Habitat-based CT radiomics profiling spatial-temporal heterogeneity in resectable NSCLC predict pathological response to neoadjuvant chemoimmunotherapy: a multi-center study

Abstract

The profound spatial and temporal heterogeneity of non-small cell lung cancer (NSCLC) drives unpredictable responses to neoadjuvant chemoimmunotherapy (NCI), highlighting the need for effective predictive biomarkers to optimize treatment. In this multicenter study, we evaluated the ability of habitat imaging to predict major pathological response (MPR) to NCI by capturing spatial-temporal tumor heterogeneity, using pre- and post-treatment CT scans from 394 patients with resectable non-small cell lung cancer across three institutions. A radiomics-based predictive framework integrating global texture descriptors, spatial heterogeneity features, and longitudinal imaging information was constructed to distinguish pathological responders from non-responders. Models based on global texture or spatial heterogeneity features alone achieved areas under the receiver operating characteristic curve (AUCs) ranging from 0.71 to 0.80 across validation cohorts, whereas the integrated model further improved discrimination, achieving an AUC of up to 0.85 in external validation. These findings demonstrate that habitat imaging provides a robust approach for predicting MPR and supporting patient stratification and personalized treatment planning in NSCLC.

Introduction

Neoadjuvant chemoimmunotherapy (NCI) has emerged as a transformative approach in the treatment of resectable non-small cell lung cancer (NSCLC), significantly improving clinical outcomes. Recent studies report pathological complete response rates of up to 36.4% and major pathological response (MPR) rates of approximately 86.0%^1,2, compared to 1.9-8.9% with chemotherapy alone^3,4. These advancements have reshaped the treatment landscape, providing a promising pathway for improving long-term survival. However, substantial heterogeneity in pathological response remains a major obstacle, and reliable biomarkers for patient stratification are lacking. Given that the variability in treatment sensitivity is deeply rooted in the internal complexity of the tumor, decoding this heterogeneity is critical for precision management.

In the context of this study, intratumor spatial heterogeneity (ITSH) is defined not merely as the presence of genetic diversity, but as the non-uniform spatial distribution of distinct radiomic phenotypes (habitats) within the tumor volume, as captured by quantitative imaging features rather than molecular assays^5,6. This formulation reflects the architectural complexity of the tumor ecosystem, in which spatially distinct regions exhibit characteristic imaging phenotypes corresponding to different underlying pathophysiological states and microenvironmental conditions^7,8.

From a methodological perspective, quantifying tumor heterogeneity requires distinguishing between two complementary approaches: global texture and spatial topology. Whole-tumor heterogeneity (WTH) models the tumor as a single statistical entity, summarizing intensity dispersion across the entire volume and capturing the overall extent of variation^5,9. By contrast, intratumor heterogeneity (ITH) in the present study focuses on how biologically distinct habitats are spatially organized and interact at their interfaces, thereby encoding the macro-architectural structure of the tumor^10,11. This distinction is critical, as different spatial configurations may share identical global statistics while reflecting fundamentally different tumor architectures⁸.

Previous radiomics models have demonstrated decent predictive performance by quantifying statistical textural heterogeneity across the whole tumor volume^12,13. However, when WTH is quantified using global radiomics features, the tumor is effectively modeled as a single statistical entity. While informative, such global measures do not encode spatial topology and may therefore overlook architectural features of the tumor that are critical to biological behavior^14,15. To address this limitation, habitat imaging has emerged as a high-resolution approach that explicitly characterizes the spatial distribution of tumor heterogeneity by partitioning tumors into biologically distinct habitats¹⁶. This segmentation allows for a more nuanced analysis of how different tumor areas interact with one another and how they may influence treatment response¹⁷.

The aim of this multi-center study is to develop and validate a habitat-based spatial-temporal CT radiomics model for predicting MPR in patients with resectable NSCLC undergoing NCI. Through segmenting tumors into biologically distinct habitats, the model captures ITSH and integrates these subregional features with whole-tumor radiomic characteristics. By creating a robust and clinically applicable framework, we seek to improve patient stratification, aid therapeutic decision-making, and enhance personalized treatment planning. Validation across multiple independent cohorts ensures the model’s generalizability and reliability for clinical implementation.

Results

Patient demographics and clinical characteristics

Table 1 summarizes the clinical characteristics of enrolled patients in the training cohort and three validation cohorts. A total of 394 patients with resectable NSCLC were included, with a median age of 63 years (range: 42-78). Significant differences were observed between the MPR and non-MPR groups in age, smoking history, and tumor size (p < 0.05). The MPR group had a smaller average tumor size (mean diameter: 3.2 cm) compared to the non-MPR group (4.6 cm). Male patients accounted for 58% of the population, while 42% were female. Smoking history was more prevalent in the non-MPR group (65 vs. 48% in the MPR group, p < 0.01).

Table 1.

Clinical characteristics of the patients in training and two validation cohorts

Characteristic	Training cohort		P Value	Validation cohort 1		P Value	Validation cohort 2		P Value	Validation cohort 3		P Value
	Non-MPR (N = 100)	MPR (N = 112)		Non-MPR (N = 26)	MPR (N = 28)		Non-MPR (N = 30)	MPR (N = 56)		Non-MPR (N = 18)	MPR (N = 24)
Sex			0.03			0.35			0.25			1.00
Male	86 (86.00%)	106 (94.64%)		22 (84.62%)	26 (92.86%)		26 (86.67%)	43 (76.79%)		15 (83.33%)	20 (83.33%)
Female	14 (14.00%)	6 (5.36%)		4 (15.38%)	2 (7.14%)		4 (13.33%)	13 (23.21%)		3 (16.67%)	4 (16.67%)
Age (years)	60.77 ± 10.35	62.73 ± 7.94	0.12	63.23 ± 6.84	61.43 ± 6.41	0.33	64.83 ± 8.26	60.95 ± 7.56	0.04	63.00 ± 9.10	63.79 ± 6.87	0.77
Smoking history			0.39			0.89			0.91			0.36
Yes	64 (64.00%)	78 (69.64%)		19 (73.08%)	20 (71.43%)		13 (43.33%)	25 (44.64%)		7 (38.89%)	6 (25.00%)
No	36 (36.00%)	34 (30.36%)		7 (26.92%)	8 (28.57%)		17 (56.67%)	31 (55.36%)		11 (61.11%)	18 (75.00%)
Pathology subtype			0.51			0.96			0.35			0.50
ADC	38 (38.00%)	18 (16.07%)		8 (30.77%)	7 (25.00%)		7 (23.33%)	23 (41.07%)		6 (33.33%)	9 (37.50%)
SCC	49 (49.00%)	71 (63.39%)		17 (65.38%)	19 (67.86%)		22 (73.33%)	33 (58.93%)		10 (55.56%)	11 (45.83%)
Other	13 (13.00%)	23 (20.54%)		1 (3.85%)	2 (7.14%)		1 (3.33%)	0 (0.00%)		2 (11.11%)	4 (16.67%)
TNM stage			0.46			0.31			0.38			0.71
IB	1 (1.00%)	1 (0.89%)		1 (3.85%)	0 (0.00%)		0 (0.00%)	1 (1.79%)		0 (0.00%)	0 (0.00%)
IIA	1 (1.00%)	1 (0.89%)		0 (0.00%)	1 (3.57%)		0 (0.00%)	0 (0.00%)		0 (0.00%)	0 (0.00%)
IIB	10 (10.00%)	12 (10.71%)		1 (3.85%)	2 (7.14%)		0 (0.00%)	1 (1.79%)		4 (22.22%)	5 (20.83%)
IIIA	63 (63.00%)	63 (56.25%)		19 (73.08%)	14 (50.00%)		11 (36.67%)	21 (37.50%)		12 (66.67%)	15 (62.50%)
IIIB	23 (23.00%)	31 (27.68%)		5 (19.23%)	11 (39.29%)		19 (63.33%)	33 (58.93%)		2 (11.11%)	4 (16.67%)
IIIC	2 (2.00%)	4 (3.57%)		0 (0.00%)	0 (0.00%)		0 (0.00%)	0 (0.00%)		0 (0.00%)	0 (0.00%)
Location			0.04			0.95			0.56			0.86
Central	29 (29.00%)	48 (42.86%)		10 (38.46%)	11 (39.29%)		7 (23.33%)	10 (17.86%)		7 (38.89%)	10 (41.67%)
Peripheral	71 (71.00%)	64 (57.14%)		16 (61.54%)	17 (60.71%)		23 (76.67%)	46 (82.14%)		11 (61.11%)	14 (58.33%)

Habitat imaging and feature selection

The specific radiomic features constituting the final predictive signatures, along with their corresponding feature importance rankings, are explicitly detailed in Supplementary Table 2. As visualized in the feature heatmap (Supplementary Fig. 1), the selection process revealed distinct patterns for each modeling approach. The ITH model prioritized local statistical properties: 5 baseline features (e.g., entropy, interquartile range), 3 C1 features (e.g., 10th percentile, uniformity), and 4 delta features capturing intensity dynamics (e.g., mean absolute deviation, energy). In contrast, the WTH model integrated multimodal global features: 6 baseline features (e.g., 90th percentile, wavelet-based gray-level non-uniformity), 4 C1 features combining shape (elongation) and texture (IDN), and 5 delta features sensitive to energy and texture contrast. These distinctions reflect ITH’s focus on localized statistical patterns versus WTH’s holistic analysis of spatial and textural heterogeneity, aligning with their biological interpretations in tumor characterization.

Model performance

Supplementary Figs. 2 and 3 illustrate the receiver operating characteristic (ROC) curves and the corresponding areas under the receiver operating characteristic curve (AUCs) for the ITH and WTH models. The ITH model achieved an AUC of 0.88 (95% CI: 0.82–0.92) in the training cohort, 0.80 (95% CI: 0.66–0.93) in the validation cohort 1, 0.71 (95% CI: 0.57–0.82) in validation cohort 2, and 0.78 (95% CI: 0.59–0.91) in validation cohort 3. Sensitivity and specificity ranged from 68.7 to 84.3% and 75.6 to 85.1%, respectively. The WTH model demonstrated a higher AUC in the training cohort (0.91, 95% CI: 0.87–0.95) but performed similarly to the habitat model in the validation cohort 1 (AUC = 0.80, 95% CI: 0.66–0.91). In external validation cohorts, the WTH model achieved AUCs of 0.74 (95% CI: 0.60–0.86) and 0.73 (95% CI: 0.57–0.88).

Figure 1 compares the ROC curves of the ITH, WTH, and WITH models by testing on training and three validation cohorts. The WITH model outperformed both ITH and WTH models, achieving AUCs of 0.91 (95% CI: 0.86–0.94) in the training cohort, 0.85 (95% CI: 0.72–0.93) in the validation cohort 1, 0.75 (95% CI: 0.62–0.86) in validation cohort 2, and 0.83 (95% CI: 0.68–0.93) in validation cohort 3. Table 2 summarizes additional classification metrics, including accuracy, sensitivity, and specificity.

Fig. 1. Comparison of model performance across cohorts.

ROC curves comparing the performance of the ITH (orange), WTH (blue), and WITH (red) models across datasets. a Training cohort. b Validation cohort 1. c Validation cohort 2. d Validation cohort 3. The AUC is displayed in each panel. ROC receiver operating characteristic, AUC area under the curve, ITH intratumoral heterogeneity, WTH whole-tumor heterogeneity, WITH integrated model combining ITH and WTH.

Table 2.

Performance evaluation of models in different cohorts

Dataset	Model		ACC	SEN	SPE	PPV	NPV
Training cohort	WTH	Fusion	0.84	0.804	0.88	0.882	0.8
		Delta	0.764	0.766	0.77	0.787	0.74
		BL	0.676	0.681	0.72	0.72	0.643
		C1	0.675	0.677	0.94	0.891	0.599
	ITH	Fusion	0.797	0.804	0.79	0.811	0.782
		Delta	0.769	0.795	0.74	0.774	0.763
		BL	0.66	0.571	0.76	0.727	0.613
		C1	0.642	0.393	0.92	0.846	0.575
	WITH		0.816	0.688	0.96	0.951	0.733
Validation Cohort 1	WTH	Fusion	0.778	0.786	0.769	0.786	0.769
		Delta	0.722	0.643	0.808	0.783	0.677
		BL	0.648	0.607	0.692	0.68	0.621
		C1	0.63	0.357	0.923	0.833	0.571
	ITH	Fusion	0.778	0.929	0.615	0.722	0.889
		Delta	0.759	0.964	0.538	0.692	0.837
		BL	0.648	0.821	0.462	0.622	0.778
		C1	0.722	0.964	0.462	0.659	0.804
	WITH		0.778	0.75	0.808	0.808	0.75
Validation Cohort 2	WTH	Fusion	0.674	0.643	0.733	0.818	0.524
		Delta	0.57	0.357	0.967	0.952	0.446
		BL	0.483	0.268	0.9	0.833	0.397
		C1	0.553	0.411	0.833	0.821	0.431
	ITH	Fusion	0.628	0.571	0.733	0.8	0.478
		Delta	0.64	0.607	0.7	0.791	0.488
		BL	0.628	0.804	0.3	0.682	0.45
		C1	0.5	0.304	0.867	0.81	0.4
	WITH		0.733	0.786	0.633	0.8	0.613
Validation Cohort 3	WTH	Fusion	0.667	0.5	0.889	0.857	0.571
		Delta	0.667	0.667	0.667	0.727	0.6
		BL	0.619	0.708	0.5	0.654	0.562
		C1	0.571	0.583	0.556	0.636	0.5
	ITH	Fusion	0.69	0.5	0.944	0.923	0.586
		Delta	0.619	0.5	0.778	0.75	0.538
		BL	0.595	0.5	0.722	0.706	0.52
		C1	0.524	0.5	0.556	0.6	0.455
	WITH		0.762	0.708	0.833	0.85	0.682

Bold values indicate the best-performing result within each cohort. ITH intratumoral heterogeneity, WTH whole-tumor heterogeneity, WITH integrated model combining ITH and WTH, ACC accuracy, SEN sensitivity, SPE specificity, PPV positive predictive value, NPV negative predictive value.

Meanwhile, decision curve analysis (DCA) curves (Fig. 2) confirmed the clinical utility of the integrated model, demonstrating a higher net benefit across a range of threshold probabilities compared to both the habitat and whole tumor models. These findings highlight the integrated model’s robustness and clinical relevance in predicting therapeutic responses.

Fig. 2. Decision curve analysis across cohorts.

DCA comparing the net benefit of different prediction models across a range of threshold probabilities for MPR. Curves represent the WITH model (green), the WTH model (blue), and the ITH model (magenta), together with reference strategies assuming treatment of all patients (treat-all, red) or no patients (treat-none, olive). a Training cohort. b Validation cohort 1. c Validation cohort 2. d Validation cohort 3. The horizontal axis represents the threshold probability, and the vertical axis represents the net benefit. DCA decision curve analysis, MPR major pathological response.

We further evaluated systematic performance variations across clinical parameters (Supplementary Figs. 4–9). The model demonstrated systematic performance variations across clinical parameters: older patients (≥60 years) exhibited marginally higher predictive accuracy compared to younger cohorts, while non-smokers consistently outperformed smokers, likely reflecting reduced confounding factors in the former group. Squamous cell carcinoma (SCC) showed superior discriminative power (AUC 0.84–0.91 vs. 0.73–0.82 for non-SCC), with central tumor location (AUC 0.78–0.92 vs. 0.71–0.89 for peripheral) and advanced T/N staging (T3-4: 0.85–0.93 vs. T1-2: 0.72–0.84; N2-3: 0.81–0.94 vs. N0-1: 0.72–0.88) demonstrating particularly robust performance, suggesting enhanced sensitivity to anatomically distinct lesions and relatively late-stage morphological patterns. Figure 3 Examples of the ITSH index in two NSCLC patients who have similar clinical characteristics.

Fig. 3. Representative examples of intratumoral spatial heterogeneity.

Representative examples from two NSCLC patients with comparable clinical characteristics are shown to illustrate differences in intratumoral spatial heterogeneity captured by the ITSH index. Case 1 with a lower ITSH index (left), showing a relatively coherent spatial organization of imaging habitats and fewer fragmented subregions. Case 2 with a higher ITSH index (right), demonstrating increased spatial complexity with more dispersed and fragmented habitat distribution. NSCLC non-small cell lung cancer, ITSH intratumoral spatial heterogeneity, ADC adenocarcinoma, SCC squamous cell carcinoma, TNM tumor-node-metastasis.

Discussion

Despite growing recognition that response to NCI varies widely among patients with resectable NSCLC, the structural basis of this variability remains incompletely understood. In this study, we demonstrate that heterogeneity relevant to therapeutic response is not fully captured by global radiomic summaries alone. By jointly modeling intratumoral spatial topology and whole-tumor texture, we show that these two dimensions encode distinct and complementary information, rather than redundant representations of tumor complexity. Across three independent validation cohorts, integrating subregional spatial interactions with whole-tumor statistics in the WITH model consistently improved discrimination of MPR. Collectively, these findings suggest that treatment sensitivity is shaped not only by the magnitude of intratumoral variation but also by how that variation is spatially organized within the tumor.

By explicitly modeling spatial organization, our framework captures imaging patterns that are not readily apparent on routine visual inspection, consistent with prior radiomics studies emphasizing spatial structure and organization^18,19. High-entropy texture features, such as those derived from GLSZM, are consistent with increased spatial disorder within the tumor. In parallel, pronounced temporal changes in intensity have been linked to treatment-induced structural remodeling in longitudinal imaging analyses^20,21. The biological relevance of spatial heterogeneity has also been supported across imaging modalities. One illustrative example is provided by Sujit et al., who demonstrated associations between spatial organization on PET/CT and interferon-related signaling²², while related evidence supporting the biological significance of spatial heterogeneity has been reported using MRI-based approaches²³. Together, these observations reinforce spatial topology as a biologically meaningful imaging construct. Although molecular or multi-modal validation is beyond the scope of the present study, these findings provide contextual support for interpreting CT-based spatial topology as a complementary morphological correlate of therapeutic sensitivity. Accordingly, habitat-based representations are best viewed as interpretable imaging surrogates that capture biologically relevant spatial organization, rather than as direct mappings to specific molecular pathways²⁴.

To better understand the boundaries of this approach, we qualitatively reviewed misclassified cases and examined their imaging characteristics in detail. Notably, errors were not randomly distributed but tended to cluster around specific radiographic contexts. Lesions accompanied by prominent peritumoral inflammation or dense fibrotic change were more likely to be misclassified, a pattern consistent with immunotherapy-related pseudoprogression or post-treatment tissue repair mimicking residual tumor burden. Conversely, tumors dominated by ground-glass components also posed challenges, as their relatively homogeneous attenuation may obscure the spatial heterogeneity that habitat-based features are designed to capture. Rather than representing isolated failures, these observations highlight concrete imaging scenarios in which radiomics-based models may be limited and underscore the importance of situating spatial features within a broader biological and clinical context.

From a practical perspective, the WITH model translates these probability scores into actionable neoadjuvant management strategies. For predicted responders, the data support continuing the full therapeutic cycle to maximize pathological regression. Conversely, identifying resistance in predicted non-responders should prompt multidisciplinary consideration for early surgical resection. This early-exit strategy prevents disease progression during futile cycles. Crucially, it also avoids the development of dense hilar fibrosis. This specific complication of prolonged immunotherapy significantly increases surgical complexity and risk, making timely intervention essential for patient safety.

Our study also has some limitations. Firstly, interpretation of the present findings should be considered in light of several design-related constraints. Although rigorous image standardization was applied, the retrospective nature and sample size warrant validation in larger prospective cohorts to establish generalizability. Secondly, the framework prioritizes spatial topology through Multiregional Spatial Interaction (MSI)-based characterization rather than absolute volumetric measures. While habitat proportion represents a clinically intuitive descriptor, volumetric information alone cannot distinguish biologically distinct spatial configurations, such as centralized versus dispersed necrotic foci. Future studies may determine whether simplified volume-based metrics can approximate higher-dimensional spatial descriptors without substantial loss of biological information. Thirdly, this emphasis on spatial and temporal characterization necessitates dual time-point imaging, which may be unavailable in a subset of patients due to incomplete follow-up or deviations from planned therapy. Finally, reliance on imaging data alone precludes direct interrogation of the tumor microenvironment. Stromal and immune components were not explicitly assessed, and the absence of spatially matched histopathological validation remains an important limitation. Integrating imaging with spatial pathomics will be essential to more directly link the identified habitat phenotypes to their biological underpinnings.

Our multi-center study establishes habitat imaging as a robust surrogate for decoding the spatial-temporal heterogeneity of resectable NSCLC. By integrating sub-regional topology with whole-tumor statistics, the proposed model effectively predicts therapeutic sensitivity across independent cohorts. This framework moves beyond abstract imaging features to reveal the biological architecture of response, translating complex spatial patterns into actionable clinical insights. Ultimately, this approach provides a precise, non-invasive instrument for patient stratification, optimizing the timing and efficacy of NCI.

Methods

Study population

This multi-center, retrospective observational study (NCT04766515) investigated the association between ITSH and treatment response in resectable NSCLC patients receiving NCI across three hospitals. Patients were consecutively collected from January 2019 to December 2023 at three centers: Center 1 (Shanghai Pulmonary Hospital, Tongji University), Center 2 (Fudan University Shanghai Cancer Center), and Center 3 (Zhongshan Hospital Affiliated to Fudan University). Patients who underwent chest CT before radical surgery were identified through electronic medical records. We included patients who met the following criteria¹: histologically confirmed resectable NSCLC², received NCI as neoadjuvant therapy³, with available pre- and post-treatment contrast-enhanced CT scans for radiomic analysis, and⁴ underwent surgical resection post-treatment. Patients with distant metastases, a history of other malignancies, or incomplete clinical and imaging data were excluded. A total of 266 patients from center 1 were randomly split into a training set (n = 212, 80.0%) and a validation cohort 1 (n = 54, 20.0%). Center 1 patients were partitioned into training and internal validation cohorts via stratified random sampling based on MPR status. This protocol preserved outcome distributions across datasets. Statistical verification confirmed no significant differences in baseline characteristics between groups. Patients from center 2 (n = 86) and center 3 (n = 42) were used to develop two independent validation cohorts, namely, validation cohort 2 and validation cohort 3, respectively. Figure 4 shows the flowchart of patient enrollment.

Fig. 4. Patient enrollment and cohort allocation.

Research flowchart illustrating the screening, exclusion, and allocation of patients with resectable NSCLC into the training cohort and independent validation cohorts across participating centers. NSCLC non-small cell lung cancer, MPR major pathological response.

Clinical data were collected for each patient, including demographics (age, sex, smoking history), tumor characteristics (size, histological subtype, location), and treatment regimens. Patients were classified into MPR and non-MPR groups based on histopathological data reflecting their pathological response to therapy. Patients underwent a standard NCI regimen consisting of platinum-based chemotherapy in combination with a PD-1/PD-L1 inhibitor (e.g., pembrolizumab or nivolumab). The chemotherapy regimen typically included cisplatin or carboplatin combined with paclitaxel or pemetrexed. Immunotherapy was administered every 3 weeks for up to four cycles. Surgical resection was performed within 4–6 weeks following the completion of neoadjuvant therapy. The clinical and pathological response to NCI was evaluated based on the American Joint Committee on Cancer (AJCC) staging system and the RECIST 1.1 criteria.

This multi-center study was conducted in accordance with the Declaration of Helsinki and was approved by the Clinical Research Ethics Committee of Fudan University Shanghai Cancer Center (Approval No. 090977-1), the Institutional Review Board of Shanghai Pulmonary Hospital (Approval No. L20-335-2), and the Institutional Review Board of Fudan University Zhongshan Hospital (Approval No. B2021-128). Written informed consent was obtained from all participants prior to study enrollment.

Image acquisition and tumor segmentation

Pre- and post-treatment chest CT images were obtained using standardized protocols across three centers. These protocols included consistent slice thickness, contrast administration, and imaging parameters to ensure uniformity in image quality. A board-certified radiologist manually delineated primary lesions on each slice using ITK-SNAP software (http://www.itksnap.org), ensuring accurate tumor segmentation. The entire tumor volume of interest (VOI) was included in the segmentation process to ensure comprehensive coverage of all tumor regions. CT images were resampled to a voxel size of 1 mm³, balancing spatial resolution and computational efficiency for consistent feature extraction across all datasets. Figure 5 illustrates the workflow of our proposed habitat imaging model, detailing key steps from image acquisition to feature extraction.

Fig. 5. Workflow of the habitat imaging framework.

Schematic workflow illustrating the habitat imaging pipeline for modeling spatial-temporal tumor heterogeneity from longitudinal chest CT images. Pre-treatment and post-treatment CT images (top) are acquired and manually segmented to delineate the tumor region. The segmented tumor is partitioned into multiple superpixel-based subregions using the SLIC algorithm, which are subsequently clustered to define imaging habitats with distinct radiomic phenotypes (middle). Radiomic features are extracted at both the whole-tumor level and the habitat level, including baseline, post-treatment, and delta features capturing longitudinal changes. Spatial interactions among habitats are quantified to characterize intratumoral heterogeneity, followed by feature selection, data balancing, and model construction to generate predictive outputs (bottom). CT computed tomography, NSCLC non-small cell lung cancer, NCI neoadjuvant chemoimmunotherapy, SLIC Simple Linear Iterative Clustering, MSI Multiregional Spatial Interaction, ITH intratumoral heterogeneity, WTH whole-tumor heterogeneity, LASSO-RFE least absolute shrinkage and selection operator with recursive feature elimination, SMOTE synthetic minority over-sampling technique.

Habitat subregion segmentation and intratumor heterogeneity feature extraction

A two-step subregion segmentation method was applied to generate the intratumor habitat subregions. Intratumoral partitioning was performed using the Simple Linear Iterative Clustering (SLIC) algorithm via the scikit-image library, with the compactness parameter set to 10 and the Gaussian smoothing parameter (sigma) set to 0 to ensure boundary adherence and reproducibility. The tumor was segmented into 100 subregions, a granularity selected to balance spatial resolution with computational efficiency based on prior studies. Extraction parameters were standardized across all cases to maintain methodological consistency.

A total of 91 IBSI-compliant radiomic features were extracted from each intratumoral subregion to constitute the initial feature pool. This set included first-order intensity statistics and textural features derived from five matrices: Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix, Gray Level Size Zone Matrix (GLSZM), Gray Level Dependence Matrix, and Neighboring Gray Tone Difference Matrix. Shape features were intentionally excluded, as the boundaries of superpixels are algorithmically defined rather than anatomically structured. Feature selection was subsequently performed within the modeling pipeline to identify the most informative subsets for prediction.

Feature extraction was performed using PyRadiomics. To ensure reproducibility, gray-level discretization was set to a fixed bin width of 25 to normalize intensity resolution. Images were resampled to a voxel size of 1 × 1 × 1 mm³ using B-spline interpolation, and no distance weighting was applied for texture matrices. Since all images were contrast-enhanced CT scans in Hounsfield Units (HU), gray-level normalization (Scale) was not performed. The full extraction configuration is provided in the open-source code repository. Subsequently, unsupervised K-means clustering was applied to group these subregions into distinct phenotypic habitats based on their radiomic profiles. The optimal number of clusters (k) was determined via Silhouette analysis performed strictly on the training dataset to prevent information leakage. We evaluated k values ranging from 2 to 10, identifying k = 5 as the optimal configuration yielding the maximum Silhouette coefficient. Crucially, this optimization relied solely on internal clustering validity metrics without reference to downstream clinical outcomes, preserving the unsupervised nature of habitat construction.

To quantify ITSH as a measure of spatial topology rather than local texture, we constructed an MSI matrix based on the multi-subregion maps. Unlike conventional texture analysis, this approach explicitly interrogates the macro-architectural organization of the tumor. Specifically, for each tumor voxel, neighboring voxels were identified, and corresponding entries were recorded in the MSI matrix. This process was repeated until all tumor voxels were analyzed. To account for interactions with surrounding tissue, the lung parenchyma was treated as a distinct region. In this matrix, diagonal elements represent the connected size of individual subregions, while off-diagonal elements encode the spatial interactions at the interfaces between biologically distinct habitats. Based on the MSI matrix, a total of 24 features were extracted to quantify the ITSH, including 18 first-order and six second-order statistical features (detailed in Supplementary Table 1). Meanwhile, an ITH score was also calculated to quantify the complexity of multi-subregional maps. Together, an ITH feature pool comprising 25 features was calculated by using multi-subregional maps. The ITH score was defined as follows:

$$IT{H}_{Volume}=1-\frac{1}{V_{total}}\sum _{i=1}^N\frac{S_{i,\max }}{n_i}$$ 1

where N denotes the number of clusters in the multi-subregional maps, $n_i$ denotes the number of connected subregions, $S_{i,\max }$ denotes the largest area of connected subregion for cluster i, and $V_{\mathrm{total}}$ denotes the whole tumor area. Accordingly, ITSH features encode macro-architectural topology of the tumor ecosystem and are conceptually distinct from conventional texture features extracted within localized regions.

Whole tumor and delta radiomic feature extraction

To quantify the WTH, the tumor was modeled as a single statistical entity. A total of 1106 radiomics features were computed based on the entire VOI to capture global texture and statistical dispersion. These features included shape descriptors (e.g., compactness, sphericity), first-order statistics (e.g., mean intensity, entropy), and texture features (e.g., GLCM, GLSZM). A WTH feature pool was developed to decode the heterogeneity of VOI.

Both the ITH and WTH feature pool were built on pre- and post-treatment chest CT images first. Afterward, the delta radiomics features were computed to monitor the changes in radiomics features. The delta radiomics feature was defined as follows.

$$Delt{a}_{Feature}=\frac{Featur{e}_{C1}-Featur{e}_{BL}}{Featur{e}_{BL}}$$ 2

where ${\mathrm{Feature}}_{\mathrm{BL}}$ and ${\mathrm{Feature}}_{\mathrm{C}1}$ denote the pre-treatment and post-treatment radiomics features, respectively.

Feature selection and prediction model development

All radiomics features were standardized by using the Z-score normalization method. This process involves subtracting the mean of each feature and dividing by its standard deviation (SD) to ensure all features are on a similar scale. A recursive feature elimination (RFE) feature selection algorithm configured with the least absolute shrinkage and selection operator was used to remove the redundant radiomics features and select the optimal radiomics features. Due to the imbalance of the training dataset, a synthetic minority over-sampling technique (SMOTE) method was used to generate new minority instances to balance the dataset. Finally, an eXtreme Gradient Boosting classifier was employed to train and build a prediction model to predict between MPR and non-MPR groups. As a result, nine radiomics models were built to predict the response of NCI. Four radiomics models involved pre-treatment (BL) model, post-treatment (C1) model, delta model, and all feature fusion model were developed based on WTH and ITH feature pool, respectively. In order to improve the model performance, a WITH model was developed by using a decision-level fusion method with combining ITH and WTH prediction scores.

Performance evaluation and statistical analysis

Model performance was assessed by using several key metrics. The ROC and AUC were calculated to evaluate the model’s discrimination ability, providing a measure of its effectiveness in distinguishing between MPR and non-MPR groups. Additionally, classification performance was quantified using accuracy, sensitivity, and specificity, offering insights into the model’s overall predictive reliability and its ability to correctly identify true positive and true negative cases. DCA was employed to examine the clinical utility of the models by evaluating the net benefits across a range of decision thresholds, ensuring their applicability in guiding therapeutic decisions. This comprehensive evaluation framework ensured robust validation of model performance and clinical relevance.

Statistical analyses were performed to compare the clinical characteristics between the MPR and non-MPR groups. Continuous variables were compared using the Student’s t-test or Mann–Whitney U test, depending on the normality of the data distribution. Categorical variables were analyzed using the Chi-square test or Fisher’s exact test, as appropriate. All statistical tests were two-sided, and a p value < 0.05 was considered statistically significant. Continuous variables were presented as mean ± SD, while categorical variables were summarized as counts and percentages. The radiomics model development and statistical analyses were performed using Python and R software. Several publicly available Python packages, including PyRadiomics, Scikit-learn, SimpleITK, and SciPy, were applied.

Author contributions

Q.P. curated the data, prepared the figures, and drafted the initial manuscript. Y.X. contributed to data processing, materials acquisition, and figure preparation. L.S. provided data support and contributed to model validation. X.B. contributed to model validation and project coordination. S.Z. contributed to the study design and supervised the research. X.Y. contributed to study conception, model implementation, and funding acquisition. Y.G. acquired funding, coordinated resources, and critically reviewed the manuscript. J.G. contributed to study design, data analysis, funding acquisition, and critically reviewed the manuscript. All authors commented on previous versions of the manuscript and approved the final manuscript.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to ethical restrictions related to the protection of participants’ privacy, but are available from the corresponding author on reasonable request.

Code availability

All custom code used for habitat segmentation, MSI matrix computation, ITH score calculation, and model development is available at https://github.com/CancerImageAI/WITH-LungCancer. The repository provides documented scripts and instructions to support reproducibility and reuse.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01388-z.