D CT Radiomic Features Performance Comparison in Characterization of Gastric Cancer: A Multi-Center Study. IEEE J Biomed Health Inform 2021;25:755-63. [Crossref] [PubMed]
30. Xie XJ, Liu SY, Chen JY, et al. Development of unenhanced CT-based imaging signature for BAP1 mutation status prediction in malignant pleural mesothelioma: Consideration of 2D and 3D segmentation. Lung Cancer 2021;157:30-9. [Crossref] [PubMed]
31. Xu L, Yang P, Yen EA, et al. A multi-organ cancer study of the classification performance using 2D and 3D image features in radiomics analysis. Phys Med Biol 2019;64:215009. [Crossref] [PubMed]
32. Xie R, Pan D, Zeng A, et al. Target area distillation and section attention segmentation network for accurate 3D medical image segmentation. Health Inf Sci Syst 2023;11:9. [Crossref] [PubMed]
33. Tunali I, Hall LO, Napel S, et al. Stability and reproducibility of computed tomography radiomic features extracted from peritumoral regions of lung cancer lesions. Med Phys 2019;46:5075-85. [Crossref] [PubMed]
34. Kapoor-Narula U, Lenka N. Cancer stem cells and tumor heterogeneity: Deciphering the role in tumor progression and metastasis. Cytokine 2022;157:155968. [Crossref] [PubMed]
35. Ganguly A, Mukherjee S, Spada S. Editorial: Spatial immune cell heterogeneity in the tumor microenvironment. Front Immunol 2024;15:1377532. [Crossref] [PubMed]
36. Malta TM, Sokolov A, Gentles AJ, et al. Machine Learning Identifies Stemness Features Associated with Oncogenic Dedifferentiation. Cell 2018;173:338-354.e15. [Crossref] [PubMed]
37. Peng J, Zou D, Zhang X, et al. A novel sub-regional radiomics model to predict immunotherapy response in non-small cell lung carcinoma. J Transl Med 2024;22:87. [Crossref] [PubMed]
38. Zhou C, Hou L, Tang X, et al. CT-based radiomics nomogram may predict who can benefit from adaptive radiotherapy in patients with local advanced-NSCLC patients. Radiother Oncol 2023;183:109637. [Crossref] [PubMed]
39. Wang P, Zhu M, Zhang D, et al. The relationship between chronic obstructive pulmonary disease and non-small cell lung cancer in the elderly. Cancer Med 2019;8:4124-34. [Crossref] [PubMed]
40. Schussler O, Bobbio A, Dermine H, et al. Twenty-Year Survival of Patients Operated on for Non-Small-Cell Lung Cancer: The Impact of Tumor Stage and Patient-Related Parameters. Cancers (Basel) 2022;14:874. [Crossref] [PubMed]
41. Lee SJ, Lee J, Park YS, et al. Impact of chronic obstructive pulmonary disease on the mortality of patients with non-small-cell lung cancer. J Thorac Oncol 2014;9:812-7. [Crossref] [PubMed]
42. Yan M, Zhang Z, Tian J, et al. Whole lung radiomic features are associated with overall survival in patients with locally advanced non-small cell lung cancer treated with definitive radiotherapy. Radiat Oncol 2025;20:9. [Crossref] [PubMed]
43. Luo T, Yan M, Zhou M, et al. Improved prognostication of overall survival after radiotherapy in lung cancer patients by an interpretable machine learning model integrating lung and tumor radiomics and clinical parameters. Radiol Med 2025;130:96-109. [Crossref] [PubMed]
44. Zhu B, Tapinos A, Koka H, et al. Genomes and epigenomes of matched normal and tumor breast tissue reveal diverse evolutionary trajectories and tumor-host interactions. Am J Hum Genet 2024;111:2773-88. [Crossref] [PubMed]
45. Huang Q, Xu N, Yin J, et al. Radiomics reveals the biological basis for non-small cell lung cancer prognostic stratification by reflecting tumor immune microenvironment heterogeneity. Front Immunol 2025;16:1708692. [Crossref] [PubMed]
46. Liu Y, Pan Y, Wang Q, et al. Integration of intratumoral/peritumoral radiomics and deep learning for predicting overall survival in non-small cell lung cancer patients: a multicenter study. Front Oncol 2025;15:1669200. [Crossref] [PubMed]