Development and validation of a radiomics-habitat model for preoperatively predicting poorly differentiated stage IA lung adenocarcinoma Authors: Gang Xiang, Zi Yang, Xiang Wang, Jie Gao, Bin Yang, Qin-Ling Jiang, Yu Ding, Wen-Man Li, Guang-Yan Si Published: 2026 ABSTRACT BackgroundStage IA poorly differentiated lung adenocarcinoma has stronger invasiveness and shorter survival time. Its treatment methods are related to the survival time, and preoperative prediction can assist clinical practice in formulating personalized treatment options. The aim of this study was to develop and validate a computed tomography (CT)-based radiomics-habitat fusion model for preoperative prediction of stage IA poorly differentiated lung adenocarcinoma.MethodsA retrospective dual-center cohort of 507 patients with surgically confirmed stage IA lung adenocarcinoma was enrolled. The 419 patients from The Affiliated Traditional Chinese Medicine Hospital, Southwest Medical University (Center 1) were randomly assigned to a training cohort (n=294) and an internal test cohort (n=125), whereas 88 independent patients from Second Affiliated Hospital of Naval Medical University (Center 2) formed the external validation cohort. Tumor pathological differentiation graded by the 2021 World Health Organization (WHO) classification of thoracic tumors was set as the reference standard. Voxel-wise k-means clustering divided intratumoral regions into three distinct habitats (H1–3). Four machine learning algorithms [logistic regression, random forest, ExtraTrees, eXtreme Gradient Boosting (XGBoost)] were separately used to construct whole-tumor radiomics models, habitat subregional models and the final combined model. Model discrimination was comprehensively evaluated via area under the curve (AUC), sensitivity and specificity, etc.; calibration and clinical utility were assessed by calibration curves and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature contributions of the combined model.ResultsPoorly differentiated lesions accounted for 26.2% (133/507) of all participants. In the independent external validation cohort, the combined model demonstrated robust discriminatory performance with an AUC of 0.880 [95% confidence interval (CI): 0.794–0.966], significantly outperforming single models (P<0.05). At the optimized decision threshold, the combined model achieved a sensitivity of 0.84 and a specificity of 0.81. DCA confirmed that the combined model provided substantial net clinical benefit.ConclusionsThe developed and externally validated CT-based radiomics-habitat combined model serves as a promising non-invasive tool for preoperatively identifying poorly differentiated stage IA lung adenocarcinoma. Nevertheless, our external validation cohort only included 88 patients from a single independent center, which limits the generalizability of the model; further multi-center validation is warranted prior to clinical use. Highlight box Key findings • A combined computed tomography (CT)-based radiomics-habitat model was constructed, and the area under the curve (AUC) for predicting stage IA poorly differentiated lung adenocarcinoma in the external validation cohort was 0.880. What is known and what is new? • Traditional radiomics regards tumors as a single entity, potentially overlooking the microenvironmental heterogeneity associated with grading. Habitat imaging has not been applied to the prediction of the differentiation degree of stage IA lung adenocarcinoma. • This multicenter study combined habitat analysis with whole-tumor radiomics for the first time. The results showed that the combined model was significantly superior to the single model (DeLong P<0.05). What is the implication, and what should change now? • This non-invasive model is expected to assist in the preoperative risk stratification and surgical decision-making of stage IA lung adenocarcinoma, but prospective validation is still required. Introduction Lung adenocarcinoma is the most prevalent histologic subtype of non-small cell lung cancer, with its incidence rising globally. According to the 2021 World Health Organization (WHO) classification of lung tumors, adenocarcinomas characterized by more than 20% of micropapillary, solid, or occasionally colloid, enteric, or fetal type components are classified as poorly differentiated adenocarcinomas (1). High-grade micropapillary and solid histological components are closely linked to aggressive tumor phenotypes including spread through air spaces (STAS) and lymphovascular invasion, which substantially elevate early postoperative recurrence risk (2). Studies confirm that patients with poorly differentiated lung adenocarcinoma experience significantly shorter progression-free survival (3). The 5-year survival rate for patients undergoing lobectomy surpasses that of those undergoing segmentectomy (P<0.01) (4). Given that sublobar resection results in unsatisfactory oncologic control for lesions harboring these adverse pathological features, non-invasive preoperative identification of poorly differentiated stage IA adenocarcinoma bears critical clinical significance (5). Radiomics has developed as a significant method for analyzing intratumoral heterogeneity by measuring and extracting high-dimensional features from standard computed tomography (CT) images. The clinical usefulness of Radiomics has been shown in a number of studies that predict invasive components and lymphovascular invasion, spread through air spaces, and high-grade pattern tumors in lung adenocarcinoma (6,7). On the other hand, conventional radiomics approaches the tumor as a single, uniform entity and likely does not capture the distinct microenvironmental heterogeneities (8). Recently developed Habitat Imaging helps overcome the limitations of conventional radiomics by utilizing an unsupervised clustering technique which subsections the tumors into varying biological ‘habitats’ in accordance with the differing local microenvironmental conditions (9,10). Built on multiparametric imaging metrics reflecting cellularity, vascularity, necrosis and hypoxia, unsupervised habitat clustering groups voxels with matching functional signals to isolate biologically distinct tumor subregions missed by uniform whole-tumor radiomics (11). The identified subregions enhance the predictive capabilities regarding treatment responses, metastatic potential, and prognostic assessment across various cancer types. Thus, habitat-driven radiomics provides a more detailed characterization of tumor ecology and spatial heterogeneity, resulting in a more accurate biological understanding, as well as enhanced predictive power (12,13). However, its use in differentiating lung adenocarcinoma, especially in early-stage disease, has yet to be fully investigated, and no previous research has combined habitat imaging with standard radiomics to forecast stage IA tumors with poor differentiation. Accordingly, we developed and validated a CT-based radiomics-habitat combined model for the preoperative prediction of poorly differentiated stage IA lung adenocarcinoma. We hypothesized that combining whole-tumor radiomics with habitat-derived subregional descriptors would improve discrimination compared with either approach alone while retaining model interpretability. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1254/rc). Methods Patients This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective multi-center study was approved by the Ethics Committee of The Affiliated Traditional Chinese Medicine Hospital, Southwest Medical University (No. BY2025078), and the requirement for informed consent was waived due to the retrospective nature of the study. Second Affiliated Hospital of Naval Medical University was also informed of and agreed to the study. A retrospective study was conducted on 507 patients from two centers, each with surgically verified primary lung adenocarcinoma: 419 patients at The Affiliated Traditional Chinese Medicine Hospital, Southwest Medical University (Center 1; April 2021–March 2025), and 88 at Second Affiliated Hospital of Naval Medical University (Center 2; March 2023–May 2024). Patients at Center 1 were divided into a training set (n=294) and an internal testing set (n=125) at random, whereas patients from Center 2 made up the external validation cohort (n=88) entirely. The inclusion criteria are as follows: (I) the lesion was pathologically diagnosed as lung adenocarcinoma after resection; (II) the clinical stage was IA; (III) the preoperative CT images were thin-slice (slice thickness <1.5 cm). The exclusion criteria are: (I) the time interval between CT and surgery was >15 days; (II) neoadjuvant chemotherapy or radiotherapy was applied; (III) a history of other malignancies; (IV) CT artifacts are so extensive that they preclude the extraction of radiomic features. The patient selection flowchart is shown in Figure 1. Flowchart for collecting lung cancer cases. Center 1: Southwest Medical University Affiliated Hospital of Traditional Chinese Medicine; Center 2: Second Affiliated Hospital of Naval Medical University. Each specimen was examined independently by two pathologists with 5 and 10 years of experience, respectively. Per the 2021 WHO Classification of Thoracic Tumors, tumors were graded as high, moderately, or poorly differentiated adenocarcinomas. Discrepancies were settled by consensus. For analysis, well- and moderately differentiated tumors were combined into one group and compared to poorly differentiated tumors. Image preprocessing and segmentation Preoperative thin-slice chest CT data were retrieved from the PACS in DICOM format, converted to NIfTI format, and standardized using lung-window settings. To reduce variability across scanners and acquisition protocols, voxel intensities were clipped to −800 to 800 Hounsfield units (HU) and all images were resampled to an isotropic voxel spacing of 1×1×1 mm3 before radiomic and habitat-feature extraction. Two senior thoracic radiologists with 13 and 15 years of experience independently delineated the tumor regions of interest (ROIs) in ITK-SNAP. Discrepancies were resolved by joint review to obtain a consensus volume of interest (VOI), which was checked in the sagittal and coronal planes. All segmentations were subsequently reviewed and calibrated by a third senior radiologist with 15 years of experience to ensure boundary consistency before feature extraction. Radiomics feature extraction Whole-tumor radiomic features were extracted using the open-source PyRadiomics package (version 3.1.0). A total of 1,834 features were obtained, including 14 shape features, 360 first-order features, 440 gray-level co-occurrence matrix (GLCM) features, 280 gray-level dependence matrix (GLDM) features, 320 gray-level run-length matrix (GLRLM) features, 320 gray-level size zone matrix (GLSZM) features, and 100 neighboring gray-tone difference matrix (NGTDM) features. The same preprocessing settings were applied before whole-tumor and habitat-level feature extraction. Habitat construction and feature extraction Intratumoral habitats were partitioned using a voxel-wise approach. First, 19 local features were computed within a 5×5×5 sliding window to characterize the per-voxel imaging microenvironment (Table S1). These voxel-level vectors were subsequently clustered using the unsupervised k-means algorithm. The optimal cluster number (k=3)—defining habitats H1, H2, and H3—was determined via the Calinski-Harabasz, Silhouette, and Davies-Bouldin indices (Appendix 1). Following partitioning, habitat labels were projected back to the original image space for subregional radiomic feature extraction using PyRadiomics. To construct the final habitat feature matrix, these subregional descriptors were integrated at the patient level. For habitats with limited volumes, missing values were handled via K-nearest neighbors (KNN) imputation to ensure numerical stability during model training (Appendix 1). Feature selection All features were standardized via z-score normalization. All feature-selection procedures were performed exclusively in the training cohort. Features associated with tumor differentiation were first screened using two-sample t-tests (P<0.05). To reduce redundancy, for any pair of strongly positively correlated features (Pearson r>0.9), one feature was randomly retained. Subsequently, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was performed to select an optimal regularization parameter (λ) and finalize the feature set. Construction of radiomics, habitat and combined models Four algorithms were evaluated for model construction: ExtraTrees, eXtreme Gradient Boosting (XGBoost), random forest, and logistic regression. A radiomics model was trained using whole-tumor features, whereas habitat-specific models were first derived from H1, H2, and H3 and then combined to form the overall habitat model. Because some habitat regions occupied very small volumes in individual tumors, a subset of habitat-level features could not be stably computed in all cases; these missing values were handled by K-nearest-neighbors imputation during model development before classifier training. The predicted probabilities generated by the whole-tumor radiomics model and the overall habitat model were defined as the Rad-score and Habitat-score, respectively. A logistic regression model was then fitted using the Rad-score and Habitat-score as inputs to construct the combined model and generate the final predicted probability. Pathological evaluation Histopathologic slides were independently evaluated by two senior thoracic pathologists (10 and 5 years of diagnostic experience, respectively), who were completely blinded to all preoperative CT images, clinical data, and machine learning model predictions. Any discrepancies were resolved through joint consensus review. Histological grading strictly followed the 2021 World Health Organization (WHO) thoracic tumor classification and the 2020 International Association for the Study of Lung Cancer (IASLC) grading systems for lung adenocarcinoma. Poorly differentiated (high-grade) tumors were explicitly defined as those containing more than 20% high-grade architectural patterns, including micropapillary, solid, cribriform, and complex glandular subtypes. Statistical analysis When comparing the training set, internal test set, and external validation set, a normality test was performed on continuous variables. If the data followed a normal distribution, they were presented as mean ± standard deviation, and one-way analysis of variance was used. If the data did not follow a normal distribution, they were presented as median and interquartile range, and the Kruskal-Wallis test was used. Categorical variables were presented as frequencies and percentages, and the chi-square test was used. To alleviate the problem of sample size imbalance, we employed the Synthetic Minority Oversampling Technique (SMOTE) method to resample the training dataset. This method generates new synthetic samples by interpolating between samples of the minority class, thereby increasing the number of cases of poorly differentiated lung adenocarcinoma and achieving a balanced ratio between the two classes of samples. Model performance was measured by area under the curve (AUC), accuracy, sensitivity, specificity, and F1 score. The DeLong test was performed for AUC comparisons. The Hosmer-Lemeshow test and calibration curve assessed calibration. Decision curve analysis (DCA) estimates the clinical relevance. To analyze the explanation models, we used SHapley Additive exPlanations (SHAP) and determined the contribution of each chosen attribute to the ultimate prediction. All statistical analysis were conducted in Python (version 3.9) and R (version 4.1.0). All statistical hypothesis tests in this study were two-sided, and a P value <0.05 was considered statistically significant. The process is illustrated in Figure 2. The entire technical process. 2D, two dimensional; 3D, three dimensional; ROI, region of interest; VOI, volume of interest. Results Patient characteristics This study incorporated 507 participants: 374 (73.8%) were in the moderately/high differentiated category and 133 (26.2%) were in the poorly differentiated category. Table 1 summarizes the baseline characteristics for both groups. No significant between-cohort differences were observed in age, sex, differentiation status, tumor diameter, or lesion location (P>0.05). Detailed baseline characteristics are provided in Table 1. Table 1The baseline clinical and lesion characteristicsCharacteristicsTotal (n=507)Train set (n=294)Test set (n=125)Validation set (n=88)P valueAge (years)61 (54.00–68.00)61 (55.00–68.00)60 (53.00–68.00)61 (54.00–69.75)0.93Gender0.56 Male224 (44.18)124 (42.18)59 (47.20)41 (46.59) Female283 (55.82)170 (57.82)66 (52.80)47 (53.41)Differentiate0.23 High/moderately differentiated374 (73.77)225 (76.53)86 (68.80)63 (71.59) Poorly differentiated133 (26.23)69 (23.47)39 (31.20)25 (28.41)Tumor diameter (cm)1.6 (1.00–2.20)1.50 (1.00–2.20)2.00 (1.50–2.45)1.65 (1.20–2.10)0.06Tumor location0.43 Right upper169 (33.33)97 (32.99)44 (35.20)28 (31.81) Right middle38 (7.50)24 (8.16)7 (5.60)7 (7.95) Right lower102 (20.12)55 (18.71)30 (24.00)17 (19.32) Left upper116 (22.88)64 (21.77)25 (20.00)27 (30.68) Left lower82 (16.17)54 (18.37)19 (15.20)9 (10.23)Data are presented as median (interquartile range) or number (percentage). The baseline clinical and lesion characteristics Data are presented as median (interquartile range) or number (percentage). Feature selection and habitat visualization The feature reduction pipeline identified 13 optimal descriptors for the whole-tumor radiomics model and 21 selected features for the integrated habitat model (detailed stepwise reduction counts are provided in Table S2). The characterization of the 19 local radiomic features utilized for habitat construction is illustrated in Figure 3. For subregional partitioning, internal validation metrics—including the Calinski-Harabasz, Silhouette Coefficient, and Davies-Bouldin indices—confirmed that a three-cluster configuration was the optimal solution (Figure 4). Visual analysis of the resulting habitat maps demonstrated that Habitat 3 (H3), characterized by high-intensity and high-density voxel distributions, was more prevalent in poorly differentiated lesions. The volumetric proportions and spatial distributions of these habitats (H1–3) are visualized in Figure 5, with representative CT images of subregional distributions in six patients presented in Figure 6. Nineteen local features were used for habitat construction. Features 00–02 represent first-order features, 03–11 GLCM features, 12–14 GLRLM features, 15–16 GLSZM features, and 17–18 NGTDM features. The final panel shows the voxel-wise habitat map after clustering. GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size zone matrix; NGTDM, neighboring gray tone difference matrix. Clustering evaluation data graph. When the number of clusters is 3, it represents the optimal clustering. CH, Calinski-Harabasz; DB, Davies-Bouldin; SC, Silhouette Coefficient. Distribution within a three-dimensional feature space after dimensionality reduction. Distribution of tumor sub-regions in CT images of 6 lung cancer patients. Different colors represent different sub-regions, and sub-region 2 is missing in the 6th case. CT, computed tomography. Model performance and multicenter validation The radiomics-habitat combined model consistently demonstrated superior discriminative power across all datasets. The comparative performance of the standalone machine learning classifiers (ExtraTrees, XGBoost, random forest, and logistic regression) for the radiomics component is detailed in Table 2. In the independent external validation cohort, the combined model achieved an AUC of 0.880 [95% confidence interval (CI): 0.794–0.966]. This performance was significantly higher than that of the standalone habitat model (AUC: 0.819, 95% CI: 0.738–0.900; DeLong P<0.05) and showed a statistically significant incremental improvement over the whole-tumor radiomics model (AUC: 0.841, 95% CI: 0.765–0.917). Comprehensive performance metrics for all models, including accuracy, sensitivity, and specificity, are provided in Table 3. The ROC curves, calibration plots, and DCA for the training, internal test, and external validation sets are collectively displayed in Figure 7. Table 2Detailed information on model construction using ExtraTrees, LR, RF, and XGBoostSignatureAccuracyAUC95% CISensitivitySpecificityPPVNPVPrecisionRecallF1Train set ExtraTrees0.7520.8530.8069–0.89840.8550.720.4840.9420.4840.8550.618 LR0.7210.8150.7585–0.87240.7970.6980.4470.9180.4470.7970.573 RF0.8470.8680.8209–0.91520.6670.9020.6760.8980.6760.6670.672 XGBoost0.8810.9440.9174–0.97080.8550.8890.7020.9520.7020.8550.771Test set ExtraTrees0.7120.8210.7469–0.89410.9490.6050.5210.9630.5210.9490.673 LR0.80.8090.7282–0.89020.7690.8140.6520.8860.6520.7690.706 RF0.7920.8120.7279–0.89520.7950.7910.6330.8950.6330.7950.705 XGBoost0.720.790.7085–0.87060.7690.6980.5360.870.5360.7690.632Validation set ExtraTrees0.7610.8410.7375–0.94380.920.6980.5480.9570.5480.920.687 LR0.7950.8370.7356–0.93930.920.7460.590.9590.590.920.719 RF0.7730.7550.6383–0.87220.680.810.5860.8640.5860.680.63 XGBoost0.7730.8260.7226–0.92880.840.7460.5680.9220.5680.840.677AUC, area under the curve; CI, confidence interval; LR, logistic regression; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; XGBoost, eXtreme Gradient Boosting. Detailed information on model construction using ExtraTrees, LR, RF, and XGBoost AUC, area under the curve; CI, confidence interval; LR, logistic regression; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; XGBoost, eXtreme Gradient Boosting. Table 3Performance of radiomics models and habitat modelsSignatureAccuracyAUC95% CISensitivitySpecificityPPVNPVPrecisionRecallF1Train set Rad0.7520.8530.8069–0.89840.8550.720.4840.9420.4840.8550.618 Habitat0.8440.9080.8720–0.94420.8410.8440.6240.9450.6240.8410.716 HabitatH10.8230.8870.8461–0.92750.8260.8220.5880.9390.5880.8260.687 HabitatH20.7550.7760.7137–0.83850.710.7690.4850.8960.4850.710.576 HabitatH30.8330.8710.8237–0.91800.7830.8490.6140.9270.6140.7830.688 Combined0.8330.9280.8970–0.95890.8990.8130.5960.9630.5960.8990.717Test set Rad0.7120.8210.7482–0.89460.9490.6050.5210.9630.5210.9490.673 Habitat0.7360.8030.7285–0.87790.9490.6470.5440.9650.5440.9490.692 HabitatH10.7280.7740.6918–0.85560.8970.6510.5380.9330.5380.8970.673 HabitatH20.6320.6980.6043–0.79080.7690.570.4480.8450.4480.7690.566 HabitatH30.7440.8020.7272–0.87770.9490.6590.5520.9660.5520.9490.698 Combined0.8080.9070.8579–0.95670.8720.7790.6420.9310.6420.8720.739Validation set Rad0.7610.8410.7375–0.94380.920.6980.5480.9570.5480.920.687 Habitat0.7840.8190.7156–0.92250.840.7740.5830.9230.5830.840.689 HabitatH10.8180.8140.7125–0.91540.720.8710.6670.8850.6670.720.692 HabitatH20.6250.6840.5672–0.80050.920.5160.4260.9410.4260.920.582 HabitatH30.8070.7980.6865–0.90970.800.810.6250.9110.6250.800.702 Combined0.8180.8800.7936–0.96580.840.810.6360.9270.6360.840.724AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Performance of radiomics models and habitat models AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Model’s performance of the training set, internal test set, and external validation set. (A) Receiver operating characteristic curve, (B) DeLong test; (C) decision curve analysis; (D) the calibration curve. AUC, area under the curve; CI, confidence interval. SHAP analysis To enhance the interpretability of the radiomics-habitat combined model, SHAP analysis was employed to quantify the contribution of individual features to the final prediction. Figure 8A illustrates the most influential predictors, notably wavelet_LLL_glcm_DifferenceVariance, gradient_firstorder_10 Percentile_h1, and wavelet_LHL_glcm_ClusterTendency. These features collectively reflect complex aspects of intratumoral heterogeneity. Specifically, higher values of these descriptors (represented by red points) were associated with a positive SHAP value, indicating an increased probability of poor differentiation. Conversely, lower values—corresponding to reduced textural and density heterogeneity—decreased the likelihood of a poorly differentiated diagnosis. SHAP analysis of the combined model. (A) SHAP summary plot illustrates the feature relevance and attributions to the model’s predictive performance. (B) SHAP dependence plot explains how a certain sample feature predicts poorly differentiated adenocarcinoma. SHAP, SHapley Additive exPlanations. The SHAP force plot (Figure 8B) further visualizes the decision-making process for an individual patient. In this representative case, the feature log_sigma_3__mm_3D_glcm_DifferenceVariance exerted the most substantial positive influence on the model. Consequently, the predicted probability reached 0.464, significantly exceeding the cohort baseline (0.235) and correctly classifying the lesion as poorly differentiated. Discussion The rapid development of artificial intelligence, radiomics, and habitat-based imaging analysis has been widely applied to describe tumor heterogeneity and assist clinical decision-making. Previous studies have demonstrated their effectiveness in predicting the benign or malignant nature of lesions, invasiveness, lymph node metastasis, EGFR mutations, and patient prognosis (14-19). In the past, various research teams have employed artificial intelligence to forecast high-grade tumor components and poorly differentiated tumors. Xing et al. proposed an early hybrid radiomics–clinical model, reporting an AUC of 0.843 (20). Du et al. obtained an AUC of 0.849 with the addition of whole-tumor habitat and some semantic imaging sign (21). Using high-resolution CT-based radiomics, Ninomiya et al. were able to successfully predict the micropapillary and solid components (22). Sun et al. developed a model integrating radiomics with clinical variables to identify poorly differentiated adenocarcinoma (23). To our knowledge, the application of habitat analysis to predict poorly differentiated lung adenocarcinoma stage IA remains unexamined. Even though both conventional radiomics and habitat analysis entail the extraction of information within a tumor, the former involves a different level of methodology and biological detail. While conventional radiomics looks at the whole tumor, potentially losing informative signal value due to the presence of hemorrhage, necrosis, cystic components, or edema, habitat analysis, through the use of unsupervised clustering, divides tumors into subregions with biological significance. This enables the description of microenvironmental heterogeneity, which can be of greater predictive value (24,25). The term ‘habitat’ has evolved from its original definition as a description of the environment an organism lives in to now include the term’s micro-ecological environments in tissues. In oncology, tumor habitats describe the cellular, vascular, stromal, and immune components that influence the evolution, therapeutic response, and cancer progression of a tumor (26,27). These environments exhibit disordered neovascularization, irregular percombined, hypoxia, and metabolic competition, all of which create selective pressures that promote intratumoral heterogeneity (28,29). Extensive research has associated this heterogeneity with more aggressive tumor features, worse prognosis, lymphovascular invasion, and resistance to therapy (16,30). In this study, we utilized the Calinski-Harabasz, Silhouette Coefficient, and Davies-Bouldin indices to evaluate the optimal clustering among 3 to 10 clusters. All three indices consistently identified the 3-cluster configuration as the optimal choice, possibly because this clustering can best capture the microenvironmental differences in stage IA lung adenocarcinoma. Interestingly, in our comparison, the overall habitat model was better than each individual subregion model in the external validation, which indicates that the overall spatial pattern of heterogeneity could be more useful than the localized microregional signals. Nevertheless, the subregions still hold some value. Future work could be directed toward exploiting hierarchical or attention-based combined methods to improve the integration of features across multiple scales. One unanticipated result was a relatively weak predictive performance for Habitat H2, despite being located centrally within the tumor. This is, in part, due to its very small volume proportion (9.8%), which may limit its contribution to a stable feature representation. Further, the central necrosis or low per-combined regions may result in signal degradation and loss of predictive value. The results emphasize the need to focus on volumetric and biological reasoning when interpreting the predictive performance of habitat-derived subregions. To address the “black box” problem in machine learning, we utilized SHAP for model interpretation. In previous lung cancer studies, SHAP has been used to identify key features affecting factors like airspace dissemination, survival prediction, and immunotherapy response (31-34). In our research, the top three features of the combined models include wavet_lll_glcm_differencevariance and wavet_lhl_glcm_clustertendency, which are linked to the internal texture and heterogeneity of tumors. Meanwhile, gradient_firstorder_10Percentile_h1 relates to the tumor margin. These characteristic meanings are similar to the signs used in clinical practice for judging the benign or malignant nature of pulmonary nodules (35). Even though its training cohort achieved an AUC of 0.923, the external validation of the Habitat model revealed poor performance, which was lower than that of the traditional radiomics model. Despite the lack of a statistically significant difference between the two models (P<0.05), the study demonstrated that the Habitat model’s performance did not exceed that of the radiomics model. Nonetheless, prior studies have noted that the strength of combining different but complementary modalities is often greater than the sum of its parts (36). In this regard, our combined model showed a significant improvement over both the radiomics-only and habitat-only models across all cohorts (P<0.05). This enhancement is attributed to the integrated macro-scale radiomics (whole-tumor structural heterogeneity) and micro-scale habitat (local ecological diversity) synergism (10,36-38). The integration of macroscopic radiomics and microscopic habitat imaging facilitates a cross-scale biological characterization of the tumor, thereby providing more comprehensive data on tumor behavior (12,17,37). In the setting of stage IA lung adenocarcinoma, the clinical consequence of a false-negative prediction is severe: mistakenly misclassifying an aggressive, poorly differentiated tumor as low-risk could lead to an inadequate sublobar resection, leaving behind occult residual disease and increasing the risk of recurrence. Conversely, a false-positive result leads to a standard lobectomy, which carries more tissue loss but remains oncologically safe. Consequently, our model is explicitly positioned as a preoperative clinical triage adjunct tool designed to augment radiological judgment rather than replace definitive post-operative pathology. We adjusted the model’s operational decision threshold to maximize sensitivity (0.84) to suppress the false-negative rate. According to our DCA, when a clinician’s risk tolerance threshold for missing high-grade components lies within a realistic range of 10% to 40%, managing patients based on the model’s guidance (i.e., performing standard lobectomy with nodal clearing if the model flags a high-grade signature) yields a significantly higher net clinical benefit compared to a blanket strategy of “all lobectomy” or “all sublobar resection”. Numerous constraints deserve attention. While qualitative CT indicators like lobulation and spiculation are clinically relevant for tumor differentiation assessment (38), their inclusion into this framework was not possible due to subjectivity. The integration of radiomics with objective morphological descriptors or deep learning-derived features would be beneficial for enhancing predictive accuracy. Second, the number of poorly differentiated cases is relatively limited even with two participating centers, and additional large-scale multi-center validation cohorts are needed to fully verify model robustness. Lastly, the prediction with the greatest improvement, and perhaps the most insightful in the context of the biological understanding of the imaging-derived habitats, will be obtained with the inclusion of multi-omics data, such as genomic, proteomic, or spatial transcriptomic data (39). Conclusions In summary, we created a thorough and explorable model for preoperative prediction of poorly differentiated stage IA lung adenocarcinoma by integrating spatial habitat analysis and conventional radiomics. This combined method captures global and microenvironmental levels of tumor diversity and demonstrates superior outcomes than each method individually. The proposed model provides a valuable, non-invasive additional layer to surgical strategies and improvements in risk stratification for early lung adenocarcinoma. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1254/prf