Multi-institutional development and validation of habitat imaging for predicting outcomes of first-line immunotherapy in advanced non-small cell lung cancer

Introduction

Lung cancer is one of the leading causes of cancer-related death worldwide, with non-small cell lung cancer (NSCLC) accounting for approximately 85% of new cases (1). Immunotherapy, a promising new treatment modality, has shown encouraging outcomes in lung cancer, particularly among patients with advanced disease (1,2). Unlike traditional chemotherapy which directly targets and destroys cancer cells, immunotherapy harnesses the patient’s immune system to inhibit tumor growth (3). Currently, immune checkpoint inhibitors (ICIs) represent the most extensively studied immunotherapeutic agents for lung cancer, functioning primarily by blocking inhibitory signals on T cells (4,5). Although ICI therapies have revolutionized cancer treatment, not all patients derive benefit (6), and these treatments can also lead to severe immune-related side effects, such as pneumonia. Consequently, it is essential to identify patients most likely to benefit from immunotherapy and to monitor predictive biomarkers of treatment response (7).

At present, the most commonly used biomarker for predicting the effectiveness of ICIs is anti-programmed death receptor 1 (anti-PD-1); however, fewer than 30% of patients experience long-term survival benefits from this treatment (8). In addition, although patients with high programmed death-ligand 1 (PD-L1) expression typically show response rates below 50% when treated with ICIs alone, those with low PD-L1 expression may still benefit from ICIs—especially when combined with chemotherapy (9). The tumor proportion score (TPS) is a widely used clinical method for assessing PD-L1 expression in NSCLC, reflecting the percentage of tumor cells exhibiting any degree of PD-L1 (10). Nonetheless, there are limitations with these biomarkers. Due to the heterogeneity of the tumor microenvironment, a single biopsy may not capture the spatial characteristics of the entire tumor, and in some cases, biopsy may not be feasible because of the tumor’s location (11). Therefore, there is a pressing need for novel biomarkers that can more accurately identify patients who are likely to benefit from immunotherapy and guide more precise treatment strategies.

Recent real-world evidence and retrospective analyses have identified multiple novel prognostic factors and predictive models for treatment response in NSCLC. The Δcycle threshold values derived from the super-amplification refractory mutation system (ARMS) quantitatively reflect epidermal growth factor receptor (EGFR) mutation abundance and demonstrate predictive value for clinical outcomes of first-line EGFR-tyrosine kinase inhibitors (TKIs) therapy (12). Real-world effectiveness of PD-1/PD-L1 inhibitors aligns with clinical trial findings, with progression-free survival (PFS) showing significant associations with Eastern Cooperative Oncology Group (ECOG) performance status, KRAS mutation status, PD-L1 TPS ≥50%, and incidence of immune-related adverse events (irAEs) (13). The prognostic immune-inflammatory-nutritional (PIIN) score has been validated as an independent prognostic indicator for stage I–III NSCLC, while nomograms incorporating PIIN score, TNM stage and forced expiratory volume in 1 second (FEV1%) demonstrate robust predictive performance (14). Machine learning models like LightGBM, trained on SEER data, have accurately predicted 1-, 3- and 5-year survival in early-stage NSCLC and may aid in individualized treatment decision-making (15).

Vascular normalization is closely related to the efficacy of tumor immunotherapy (16). Studies have shown that, under specific conditions, both anti-angiogenic drugs and ICIs can induce vascular normalization within tumors (17,18). This normalization enhances intratumoral blood flow, reduces hypoxia, and improves the delivery of therapeutic agents and immune cells, thereby ultimately boosting the efficacy of various treatments, including immunotherapy (19,20). Patients who experience improved tumor perfusion following anti-angiogenic therapy generally have better clinical outcomes (4). Moreover, in our neoadjuvant breast cancer clinical trial, we found that pre-treatment microvessel density—as an indirect marker of perfusion—was associated with the efficacy of bevacizumab (21).

Given that medical imaging can reveal the spatial heterogeneity of tumor structures, habitat analysis has emerged as an advanced imaging technique to capture regional differences within tumors, demonstrating clinical relevance across various cancer types (22-25). Habitat analysis extracts shared feature based on imaging biomarkers to identify distinct sub-regions, or “habitats”, within a tumor (26). Building on our previous work in developing predictive models for immunotherapy response, we recognize that, compared to non-contrast computed tomography (CT), arterial-phase contrast-enhanced CT (CECT)—with its rich vascular perfusion information—provides an opportunity for sub-tumor region delineation via habitat analysis (27,28). This study aims to leverage habitat analysis to examine the features within tumor sub-regions and to develop a model for predicting PFS in NSCLC patients following immunotherapy. In addition, the study compares the significance of differences in sub-region proportions between high-risk and low-risk groups. The model, which is based on arterial-phase CECT, has been externally validated in a dual-center cohort of patients. Ultimately, this research provides clinicians with a tool that captures spatial heterogeneity to identify NSCLC patients most likely to benefit from ICI therapy. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-554/rc).

Methods

Study population

The discovery cohort (training cohort) originated from the ORIENT-11 study (ClinicalTrials.gov: NCT03607539), which enrolled 397 NSCLC patients treated between August 23, 2018, and July 30, 2019. These patients provided CECT images, treatment response data, and PD-L1 TPS information for analysis. Among them, 128 patients with suitable CECT images were selected for Habitat Imaging analysis (refer to Appendix 1).

The external validation (EV) cohort consisted of real-world NSCLC patients treated with first-line immunotherapy at Institution 1: Xinqiao Hospital Third Military Medical University (XQH) between January 1, 2018 and December 31, 2021, and at Institution 2: Peking University Shenzhen Hospital (PKUSZ) between January 1, 2021 and July 1, 2023. A total of 161 patients had CECT images eligible for analysis, with 92 of these patients also having PD-L1 TPS data available providing adequate statistical power for validation (see Appendix 1) (29).

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Peking University Shenzhen Hospital (approval No. 2024100) and by the Institutional Review Board of Xinqiao Hospital (approval No. 2022-454-01). Given the retrospective nature of the study and full anonymization of patient data, the requirement for written informed consent was waived by both ethics committees.

Overall response rate (ORR) and PFS

In the discovery cohort, ORR and PFS data were obtained from a publicly available clinical trial database. ORR was defined as the proportion of patients with a confirmed complete response or partial response, while PFS was defined as the time from the start of the randomized cohort to disease progression for any reason. In the external validation cohort, tumor response was assessed at both centers by physicians with 14 and 15 years of experience, respectively, with tumor imaging evaluations performed every 12 weeks after baseline using CECT. In the discovery cohort, 55 patients experienced PFS events. In the external validation cohorts, 25 and 12 patients experienced PFS events in Validation Cohort 1 and Validation Cohort 2, respectively. Efficacy was determined according to the Response Evaluation Criteria in Solid Tumors (RECIST v1.1). Complete and partial responses were confirmed by repeat tumor imaging after the initial response was recorded, and survival status was monitored via telephone follow-up every 3 months by trained staff (30).

Image processing and segmentation

All enrolled images were processed and tumor segmented using 3D Slicer software (version 4.11.0; see Figure 1) (31). The resampling algorithm used was linear interpolation, and the voxel size obtained after resampling was 1 mm × 1 mm × 1 mm. To ensure the stability of image segmentation, a semi-automatic segmentation method was used, and the grey scale threshold was set to (−50 to 200 HU). Tumor segmentation was conducted by a radiologist with 11 years of experience in medical oncology image segmentation (Figure 1, step 1).

Figure 1 Study design for modeling of NSCLC patients using habitat imaging based radiomics analysis. Step 1: this phase includes patient inclusion and exclusion criteria, along with the processing of enhanced CT images. Preprocessing involves the 3D segmentation of CT images to prepare for subsequent analysis. Step 2: the framework employs a two-stage clustering process: (I) feature extraction is performed on pixel points from the tumor regions of each patient and pixel points from all patients are pooled and clustered to identify distinct tumor subregions. (II) Imaging radiomics features are then extracted from these subregions, followed by feature selection and model development. Step 3: multiple models were developed and compared based on TPS, clinical features, and hybrid features. PFS prognostic curves derived from habitat imaging radiomics features and TPS-based groupings are evaluated and compared using an external validation cohort. 3D, three-dimensional; CECT, contrast-enhanced computed tomography; CT, computed tomography; NSCLC, non-small cell lung cancer; PFS, progression-free survival; PKUSZ, Peking University Shenzhen Hospital; ROC, receiver operating characteristic; TPS, tumor proportion score; XQH, Xinqiao Hospital Third Military Medical University.

Habitat generation and K-means clustering

The heterogeneous subregions of the region of interest (ROI) considered as habitats were generated from the three-dimensional (3D) arterial-phase CECT images. Habitat features were first extracted for each pixel point in the ROI using a sliding window of size 3×3×3 to compute the local features for each voxel, and 19 features were extracted for each voxel, which included the mean, standard deviation, minimum, maximum, and the 10th, 25th, 50th, 75th, and 90th percentiles (32-34). The K-means clustering algorithm was then applied to the pixel values using the Euclidean distance metric for cluster at the aggregate level.

The objective of K-means is to minimize the sum of the squares of the distances between all points in the cluster and their corresponding center of mass. The objective function is:

J=∑i=1K∑x∈ci‖x−μ‖2[1]

where K is the number of clusters, i denotes the first set of the i clusters. x denotes the data point. μ_i is the first centre of mass (mean vector) of the i-th cluster. ‖x−μ‖2 denotes the data point x and the cluster center μ_i the square of the Euclidean distance.

To determine the optimal number of habitats, we varied the number of clusters from 3 to 9 and evaluated the clustering performance using the Calinski-Harabasz (CH) score, Silhouette (SC) score, and Davies-Bouldin (DB) score. Specifically, the CH score measures the ratio of between-cluster variance to within-cluster variance, with higher scores indicating better cluster separation; the SC score assesses the similarity of each sample to its own cluster relative to other clusters, with higher values indicating more compact and well-separated clusters; and the DB score evaluates the average similarity between each cluster and its most similar one, with lower scores signifying superior clustering performance.

Feature extraction and model building

The pixel values of the images were first rescaled to the range of [−1,024, 2,048]. Radiomics Imaging feature extraction was performed (35) toolkit across three sub-regions and seven feature categories, following the guidelines of the Imaging Biomarker Standardization Initiative (IBSI) (36). Intratumoral imaging analysis involved calculating histogram, geometric, and texture features from five log filters (sigma =1.0, 2.0, 3.0, 4.0, 5.0) images and the original CT images. These features included the gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and local binary pattern (LBP). The extracted radiomic feature data were normalized using Z-scores and further reduced using principal component analysis (PCA) method. The resulting data were refined using the least absolute shrinkage and selection operator (LASSO) algorithm, which identified ten radiomic features most relevant to ORR coefficients for machine learning modeling. Four classifiers—Support Vector Machine (SVM), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Random Forest—were applied to a discovery cohort of 128 patients. A five-fold cross-validation was conducted to evaluate model performance on the validation cohort. The LR model demonstrated the best performance and was subsequently used to model clinical samples, TPS, and a fusion of clinical, TPS, and habitat analysis radiomic features in the validation cohort. The results obtained from 92 patients in the validation cohort, after model evaluation, were analyzed using Kaplan-Meier (K-M) curves and compared with the K-M curves of subgroups with TPS greater than or equal to 50% and less than 50% (Figure 1). The diagnostic performance of the model was assessed using several metrics, including area under the receiver operating characteristic curve (AUC), specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV). Decision curve analysis (DCA) was performed to evaluate the net benefit of the model in predicting NAIC efficacy.

Statistical analysis

Statistical analyses were carried out using R (version 4.2.3) and Python (version 3.5.6; http://www.python.org). Independent samples t-tests were used to analyze continuous variables, which are presented as mean ± standard deviation. The Chi-squared test or Fisher’s exact test was applied for categorical variables, with results reported as ratios. The Pearson correlation test was used to examine the relationships between features, and highly correlated features were removed to avoid redundancy. If the correlation coefficient between any two features exceeded 0.9, only one feature was retained. LASSO regression was employed to select non-zero variables as key features, which were then incorporated into the prediction model. PFS as defined as the time from treatment initiation to disease progression or death. K-M survival analysis was employed to estimate PFS for each habitat group, with log-rank tests assessing statistical differences between groups. A significance level of P<0.05 was considered statistically significant.

Results

Patient demographics and clinical characteristics

According to the inclusion criteria, a total of 432 patients with stage IIIB–IV NSCLC who received first-line ICI combination therapy were screened. The exclusion criteria were as follows: 26 patients did not undergo enhanced chest CT before treatment, 18 patients had incomplete CT images or thick-section scans, and 12 patients underwent chest CT imaging more than one month prior to the start of treatment. Additionally, 19 patients were excluded due to incomplete clinical or pathological data. Ultimately, 220 patients were included in the study: 128 patients with multi-center data, 60 patients from Institution 1, and 32 patients from Institution 2. The discovery cohort comprised 128 patients from multiple centers, while the external validation cohort consisted of 92 patients from Institution 1 and Institution 2. Table 1 presents the general statistical analysis of patient data. Among the patients who received first-line ICI combination therapy, 111 achieved an objective response (OR), while 109 did not, resulting in an OR rate of 50.5%. The average PFS time for all patients was 9.3 months, with 9.89 months in the discovery cohort and 8.42 months in the external validation cohort. The study included 181 male patients (82%) and 39 female patients (18%). Multivariate analysis revealed statistically significant differences across various factors, including gender, smoking history, TPS, and disease stage. In contrast, no significant differences were observed between the two groups regarding disease type and age. Table 1 provides further detailed statistical results.

Tumor habitat subregions

The optimal clustering solution was assessed using three widely used evaluation metrics: the CH score, SC score, and DB score. These metrics provide critical insights into the quality and robustness of the clustering approach. Clustering solutions ranging from 3 to 9 clusters were systematically assessed. Based on the analysis of all three metrics, we identified 3 as the optimal number of habitats, as it consistently achieved the highest CH and SC scores (Figure 2).

Figure 2 A bar chart displaying three clustering evaluation metrics in all cohorts. The CH score on the left, the SC score in the center, and the DB score on the right, with the number of clusters as the horizontal coordinate below. CH, Calinski-Harabasz; DB, Davies-Bouldin; SC, Silhouette.

The resulting three habitat subregions displayed distinct and reproducible imaging patterns across cohorts, each correlating with differential responses to ICI therapy (Figure 3). A higher proportion of Cluster 2 was associated with early disease progression, whereas a predominance of Cluster 1 was linked to prolonged progression-free intervals (Figures 3,4).

Figure 3 Examples of habitat maps for patients with progression-free survival. Rows 1 and 2 illustrate habitat maps of patients who experienced disease progression after 6 and 2.26 months, respectively, characterized by a high volume of Cluster 2. In contrast, Rows 3 and 4 depict habitat maps of patients with progression-free survival of 18.3 and 12.4 months, respectively, showing a high volume of Cluster 1 and low volume of Cluster 2 (Cluster 1 is light gray, Cluster 2 is dark red, and Cluster 3 is dark gray). CECT, contrast-enhanced computed tomography; GLCM, gray level co-occurrence matrix.

Figure 4 Distribution presents box plots comparing three cluster groups in both discovery cohort and validation cohort: ratio_1_L vs. ratio_1_H, ratio_2_L vs. ratio_2_H, and ratio_3_L vs. ratio_3_H. The Mann-Whitney test was used to assess statistical significance between the cluster groups, with significance levels indicated by asterisks: P>0.05 (NS), P<0.01 (**), and P<0.001 (***). The results show a highly significant difference between ratio_1_L and ratio_1_H (P<0.001), a significant difference between ratio_2_L and ratio_2_H (P=0.008), and no significant difference between ratio_3_L and ratio_3_H (P=0.12). These findings suggest that the ratios differ significantly across the cluster groups. H, high risk; L, low risk; NS., not significant.

Habitat imaging model building

From each of the three tumor subregions (Habitat 1, 2, and 3), 1,834 radiomics features were extracted, yielding a total of 5,502 features. A t-test identified 189 features significantly associated with ORR. After removing highly correlated features (Pearson correlation coefficient >0.9), 110 features were retained.

LASSO regression was then applied for feature selection, resulting in 10 non-zero coefficient features, which were used to build a LR model. Figure 1 (step 2) illustrates the feature selection process, including MSE plots and cross-validated coefficients.

To visually demonstrate the relationship between imaging features and progression risk, a volcano plot was generated (Figure 5), highlighting key discriminative features from different subregions.

Figure 5 Volcano plot illustrating the correlation between imaging radiomics features and the risk of progression in all NSCLC patients. Data points are colored based on their significance, with clusters identified as Cluster 1 and Cluster 2. 3D, three-dimensional; FC, fold change; NSCLC, non-small cell lung cancer.

The final 10 selected habitat imaging radiomics features were integrated into a LR model to characterize intra-tumor heterogeneity. Additionally, clinically significant features identified through logistic regression screening were independently utilized to develop distinct LR prediction models (Figure S1). Finally, these three features were combined to construct a fused model, which demonstrated an AUC of 0.869 [95% confidence interval (CI): 0.797–0.939], with a sensitivity of 0.721, specificity of 0.857, PPV of 0.816, and NPV of 0.778 in the external validation cohort (Figure 6A and Table 2). The detail of internal and external validation cohort was demonstrated in Figure S2. Figure 6B shows the decision curves for the four prediction models in the external validation cohort. The combined prediction model offers a higher overall net benefit compared to the other models. Figure 6C illustrates the distribution of predicted OR status for all patients in the external validation cohort after first-line ICI combination therapy, as determined by the combined model. The therapeutic response predicted by the model aligns with the actual clinical therapeutic response status. The results obtained by habitat analysis imaging were grouped, and K-M curves were established for the groups with TPS ≥50%, and the P values were 0.007 and 0.03, respectively, as shown in Figure 7 (Figure 1, step 3). In addition, two patients with stage IV NSCLC were classified into different imaging subtypes, each with distinct risks of cancer progression, based on their preoperative CECT images, as shown in Figure 8. This risk stratification could help identify high-risk patients who may require closer monitoring and more personalized ICI-based treatments.

Figure 6 ROC curves for all features in the validation cohort (A), DCA (B), and waterfall plots of the combined model showing the predicted outcomes for each patient in the validation cohort (C). AUC, area under the receiver operating characteristic curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic; TPS, tumor proportion score.

Figure 7 Habitat analysis of the prognostic predictive role of imaging groupings for PFS in PD-1 patients. Kaplan-Meier curves comparing prognostic outcomes based on TPS groupings: 0–49% TPS (blue) versus ≥50% TPS (red), with a P value of 0.03 in the log-rank test for the validation cohort (A). Kaplan-Meier curves comparing the prognostic outcomes based on habitat imaging groupings: low risk (blue) and high risk (red), with a P value of 0.007 in the log-rank test for the validation cohort (B). PD-1, programmed death receptor 1; PFS, progression-free survival; TPS, tumor proportion score.

Figure 8 Example of intra-regional delineation of tumors on arterial-phase CECT based on habitat imaging. Patients with similar baseline clinical characteristics but differing TPS are classified into low- and high-risk groups. CT scans are shown with feature annotations based on voxel-level imaging characteristics displayed beneath the images. Clinical features are provided below, and the subregions of the 3D tumor, identified through K-means clustering, are also depicted. 3D, three-dimensional; CECT, contrast-enhanced computed tomography; CT, computed tomography; PFS, progression-free survival; TPS, tumor proportion score.

Discussion

This study highlights the transformative potential of habitat analysis in the stratifying patients with advanced stage IIIB–IV NSCLC, particularly in optimizing immunotherapy efficacy. In this multi-center study, we employed a habitat imaging framework based on arterial-phase CECT data to identify phenotypically distinct regions (habitats) within tumors. The imaging features from these intratumoral habitats showed strong predictive capabilities, effectively classifying patients into high-risk and low-risk groups, with an AUC value superior to models based on TPS ≥50% stratification. Additionally, K-M survival curve analysis of model results further demonstrated its predictive ability for long-term prognosis. Validation in an independent cohort showed that the habitat imaging subtype model maintained stable predictive performance. Although this study was a retrospective proof-of-concept study, it provides important guidance for future clinical applications and lays the groundwork for prospective studies.

Traditional clinical factors, TPS scores, and genetic testing can predict disease progression to some extent, but they often fail to fully account for the individual tumor microenvironment and its heterogeneity. Arterial-phase CECT, which is more sensitive in detecting metabolically active vascular regions, shows greater potential in early progression risk prediction compared to routine clinical examinations or conventional CT imaging radiomics. This study confirms that this method can identify high-risk progression patients, guiding personalized treatment strategies such as intensified therapy, close post-operative follow-up, and inclusion in prospective clinical trials for more aggressive systemic treatment.

A growing body of research indicates significant intratumoral heterogeneity being more aggressive and playing key roles in disease progression (37-39). Unsupervised clustering techniques, such as the K-means clustering algorithm, group based on the similarity of imaging radiomics voxel features, providing a way to quantify tumor heterogeneity (32). In this study, we found that a clustering number of three better delineated distinct habitats, the model could better define the boundaries of different habitats, showing superior inter-group differentiation and smaller intra-group variance. Key features were primarily concentrated in the h1 and h2 regions near the tumor boundary, suggesting that the high heterogeneity of tumor boundary areas may be associated with better immune therapy efficacy. This finding aligns with the study by Liu et al., who extracted an onion-mode perfusion (OMP) feature from CECT to build a predictive model, with an AUC significantly higher than TPS score models (27). Liao et al. combined tumor internal heterogeneity, peritumoral vascular radiomics, and clinical factors to improve the prediction accuracy of the immunotherapy response (40). Ye et al. successfully predicted pathological complete response (pCR) rates in NSCLC patients receiving neoadjuvant chemotherapy through CT radiomics habitat quantitative analysis (41). Yolchuyeva et al. emphasized the importance of feature selection methods combined with machine learning strategies in constructing robust survival prediction models (42). We further analyzed the proportional volumes of different habitat sub-regions for disease progression groups, showing significant differences, and presenting several clear examples. This type of sub-region volume proportion analysis has been less studied in previous work. This study is a multi-center habitat imaging genomics study based on arterial-phase CECT, focusing on advanced NSCLC patients. By dividing the entire tumor into multiple habitat sub-regions and analyzing their spatial interactions, we emphasize the impacts of tumor internal heterogeneity on immune therapy efficacy.

Compared with previous whole-tumor radiomics approaches, our study has the following advantages: first, traditional radiomics genomics studies often rely on large-scale feature extraction and screening to build predictive models. In contrast, this study employs an unsupervised clustering method based on tumor habitat features to define imaging radiomics subtypes, thus more accurately depicting the relationship between tumor internal structure and immune therapy response and improving model interpretability. By capturing the intrinsic structure of the data, this method is less susceptible to overfitting and exhibits superior robustness in small, heterogeneous cancer cohorts compared to traditional feature engineering. Second, this study used a multi-center discovery cohort and an external validation cohort from two centers, ensuring model generalizability. We normalized the imaging data and standardized the radiomic extraction process, ensuring that the external validation results from multi-center trials showed high consistency, further verifying the model’s stability.

In the external validation cohort, K-M survival analysis further confirmed that patients stratified by habitat radiomics imaging had better long-term survival prediction outcomes than those stratified by TPS ≥50%, further validating the clinical value of the habitat imaging genomics model in high- and low-risk patient stratification. Given that arterial-phase CECT is routinely performed in clinical practice, this habitat imaging genomics model can be used alone or combined with other biomarkers, greatly improving its clinical feasibility.

Although this study yielded important findings, it still has the following limitations: Studies have shown that immunotherapy can improve tumor vascular perfusion and is associated with increased CD8⁺ T cell counts, suggesting that habitat imaging genomics features may relate to different lymphocyte counts and states within the tumor. Future studies should further explore the relationship between habitat imaging and the tumor immune microenvironment. This study was based on retrospective data analysis of over 200 patients, which may introduce data bias. Prospective validation should be conducted in larger datasets. This study only focused on primary tumor lesions and did not include lymph nodes or distant metastatic lesions. Furthermore, different lung cancer subtypes, such as adenocarcinoma and squamous cell carcinoma, were not analyzed independently. Future research should further refine tumor classification and build more comprehensive predictive models. Although external validation was performed, large-scale prospective studies are still needed to confirm the generalizability of the model. In addition, its applicability across different ethnicities and geographic populations requires further investigation.

Conclusions

In conclusion, our study develops a novel model that combines spatial habitat imaging to predict the efficacy of first-line ICI-based therapy in NSCLC patients. By integrating TPS, clinical features, and spatial habitat imaging, this approach represents a new method that significantly improves predictive performance compared to using TPS alone. Grouping analysis and prognostic risk stratification indicate that the model can reflect the risk of disease progression. Our findings provide a non-invasive, personalized treatment tool, addressing the urgent need for early identification of NSCLC patients who would benefit from ICI-based therapy. However, further biological validation is required to confirm its clinical value across different subtypes and larger patient populations.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-554/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-554/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-554/prf

Funding: This work was supported by grants from the National Natural Science Foundation of China (Nos. 82473305 and 82150106 to Y.H.), Provincial-Level Talent Program for National Center of Technology Innovation for Biopharmaceuticals (No. NCTIB2024JS0101), Jiangsu Science and Technology Program (No. BK20241815 to S.W.), Key Program for Basic Research in Jiangsu Province (No. BK20243008 to Y.H.), Interdisciplinary Basic Frontier Innovation Program of Suzhou Medical College of Soochow University (No. YXY2303028 to Y.H.), the Yancheng Municipal Health Commission Medical Research Project Fund (No. YK2023061 to M.Z.), Shenzhen City San Ming project (No. SZSM202211036 to Y.L.), the Collaborative Innovation Center of Hematology, and the Priority Academic Program Development of Jiangsu Higher Education Institutions.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-554/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Peking University Shenzhen Hospital (approval No. 2024100) and by the Institutional Review Board of Xinqiao Hospital (approval No. 2022-454-01). Given the retrospective nature of the study and full anonymization of patient data, the requirement for written informed consent was waived by both ethics committees.

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. Gridelli C, Rossi A, Carbone DP, et al. Non-small-cell lung cancer. Nat Rev Dis Primers 2015;1:15009. [Crossref] [PubMed]
2. Herbst RS, Morgensztern D, Boshoff C. The biology and management of non-small cell lung cancer. Nature 2018;553:446-54. [Crossref] [PubMed]
3. Doroshow DB, Sanmamed MF, Hastings K, et al. Immunotherapy in Non-Small Cell Lung Cancer: Facts and Hopes. Clin Cancer Res 2019;25:4592-602. [Crossref] [PubMed]
4. Jain P, Jain C, Velcheti V. Role of immune-checkpoint inhibitors in lung cancer. Ther Adv Respir Dis 2018;12:1753465817750075. [Crossref] [PubMed]
5. Hirano S, Fujita T, Takahashi H, et al. Rechallenge of immune checkpoint inhibitor after local therapy for immune checkpoint inhibitor-resistant non-small cell lung cancer. Chin Clin Oncol 2025;14:29. [Crossref] [PubMed]
6. Gettinger S, Choi J, Hastings K, et al. Impaired HLA Class I Antigen Processing and Presentation as a Mechanism of Acquired Resistance to Immune Checkpoint Inhibitors in Lung Cancer. Cancer Discov 2017;7:1420-35. [Crossref] [PubMed]
7. Brahmer JR, Lacchetti C, Schneider BJ, et al. Management of Immune-Related Adverse Events in Patients Treated With Immune Checkpoint Inhibitor Therapy: American Society of Clinical Oncology Clinical Practice Guideline. J Clin Oncol 2018;36:1714-68. [Crossref] [PubMed]
8. Sharma P, Goswami S, Raychaudhuri D, et al. Immune checkpoint therapy-current perspectives and future directions. Cell 2023;186:1652-69. [Crossref] [PubMed]
9. Lesterhuis WJ, Haanen JB, Punt CJ. Cancer immunotherapy--revisited. Nat Rev Drug Discov 2011;10:591-600. [Crossref] [PubMed]
10. Riely GJ, Wood DE, Ettinger DS, et al. Non-Small Cell Lung Cancer, Version 4.2024, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2024;22:249-74. [Crossref] [PubMed]
11. McLaughlin J, Han G, Schalper KA, et al. Quantitative Assessment of the Heterogeneity of PD-L1 Expression in Non-Small-Cell Lung Cancer. JAMA Oncol 2016;2:46-54. [Crossref] [PubMed]
12. Huang Z, Wu Y, Ye H, et al. Prediction of the efficacy and clinical prognosis of first-line EGFR-tyrosine kinase inhibitors in non-small cell lung cancer patients based on ΔCt values derived from the super-amplification refractory mutation system (ARMS): a real-world retrospective study. J Thorac Dis 2025;17:3897-911. [Crossref] [PubMed]
13. Zhu Y, Wu S, Ma F, et al. Predictive factors for the efficacy of immune checkpoint inhibitors in advanced non-small cell lung cancer: a retrospective study. J Thorac Dis 2025;17:4159-76. [Crossref] [PubMed]
14. Gu Y, Zheng H, Wang J, et al. The value of prognostic immune-inflammatory-nutritional score in predicting survival after surgery in stage I-III non-small cell lung cancer. J Thorac Dis 2025;17:3886-96. [Crossref] [PubMed]
15. Ding Q, Wang C, Zhang Z, et al. An explainable AI approach to surgical and radiotherapy interventions for optimized treatment decision-making in early-stage non-small cell lung cancer. Transl Lung Cancer Res 2025;14:2011-30. [Crossref] [PubMed]
16. Jain RK. Normalizing tumor microenvironment to treat cancer: bench to bedside to biomarkers. J Clin Oncol 2013;31:2205-18. [Crossref] [PubMed]
17. Martin JD, Seano G, Jain RK. Normalizing Function of Tumor Vessels: Progress, Opportunities, and Challenges. Annu Rev Physiol 2019;81:505-34. [Crossref] [PubMed]
18. Shigeta K, Datta M, Hato T, et al. Dual Programmed Death Receptor-1 and Vascular Endothelial Growth Factor Receptor-2 Blockade Promotes Vascular Normalization and Enhances Antitumor Immune Responses in Hepatocellular Carcinoma. Hepatology 2020;71:1247-61. [Crossref] [PubMed]
19. Tian L, Goldstein A, Wang H, et al. Mutual regulation of tumour vessel normalization and immunostimulatory reprogramming. Nature 2017;544:250-4. [Crossref] [PubMed]
20. Zheng X, Fang Z, Liu X, et al. Increased vessel perfusion predicts the efficacy of immune checkpoint blockade. J Clin Invest 2018;128:2104-15. [Crossref] [PubMed]
21. Tolaney SM, Boucher Y, Duda DG, et al. Role of vascular density and normalization in response to neoadjuvant bevacizumab and chemotherapy in breast cancer patients. Proc Natl Acad Sci U S A 2015;112:14325-30. [Crossref] [PubMed]
22. Napel S, Mu W, Jardim-Perassi BV, et al. Quantitative imaging of cancer in the postgenomic era: Radio(geno)mics, deep learning, and habitats. Cancer 2018;124:4633-49. [Crossref] [PubMed]
23. Wu J, Gensheimer MF, Dong X, et al. Robust Intratumor Partitioning to Identify High-Risk Subregions in Lung Cancer: A Pilot Study. Int J Radiat Oncol Biol Phys 2016;95:1504-12. [Crossref] [PubMed]
24. Zou Y, Hu X, Zheng S, et al. Discordance of immunotherapy response predictive biomarkers between primary lesions and paired metastases in tumours: A systematic review and meta-analysis. EBioMedicine 2021;63:103137. [Crossref] [PubMed]
25. Shi Z, Huang X, Cheng Z, et al. MRI-based Quantification of Intratumoral Heterogeneity for Predicting Treatment Response to Neoadjuvant Chemotherapy in Breast Cancer. Radiology 2023;308:e222830. [Crossref] [PubMed]
26. Chen MM, Terzic A, Becker AS, et al. Artificial intelligence in oncologic imaging. Eur J Radiol Open 2022;9:100441. [Crossref] [PubMed]
27. Liu Z, Ma K, Jia Q, et al. Baseline tumour vessel perfusion as a non-invasive predictive biomarker for immune checkpoint therapy in non-small-cell lung cancer. BMJ Oncol 2024;3:e000473. [Crossref] [PubMed]
28. Liu Z, Yao Y, Zhao M, et al. Radiomics Models Derived From Arterial-Phase-Enhanced CT Reliably Predict Both PD-L1 Expression and Immunotherapy Prognosis in Non-small Cell Lung Cancer: A Retrospective, Multicenter Cohort Study. Acad Radiol 2025;32:493-505. [Crossref] [PubMed]
29. Paver EC, Cooper WA, Colebatch AJ, et al. Programmed death ligand-1 (PD-L1) as a predictive marker for immunotherapy in solid tumours: a guide to immunohistochemistry implementation and interpretation. Pathology 2021;53:141-56. [Crossref] [PubMed]
30. Yang Y, Wang Z, Fang J, et al. Efficacy and Safety of Sintilimab Plus Pemetrexed and Platinum as First-Line Treatment for Locally Advanced or Metastatic Nonsquamous NSCLC: a Randomized, Double-Blind, Phase 3 Study (Oncology pRogram by InnovENT anti-PD-1-11). J Thorac Oncol 2020;15:1636-46. [Crossref] [PubMed]
31. Fedorov A, Beichel R, Kalpathy-Cramer J, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging 2012;30:1323-41. [Crossref] [PubMed]
32. Torheim T, Groendahl AR, Andersen EK, et al. Cluster analysis of dynamic contrast enhanced MRI reveals tumor subregions related to locoregional relapse for cervical cancer patients. Acta Oncol 2016;55:1294-8. [Crossref] [PubMed]
33. Xie P, Huang Q, Zheng L, et al. Sub-region based histogram analysis of amide proton transfer-weighted MRI for predicting tumor budding grade in rectal adenocarcinoma: a prospective study. Eur Radiol 2025;35:1382-93. [Crossref] [PubMed]
34. Li HJ, Qiu ZB, Wang MM, et al. Radiomics-Based Support Vector Machine Distinguishes Molecular Events Driving the Progression of Lung Adenocarcinoma. J Thorac Oncol 2025;20:52-64. [Crossref] [PubMed]
35. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
36. Zwanenburg A, Vallières M, Abdalah MA, et al. The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology 2020;295:328-38. [Crossref] [PubMed]
37. Pan Z, Men K, Liang B, et al. A subregion-based prediction model for local-regional recurrence risk in head and neck squamous cell carcinoma. Radiother Oncol 2023;184:109684. [Crossref] [PubMed]
38. Coroller TP, Grossmann P, Hou Y, et al. CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. Radiother Oncol 2015;114:345-50. [Crossref] [PubMed]
39. Farjam R, Tsien CI, Feng FY, et al. Physiological imaging-defined, response-driven subvolumes of a tumor. Int J Radiat Oncol Biol Phys 2013;85:1383-90. [Crossref] [PubMed]
40. Liao CY, Chen YM, Wu YT, et al. Personalized prediction of immunotherapy response in lung cancer patients using advanced radiomics and deep learning. Cancer Imaging 2024;24:129. [Crossref] [PubMed]
41. Ye G, Wu G, Zhang C, et al. CT-based quantification of intratumoral heterogeneity for predicting pathologic complete response to neoadjuvant immunochemotherapy in non-small cell lung cancer. Front Immunol 2024;15:1414954. [Crossref] [PubMed]
42. Yolchuyeva S, Ebrahimpour L, Tonneau M, et al. Multi-institutional prognostic modeling of survival outcomes in NSCLC patients treated with first-line immunotherapy using radiomics. J Transl Med 2024;22:42. [Crossref] [PubMed]