Nodule subtype dependency of CT-based radiomics for preoperative prediction of spread through air spaces in stage I lung adenocarcinoma: a multi-centre study Authors: Meirong Li, Yachao Ruan, Yunfei Sheng, Fangfang Qiu, Huiqi Guo, Yuxuan Qiu, Feng Chen Published: 2026 ABSTRACT BackgroundSpread through air spaces (STAS) is an important invasive pattern associated with aggressive behaviour and prognosis in lung adenocarcinoma. However, whether nodule subtype influences the performance of computed tomography (CT)-based radiomics for preoperative STAS prediction remains unclear. This study aimed to investigate the effects of nodule subtype on CT-based radiomics performance for STAS prediction in stage I lung adenocarcinoma.MethodsThis was a multi-centre retrospective study, with 1,377 patients in the internal cohort and 436 in the external validation cohorts. Two radiomics models were developed: an all-subtype model (M1) and a pure-solid-specific model (M2). Radiomic features were obtained from preoperative CT images, and model development was performed using Extreme Gradient Boosting. Predictive performance was assessed using the area under the curve (AUC), calibration analysis, and decision-curve analysis. In the external pure-solid cohort, the two models were directly compared using the DeLong test, net reclassification improvement (NRI), and integrated discrimination improvement (IDI).ResultsM1 yielded an AUC of 0.839 in the internal cohort and 0.865 in the overall external cohort. After external stratification by nodule subtype, the AUCs for non-solid and pure-solid nodules were 0.797 and 0.788, respectively. M2 yielded an AUC of 0.773 in the internal cohort and 0.768 in the external cohort. In the external pure-solid cohort, M1 exhibited superior discrimination (P=0.0018), while M2 yielded more favourable reclassification performance (NRI =0.643; IDI =0.184).ConclusionsCT-based radiomics may have value for preoperative STAS prediction; however, its predictive performance likely depends on nodule subtype, with pure-solid nodules representing the primary diagnostic challenge. Introduction Lung cancer remains one of the most frequently diagnosed malignant tumours worldwide and is associated with substantial mortality. Adenocarcinoma (ADC) is the predominant histological subtype, with surgical resection being the main treatment for peripheral stage I disease (1-3). Spread through air spaces (STAS), recognised as an invasive pattern in the 2015 World Health Organization classification (4), is associated with aggressive pathological features, including high-grade histological patterns, pleural and lymphovascular invasion, and lymph node metastasis (5,6). STAS-positive lung ADC is associated with poorer survival and increased recurrence after sublobar resection (7,8). Therefore, reliable non-invasive preoperative assessment of STAS may support surgical decision-making regarding resection extent, lymph node dissection, and surgical margin management. However, accurate preoperative identification of STAS-positive tumours remains challenging. Intraoperative frozen-section analysis is not always reliable for STAS detection (9). Further, although there are reports on the association between STAS and conventional computed tomography (CT) features, including consolidation-to-tumour ratio (CTR), lobulation, spiculation, pleural invasion, and vascular invasion (10-18), their interpretation is still affected by inter-reader variability. Radiomics provides a complementary technique for quantitatively characterising tumour heterogeneity on routine CT images and is increasingly being explored as a tool for clinical decision support and personalised risk stratification in early lung cancer (19). Prior research has proposed that radiomics models are valuable for detecting STAS-positive tumours preoperatively in ADC (20-30). However, their predictive performance has varied significantly across studies. This variability may be partly attributable to differences in cohort composition, particularly differences in nodule-subtype distribution across development and validation cohorts. STAS is more frequent in solid nodules than in part-solid and pure ground-glass nodules (7,10,11,17), whereas the attenuation patterns and textural characteristics of these lesion subtypes substantially differ. Compared with subsolid nodules, pure-solid nodules are generally associated with more invasive tumour behaviour and poorer prognosis, and may influence preoperative consideration of surgical extent. Since most published radiomics studies involved mixed-subtype cohorts, some of the reported predictive ability may have reflected recognition of subtype-related imaging differences rather than true STAS-specific imaging signatures. This concern is supported by previous studies that show better performance for models derived from solid components than for those based on ground-glass components (29). In contrast, prediction performance appears more stable in cohorts with near-pure micropapillary and solid pattern, which have a relatively higher risk of STAS (26). Therefore, nodule subtype may represent an important source of heterogeneity in radiomics performance for STAS prediction. Pure-solid nodules are particularly relevant because their higher prevalence of STAS and relatively uniform CT attenuation may reduce the discriminative capacity of hand-crafted radiomic features. Subtype-stratified validation and pure-solid-specific radiomics modelling are therefore needed to determine whether CT-based radiomics provides STAS-related predictive information beyond nodule subtype composition. In this study, we developed and validated two CT-based radiomics models for preoperative STAS prediction in early-stage ADC: an all-subtype model (M1) and a pure-solid-specific model (M2). Their performance was compared in the overall and subtype-stratified cohorts to determine how nodule subtype affects radiomics-based STAS prediction and whether pure-solid-specific modelling provides complementary information for surgical risk stratification. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1123/rc). Methods Study population This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of the First Affiliated Hospital, College of Medicine, Zhejiang University (No. IIT20260147B). All participating hospitals were informed of and agreed to the study. Individual consent for this retrospective analysis was waived. The internal cohort consisted of 1,377 patients with pathologically confirmed lung ADC treated at the First Affiliated Hospital of Zhejiang University School of Medicine between January 2023 and December 2023. The external validation cohort included 436 patients selected from Shengzhou People’s Hospital, Xiaoshan Hospital of Traditional Chinese Medicine, the First People’s Hospital of Xiaoshan District, Jinyun People’s Hospital, Fuyang Hospital of Traditional Chinese Medicine and Tiantai People’s Hospital between 2024 and 2025. For prespecified subgroup analyses, the external cohort was further divided into a non-solid subgroup (n=233) and a pure-solid subgroup (n=203). The inclusion criteria were as follows: (I) maximum tumour diameter ≤3 cm; (II) surgically confirmed primary invasive lung ADC; (III) solitary lesion; and (IV) availability of preoperative non-contrast thin-section chest CT performed within 30 days before surgery. The exclusion criteria were as follows: (I) previous or concurrent malignancy (n=244); (II) receipt of neoadjuvant chemotherapy or radiotherapy before surgery (n=58); (III) pathological diagnosis of specific ADC variants, including invasive mucinous or foetal ADC (n=52); (IV) marked cavitary or cystic change within the lesion (n=72); (V) multiple ipsilateral pulmonary lesions (n=613); and (VI) CT images with motion artefacts or insufficient quality for model development (n=75). Clinical and pathological data were obtained from medical records. Cases across all of the nodule subtypes were used to construct M1, whereas the pure-solid subgroup was used to develop M2. Figure 1 illustrates the study selection process and overall design, with further details provided in the supplementary material. Flowchart of patient selection for the study cohorts. Hospital 1 refers to the First Affiliated Hospital, College of Medicine, Zhejiang University, where the internal cohort was collected. Other hospitals refer to the external validation centres. NSCLC, non-small cell lung cancer; STAS, spread through air spaces. CT imaging acquisition All of the CT images were acquired using multidetector CT scanners across the participating centres. The standard scanning was reconstructed with a high-frequency algorithm using a slice thickness of 0.625–1.25 mm. The supplementary material provides detailed acquisition parameters for each participating centre (Table S1). Histologic and CT feature evaluation Two pathologists reviewed STAS using a multiheaded microscope, and discrepancies were resolved by consensus. Clinical variables included age, sex, tumour size, and nodule subtype. In accordance with Fleischner Society terminology (31), two radiologists classified lesions on lung-window images as pure ground-glass opacities (pGGO), part-solid nodules, or solid nodules. For subsequent analyses, pGGO and part-solid nodules were grouped as non-solid nodules. Tumour size was measured on lung-window images, and the mean diameter was recorded. Tumour segmentation and radiomic feature extraction Image annotation, feature extraction, feature selection, and model development were performed using Deepwise Multimodal Research Platform version 3.1.3 https://keyan.deepwise.comBeijing Deepwise & League of PHD Technology Co., Ltd, Beijing, China. A radiologist with 5 years’ experience in chest CT manually delineated the regions of interest (ROIs) on thin-section high-resolution computed tomography (HRCT) images. During segmentation, major bronchovascular structures and adjacent normal lung parenchyma were deliberately excluded. Subsequently, the final contours were reviewed and confirmed by consensus with a second radiologist with 17 years of experience. Intraclass correlation coefficients (ICCs) were used to assess the interobserver variability in tumour segmentation between radiologist 1 and radiologist 2. Features with reproducibility (ICC >0.75) were retained for further analysis. All CT images underwent preprocessing, including B-spline-based resampling to a voxel size of 1 mm × 1 mm × 1 mm and z-score normalisation followed by linear scaling to [1,64]. For each ROI, 2,185 radiomic features were extracted and standardised as Z scores. Important first-order, shape, and texture features, including grey-level co-occurrence matrix (GLCM), grey-level run-length matrix (GLRLM), grey-level size-zone matrix (GLSZM), neighbouring grey-tone difference matrix (NGTDM), and grey-level dependence matrix (GLDM) features, were obtained from these lesions. Feature selection and model development Feature selection began with Spearman correlation analysis with a correlation coefficient threshold of 0.9 in the internal cohort. Subsequently, feature selection was conducted using a tree-based algorithm, and least absolute shrinkage and selection operator regression were used to identify the most informative feature subset (Figure S1). After filtering, 431 radiomic features were retained for the M1 and 419 for the M2. Both models were developed using the extreme gradient boosting (XGBoost) algorithm. Model stability was assessed through five-fold cross-validation performed on the training set. The selected features and final model settings were fixed before external validation, and no model refitting was performed in the external cohort. M1 was trained across all nodule subtypes, while M2 was trained only within the pure-solid subgroup. Model evaluation M1 was evaluated in the internal cohort and three external validation cohorts: the overall external cohort and two subtype-stratified external cohorts defined by nodule subtype (non-solid, n=233; pure-solid, n=203). M2 was evaluated in the internal and external pure-solid cohorts. In both the internal and external validation cohorts, predictive performance was characterised by the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). For direct comparison between M1 and M2 within the same external pure-solid cohort, the DeLong test was employed. Paired classification correctness was compared using McNemar’s test, and net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated to evaluate individual-level risk reassignment. Calibration was assessed using calibration curves, while decision-curve analysis was used to assess clinical utility across different threshold probabilities. Feature contributions were estimated using SHapley Additive exPlanations (SHAP) analysis. The complete radiomics workflow is illustrated in Figure 2. Workflow for the development and evaluation of the CT-based radiomics models M1 and M2. CT, computed tomography; DCA, decision-curve analysis; GLCM, grey-level co-occurrence matrix; GLDM, grey-level dependence matrix; GLRLM, grey-level run-length matrix; GLSZM, grey-level size-zone matrix; LASSO, least absolute shrinkage and selection operator; M1, all-subtype model; M2, pure-solid-specific model; NGTDM, neighbouring grey-tone difference matrix; ROC, receiver operating characteristic; ROI, region of interest; XGBoost, extreme gradient boosting. Statistical analysis Statistical analyses were performed using R software (version: 3.4.1; R Foundation for Statistical Computing, Vienna, Austria; http://www.Rproject.org). Continuous variables are expressed as mean ± standard deviation or median (interquartile range), as appropriate, while categorical variables are reported as counts and percentages. Between-group comparisons for STAS-positive and STAS-negative cases were performed using the independent-samples t-test, Mann-Whitney U test, or chi-squared test, as appropriate. Subtype-adjusted logistic regression was performed to determine whether the Rad-score retained STAS-related information beyond nodule subtype. Multivariable models were adjusted for nodule subtype, tumour size, age, and sex, and the Rad-score was standardised per 1-standard deviation (SD) increase within each cohort. Subgroup, reclassification, and misclassification analyses were considered exploratory, and P values were not adjusted for multiple comparisons. Statistical significance was defined as a two-sided P value <0.05. Results Patient selection and baseline characteristics A total of 1,813 patients were included in the final study, comprising 1,377 patients in the internal cohort and 436 in the external validation cohort. The external cohort was further stratified into a non-solid subgroup (n=233) and a pure-solid subgroup (n=203). Overall, 578 patients (31.9%) were classified as STAS-positive. Women represented 61.5% of the study population, and the acinar-predominant subtype was the most frequent histological pattern (68.6%, 1,244/1,813). Table 1 displays the baseline clinicopathological profile of the study cohorts. Tables S2,S3 show comparisons between STAS-negative and STAS-positive patients in the M1 and M2 analyses. Table 1Baseline characteristics of the internal cohort, overall external validation cohort, and subtype-stratified external cohortsVariableInternal cohort (n=1,377)Internal pure-solid cohort (n=770)External validation cohort, all subtypes (n=436)External validation non-solid (n=233)External validation pure-solid (n=203)P value (train vs. all)P value (train vs. non-solid)P value (train vs. pure-solid)P value (train vs. validation) (pure-solid)Age (years)61.62±8.5161.50±9.0362.13±11.3660.13±11.6264.43±10.630.8350.5580.2740.344Tumor size, mean diameter (cm)1.31±0.491.43±0.502.92±4.702.27±3.753.68±5.50<0.001<0.001<0.001<0.001CTR0.76±0.351.00±0.000.66±0.400.36±0.321.00±0.00<0.001<0.001<0.001Sex0.9420.8100.6430.536 Male529 (38.4)332 (43.1)169 (38.8)87 (37.3)82 (40.4) Female848 (61.6)438 (56.9)267 (61.2)146 (62.7)121 (59.6)Size grade0.482<0.001<0.0010.048 1 (≤1 cm)193 (14.0)80 (10.4)68 (15.6)58 (24.9)10 (4.9) 2 (1–2 cm)732 (53.2)381 (49.5)237 (54.4)135 (57.9)102 (50.2) 3 (>2 and ≤3 cm)452 (32.8)309 (40.1)131 (30.0)40 (17.2)91 (44.8)Nodule type<0.001<0.001<0.001– GGO126 (9.2)069 (15.8)69 (29.6)0 (0.0) Part-solid481 (34.9)0164 (37.6)164 (70.4)0 (0.0) Solid770 (55.9)770 (100)203 (46.6)0 (0.0)203 (100.0) NA2 (0.1)1 (0.1)1 (0.2)1 (0.4)0 (0.0)Pathologic T stage<0.001<0.001<0.001<0.001 T1a213 (15.5)86 (11.2)92 (21.1)68 (29.2)24 (11.8) T1b654 (47.5)311 (40.4)208 (47.7)124 (53.2)84 (41.4) T1c336 (24.4)210 (27.3)84 (19.3)39 (16.7)45 (22.2) T2a162 (11.8)151 (19.6)49 (11.2)2 (0.9)47 (23.2) T312 (0.9)12 (1.6)3 (0.7)0 (0.0)3 (1.5)Histologic subtype (leading component)<0.001<0.0010.4250.004 Acinar929 (67.5)470 (61.0)315 (72.2)169 (72.5)146 (71.9) Lepidic74 (5.4)9 (1.2)48 (11.0)41 (17.6)7 (3.4) Papillary216 (15.7)137 (17.8)46 (10.6)20 (8.6)26 (12.8) Micropapillary25 (1.8)24 (3.1)6 (1.4)0 (0.0)6 (3.0) Solid63 (4.6)63 (8.2)10 (2.3)1 (0.4)9 (4.4) Complex glandular60 (4.4)59 (7.7)6 (1.4)0 (0.0)6 (3.0) Other10 (0.7)8 (1.0)5 (1.1)2 (0.9)3 (1.5)Differentiation (leading component)<0.001<0.001<0.001<0.001 Well20 (1.5)4 (0.5)99 (22.7)77 (33)22 (10.8) Moderate990 (71.9)441 (57.3)236 (54.1)151 (64.8)85 (41.9) Poor365 (26.5)324 (42.1)100 (22.9)4 (1.7)96 (47.3) NA2 (0.1)1 (0.1)1 (0.2)1 (0.4)0 (0.0)STAS0.068<0.001<0.0010.856 Negative922 (67.0)391 (50.8)313 (71.8)212 (91.0)101 (49.8) Positive455 (33.0)379 (49.2)123 (28.2)21 (9.0)102 (50.2)Pleural invasion0.201<0.001<0.0010.334 Absent1,192 (86.6)599 (77.9)375 (86.0)224 (96.1)151 (74.4) Present184 (13.4)170 (22.1)60 (13.8)8 (3.4)52 (25.6) NA0 (0.0)–1 (0.2)1 (0.4)0 (0.0)Lymph node metastasis<0.0010.018<0.0010.008 Absent1,210 (88.0)624 (81.1)350 (89.3)181 (92.8)169 (85.8) Present113 (8.2)109 (14.2)42 (10.7)14 (7.2)28 (14.2) NA52 (3.8)36 (4.7)0 (0.0)0 (0.0)0 (0.0)Vascular invasion<0.001<0.001<0.0010.123 Absent1,267 (92.1)664 (86.2)361 (82.8)192 (82.4)169 (83.3) Present99 (7.2)98 (12.7)75 (17.2)41 (17.6)34 (16.7) NA10 (0.7)8 (1.0)0 (0.0)0 (0.0)0 (0.0)Values are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. Non-solid nodules were defined as GGO and part-solid nodules, whereas pure-solid nodules were defined as solid nodules. P values were calculated for train vs. validation all, train vs. validation non-solid, and train vs. validation pure-solid. Welch t-test was used for continuous variables. Chi-square test was used for categorical variables; Fisher exact test was used for 2×2 tables when appropriate. Histologic subtype and differentiation were summarized by the leading component of the original pathology text. CTR, consolidation-to-tumour ratio; GGO, ground-glass nodule; NA, not available; STAS, spread through air spaces; T, tumor. Baseline characteristics of the internal cohort, overall external validation cohort, and subtype-stratified external cohorts Values are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. Non-solid nodules were defined as GGO and part-solid nodules, whereas pure-solid nodules were defined as solid nodules. P values were calculated for train vs. validation all, train vs. validation non-solid, and train vs. validation pure-solid. Welch t-test was used for continuous variables. Chi-square test was used for categorical variables; Fisher exact test was used for 2×2 tables when appropriate. Histologic subtype and differentiation were summarized by the leading component of the original pathology text. CTR, consolidation-to-tumour ratio; GGO, ground-glass nodule; NA, not available; STAS, spread through air spaces; T, tumor. Multivariable analysis and subtype-adjusted association between Rad-score and STAS In both development and external validation cohorts, multivariable logistic regression revealed that pure-solid nodule subtype remained independently associated with STAS. After adjustment for nodule subtype, tumour size, age, and sex, the Rad-score remained independently associated with STAS in both cohorts, with odds ratios (ORs) per 1-SD increase of 3.82 [95% confidence interval (CI), 2.91–5.00] in the internal cohort and 3.22 (95% CI, 2.33–4.46) in the external validation cohort, both P<0.001; Table 2. Table 2Multivariable logistic regression for predicting STAS in the internal and external validation cohortsVariableInternal cohortExternal cohortOR (95% CI)P valueOR (95% CI)P valueNodule subtype (ref: pGGN) PSN2.11 (0.99–4.49)0.05216.21 (0.96–274.29)0.054 Pure-solid9.97 (4.83–20.62)<0.00195.18 (5.72–1,583.02)0.001Size (cm), per 1-cm increase2.83 (2.14–3.74)<0.0013.21 (1.91–5.40)<0.001Age (years), per 1-year increase0.99 (0.98–1.01)0.2530.98 (0.96–1.00)0.092Sex (male vs. female)2.04 (1.58–2.64)<0.0012.13 (1.28–3.52)0.003Rad-score, per 1-SD increase3.82 (2.91–5.00)<0.0013.22 (2.33–4.46)<0.001CI, confidence interval; OR, odds ratio; pGGN, pure ground-glass nodule; PSN, part-solid nodule; SD, standard deviation; STAS, spread through air spaces. Multivariable logistic regression for predicting STAS in the internal and external validation cohorts CI, confidence interval; OR, odds ratio; pGGN, pure ground-glass nodule; PSN, part-solid nodule; SD, standard deviation; STAS, spread through air spaces. Overall predictive performance of M1 and M2 In the internal cohort, M1 achieved an AUC of 0.839 (95% CI: 0.817–0.861), while in the overall external cohort, the AUC was 0.865 (95% CI: 0.822–0.909). After the external validation cohort was stratified by nodule subtype, the AUC of M1 was 0.797 (95% CI: 0.678–0.917) in the non-solid subgroup, where STAS prevalence was 9.0%, and 0.788 (95% CI: 0.725–0.850) in the pure-solid subgroup, where STAS prevalence reached 50.2%. In the pure-solid subgroup, M2 showed an AUC of 0.773 (95% CI: 0.740–0.806) in the internal cohort and 0.768 (95% CI: 0.701–0.836) in the external validation cohort. Additional discrimination metrics are provided in Table 3, and ROC curves are shown in Figure 3A,3B. Table 3Predictive performance of M1 and M2 in the internal and external validation cohortsModelCohortSubsetAUC (95% CI)AccuracySensitivitySpecificityPPVNPVBrierM1Internal validationAll nodules0.839 (0.817–0.861)0.783 (0.760–0.800)0.614 (0.570–0.660)0.866 (0.840–0.890)0.692 (0.650–0.740)0.821 (0.800–0.840)0.15M1External validationAll nodules0.865 (0.822–0.909)0.826 (0.790–0.860)0.557 (0.470–0.650)0.930 (0.900–0.960)0.756 (0.670–0.840)0.844 (0.810–0.880)0.122M1External validationNon-solid nodules0.797 (0.678–0.917)0.918 (0.880–0.950)0.200 (0.20–0.380)0.986 (0.970–1)0.571 (0.200–0.940)0.929 (0.900–0.960)0.064M1External validationPure-solid nodules0.788 (0.725–0.850)0.719 (0.660–0.780)0.627 (0.530–0.720)0.812 (0.740–0.890)0.771 (0.680–0.860)0.683 (0.600–0.770)0.189M2Internal validationPure-solid nodules0.773 (0.740–0.806)0.701 (0.670–0.730)0.696 (0.650–0.740)0.706 (0.660–0.750)0.697 (0.650–0.740)0.704 (0.660–0.750)0.197M2External validationPure-solid nodules0.768 (0.701–0.836)0.700 (0.640–0.760)0.578 (0.480–0.670)0.822 (0.750–0.900)0.766 (0.670–0.860)0.659 (0.580–0.740)0.208AUC, area under the curve; CI, confidence interval; M1, all-subtype model; M2, pure-solid-specific model; NPV, negative predictive value; PPV, positive predictive value. Predictive performance of M1 and M2 in the internal and external validation cohorts AUC, area under the curve; CI, confidence interval; M1, all-subtype model; M2, pure-solid-specific model; NPV, negative predictive value; PPV, positive predictive value. Overall predictive performance of M1 and M2 in the internal and external validation cohort, and subtype-stratified external subsets. (A) ROC curves. (B) Bar chart of AUC values. (C) Calibration plots. (D) Binned calibration curves. (E) Decision-curve analysis. (F) Summary comparison of decision-curve analysis in the external validation setting. AUC, area under the curve; M1, all-subtype model; M2, pure-solid-specific model; ROC, receiver operating characteristic. Calibration and clinical utility For M1, calibration curves showed acceptable agreement between predicted and observed probabilities in the internal cohort, the overall external validation cohort, and the subtype-stratified external cohorts (Figure 3C,3D). Acceptable calibration was also noted for M2 in the internal and external pure-solid cohorts (Figure 3C,3D). Table S4 lists detailed calibration metrics for M1 in the overall external cohort and the subtype-stratified external cohorts. Across threshold probabilities of 5–30%, decision-curve analysis revealed positive net benefit for M1 across the internal cohort, the overall external cohort, and the subtype-stratified external cohorts. M2 also showed positive net benefit in both the internal and external pure-solid cohorts (Figure 3E,3F). In the subgroup of pure-solid nodules measuring ≤2 cm, both models retained measurable net benefit at representative threshold probabilities of 0.2, 0.3, and 0.5 (Figure S2 and Tables S5-S7). Model comparison in the pure-solid external validation cohort Although M2 did not outperform M1 in terms of AUC, reclassification analysis was performed to evaluate individual risk reassignment in the external pure-solid cohort. In the external pure-solid cohort, the AUC was 0.788 for M1 and 0.768 for M2, with the DeLong test showing a significant difference (P=0.0018). Reclassification analysis showed an NRI of 0.643 (95% CI, 0.387–0.880) and an IDI of 0.184 (95% CI, 0.125–0.242). The comparative ROC curves and corresponding AUC values are shown in Figure 4A, and the calibration plots, binned calibration curves, and decision-curve analysis are shown in Figure 4B-4D. The corresponding calibration metrics and decision-curve results are provided in Table S4, and the model comparison results are summarized in Table 4. Comparison of M1 and M2 in the external pure-solid cohort. (A) ROC curves and corresponding AUC values. (B) Calibration plots. (C) Binned calibration curves. (D) Decision-curve analysis. Curve A represents M1 in the internal cohort; curve B represents M1 in the external pure-solid cohort; curve C represents M2 in the internal pure-solid cohort; and curve D represents M2 in the external pure-solid cohort. AUC, area under the curve; M1, all-subtype model; M2, pure-solid-specific model; ROC, receiver operating characteristic; STAS, spread through air spaces. Table 4Comparison predictive performance of M1 and M2 in the external pure-solid cohortMetricValueExternal pure-solid cohort, n203STAS-positive, n (%)102 (50.2%)AUC, M10.788AUC, M20.768DeLong P0.0018Total NRI (M2 vs. M1)0.643 (95% CI: 0.387–0.880)IDI (M2 vs. M1)0.184 (95% CI: 0.125–0.242)AUC, area under the curve; CI, confidence interval; IDI, integrated discrimination improvement; M1, all-subtype model; M2, pure-solid-specific model; NRI, net reclassification improvement; STAS, spread through air spaces. Comparison predictive performance of M1 and M2 in the external pure-solid cohort AUC, area under the curve; CI, confidence interval; IDI, integrated discrimination improvement; M1, all-subtype model; M2, pure-solid-specific model; NRI, net reclassification improvement; STAS, spread through air spaces. Subtype dependency of model performance and misclassification pattern In the internal M1 cohort, the overall misclassification rate was 23.6%, and 74.8% of misclassified patients were from the pure-solid subgroup. There was a clear gradient across nodule subtypes, with misclassification rates of 31.6% in pure-solid nodules, 14.6% in part-solid nodules, and 9.5% in pure ground-glass nodules (Table S8 and Figure S3). Pure-solid nodules therefore accounted for a disproportionate share of prediction errors relative to their proportion in the internal cohort. In the paired pure-solid comparison, M2 corrected 73 errors made by M1 but introduced 62 new errors in the internal cohort, with no significant paired difference in overall correctness (McNemar P=0.390). In the external pure-solid cohort, M2 corrected 18 errors made by M1 but introduced 22 new errors, also with no significant paired difference in overall correctness (McNemar P=0.636). Compared with M1, M2 produced slightly fewer false-positive predictions (18 vs. 19) but more false-negative predictions (43 vs. 38) in the external pure-solid cohort (Figure S4 and Table S9). Further evaluation showed that false-negative predictions by M1 were concentrated mainly in tumours measuring ≤2 cm, poorly differentiated tumours, and tumours with micropapillary or solid histological patterns (Table S10). Model interpretability Model interpretability analysis revealed different feature-dependence patterns between M1 and M2. SHAP analysis showed that M1 was mainly reliant on texture-related descriptors, kurtosis, and higher-order heterogeneity features, whereas M2 was strongly dependent on intensity-based statistics such as mean and median, together with comparatively homogeneous texture descriptors including inverse variance. These patterns are displayed in the SHAP summary bar plots (Figure 5A,5B) and beeswarm plots (Figure 5C,5D). The aligned comparison signals a shift in feature-learning direction rather than a simple difference in feature ranking (Figure 5E, Figures S5,S6, and Table S11). For individual sample prediction, two representative patients were selected for force-plot-style visualisation (Figure 6). Feature analysis and model interpretability of M1 and M2. (A,B) SHAP summary bar plots. (C,D) SHAP beeswarm plots. (E) Comparison of SHAP-informed feature patterns between M1 and M2. M1, all-subtype model; M2, pure-solid-specific model; SHAP, Shapley additive explanations. Representative cases showing CT images, ROI delineation, and SHAP-based explanatory profiles for M1. CT, computed tomography; M1, all-subtype model; ROI, region of interest; SHAP, Shapley additive explanations. Discussion In this multi-centre study, two radiomics models were developed for preoperative STAS prediction in stage I lung ADC: an M1 and a model restricted to M2. Subtype-stratified validation showed that radiomics has value for preoperative STAS assessment, although model performance may differ across nodule subtypes. Among these subtypes, pure-solid nodules remained the most difficult to classify, demonstrating moderate discrimination together with a higher proportion of misclassified cases. Although M2 did not significantly outperform M1 in terms of AUC within the external pure-solid cohort, it demonstrated complementary individual-level risk reclassification. These results confirm that radiomics-based prediction of STAS is influenced by nodule subtype and that subtype-specific modelling may offer complementary value in pure-solid nodules. STAS is linked to a higher risk of local recurrence and worse recurrence-free and overall survival after sublobar resection in stage I ADC (7,8). Accordingly, accurate preoperative prediction may be clinically relevant for surgical decision-making. Prior studies have also established that STAS is more prevalent in solid nodules as well as in micropapillary- and solid-predominant ADC. A similar pattern was observed in our cohort, supporting the clinical relevance of nodule subtype in preoperative STAS prediction. In mixed-subtype cohorts, prior radiomics studies have reported considerable variation in predictive performance for STAS, with AUCs ranging from 0.63 to 0.91 (20-30). For example, Cui et al. (29) reported that a model derived from solid components achieved better performance than that based on the whole nodule region or on ground-glass components. Similarly, Chen et al. (26) derived relatively stable performance using patch-wise prediction in high-grade ADC subtypes. In addition, a number of studies reported a higher proportion of solid nodules among STAS-positive than among STAS-negative cases (20,22,25,27,28,30). Chen et al. (21) observed lower predictive performance when groups did not differ significantly in the distribution of solid nodules. In contrast, Zhuo et al. (30) reported excellent external validation performance in a cohort with marked imbalance in the proportion of solid components between the two groups. These observations indicate that part of the apparent performance of radiomics-based prediction of STAS may reflect subtype composition, particularly the proportion of solid nodules, rather than the true incremental contribution of the model itself. A key issue is whether radiomics captures STAS-related imaging phenotypes beyond subtype-related attenuation differences. In the present study, subtype-adjusted analysis suggested that the Rad-score was not solely attributable to nodule subtype. Nevertheless, subtype-stratified validation and misclassification analyses indicate that subtype remains an important driver of model performance. Subtype distribution should therefore be considered during model development and validation. M1 performed best when all of the nodule subtypes were analysed together in the internal cohort and in the overall external cohort. After subtype-stratified validation, performance declined in both the non-solid and pure-solid cohorts. Additionally, most misclassified cases were concentrated in the pure-solid subgroup, where STAS prevalence was also higher. This error pattern suggests a potential limitation of hand-crafted radiomics in pure-solid nodules. The relatively homogeneous CT appearance of pure-solid nodules may reduce the discriminative value of conventional intensity and texture features. In addition, false-negative cases were more common among small tumours and lesions with adverse pathological features, suggesting that some STAS-positive pure-solid nodules may remain radiomically inconspicuous on CT. While M2 showed a lower AUC than M1 in the external pure-solid cohort, it provided complementary individual-level risk reclassification, supporting a complementary rather than superior role for M2. Favourable NRI and IDI results should therefore be interpreted cautiously and require prospective decision-impact validation before clinical superiority can be inferred. Accordingly, negative predictions in pure-solid nodules should be interpreted cautiously. Future models may benefit from deep learning features (32), peritumoural information (33), and closer radiology–pathology correlation. SHAP analysis provided further insight into the behaviour of the two models. Both models were characterised by features such as grey-level heterogeneity and clustered high-intensity dependence, suggesting that these may represent relatively robust radiomic correlates of STAS. However, the two models appeared to capture different aspects of imaging variation. While M1 retained a broad pattern of cross-subtype structural information, M2 was heavily reliant on intensity-based statistics and subtype-specific intensity variation within pure-solid nodules. This divergence in feature dependence may help explain why M2 did not improve overall discrimination, as measured by the AUC, while still demonstrating more favourable reclassification performance. SHAP findings should be regarded as model-level explanations rather than direct biological evidence of STAS. Texture and intensity related features may reflect tumour heterogeneity, but should not be equated with specific pathological structure. The Rad-score should therefore be interpreted as a composite imaging risk score. Clinically, these findings support a stratified use of radiomics models rather than uniform application. M1 may be useful as an initial risk-stratification tool across the full spectrum of nodule subtypes in stage I ADC. Conversely, predictions in pure-solid nodules remain particularly difficult and should be regarded as supportive rather than definitive evidence. For small pure-solid nodules (≤2 cm) that may be candidates for sublobar resection (7,8), a more cautious decision-making approach may combine radiomics-based assessment with intraoperative frozen-section results, margin considerations, and other clinicopathological factors. The proposed models are not intended for all indeterminate pulmonary nodules; rather, they are intended for STAS risk assessment in patients with CT findings highly suggestive of early-stage lung ADC who are being considered for surgery. Limitations The limitations of this study must be acknowledged. First, despite external validation performed across multiple centres, the retrospective design did not allow full control of potential inter-centre heterogeneity in scanning parameters and pathological interpretation. Inter-centre variability in scanner vendor, reconstruction kernel, slice thickness, acquisition protocol, and case mix may have affected radiomic feature distributions and contributed to domain shift. Although resampling and intensity normalisation were applied, these procedures cannot fully remove centre-related variability. Prospective validation using standardised acquisition protocols or harmonisation strategies is therefore warranted before clinical implementation. Second, the distribution of cases among nodule subtypes lacked balance, which may have influenced the robustness of some subgroup analyses. This imbalance is particularly relevant for non-solid nodules, in which the lower prevalence of STAS may result in unstable sensitivity estimates. Third, the study included only surgically resected and pathologically confirmed lung ADC, which may introduce selection bias and limits the direct application of the model to unselected screening populations or to nodules with benign, metastatic, or non-ADC pathology. Fourth, the present study was primarily based on radiomics. Therefore, future studies should explore whether predictive performance can be enhanced by incorporating deep learning, multimodal pathological information, or molecular biomarkers. Finally, while radiomics-based models may be useful for preoperative risk stratification, whether such models can ultimately guide optimised surgical strategies and improve long-term outcomes requires confirmation in larger prospective multi-centre studies. Conclusions Radiomics demonstrated potential for preoperative STAS prediction in stage I lung ADC; however, model performance appeared to vary by nodule subtype. While the all-subtype model preserved favourable overall discrimination, the pure-solid-specific model altered the pattern of risk assignment and provided incremental reclassification benefit. These findings support a stratified nodule type strategy for radiomics-based STAS prediction. Supplementary The article’s supplementary files as