A radiomics-based machine learning model for the preoperative differentiation of lung adenocarcinoma subtypes Authors: Fengjuan Tian, Jing Ding, Zhenyu Cao, Dengfa Yang, Hengfeng Shi, Jian Wang, Hongjie Hu Published: 2026 ABSTRACT BackgroundThe preoperative differentiation of lung adenocarcinoma subtypes is critical for implementing personalized treatment but is difficult to accomplish with conventional imaging. This study aimed to develop an interpretable multimodal model integrating clinical, peritumoral, radiomic, and deep learning features to improve diagnostic accuracy.MethodsA total of 3,038 patients from four hospitals were divided into training (n=1,822), test (n=608), and validation (n=608) sets. Two radiologists manually segmented two-dimensional tumor regions on computed tomography using ITK-SNAP software. After Pearson correlation analysis and least absolute shrinkage and selection operator regression, the radiomic score and deep learning score were generated. Clinical features were selected via univariate analysis, the Boruta algorithm, and recursive feature elimination (RFE). Individual logistic models were built and fused with the optimal combination selected via support vector machine-synthetic minority oversampling technique and extreme gradient boosting. Performance was evaluated in terms of the Obuchowski index, accuracy, F1-score, calibration, and decision curves, while interpretability was assessed via Shapley additive explanations (SHAP) and individual conditional expectation (ICE).ResultsThe fused model achieved Obuchowski indices of 0.85 [95% confidence interval (CI): 0.84–0.87], 0.81 (95% CI: 0.78–0.83), and 0.79 (95% CI: 0.76–0.81) in the training, test, and validation sets, respectively outperforming the single-modality models. The F1-scores for the lepidic, acinar/papillary, and solid/micropapillary subtypes, respectively, were 0.77, 0.61, and 0.63 in the training set; 0.72, 0.57, and 0.59 in the test set; and 0.74, 0.54, and 0.52 in the validation set. Calibration and decision curve analysis confirmed the robustness and clinical utility of the model. SHAP analysis identified ResNet-101 feature as the best predictor, followed by peritumoral radiomic score, and lobulation. ICE plots revealed the linear and monotonic relationships between key features and predicted probabilities across subtypes.ConclusionsThe radiomics model developed in this study facilitates the accurate and interpretable preoperative classification of lung adenocarcinoma subtypes. Fusion of clinical, peritumoral, and deep learning features enhances diagnostic performance and supports clinical decision-making. Introduction Lung adenocarcinoma, the most common histological subtype of non-small cell lung cancer (NSCLC), exhibits significant heterogeneity in biological behavior and clinical outcomes (1). The frequency in detection of pulmonary ground-glass nodules (GGNs) has increased due to the widespread use of low-dose computed tomography (CT) screening, and thus accurate preoperative subtyping has become crucial for personalized management (2). According to the 2021 World Health Organization (WHO) classification, lung adenocarcinoma can be classified into lepidic, acinar/papillary, and solid/micropapillary histological patterns (3), which reflect different biological behaviors. Tumors with the lepidic pattern grow slowly, are often noninvasive or minimally invasive, and are associated with excellent survival after surgery (4,5); those with the acinar or papillary pattern, the most common subtype, are moderately aggressive (6,7); and those with the solid or micropapillary pattern are highly aggressive and associated with a higher risk of lymph node metastasis, vascular invasion, and recurrence (8). Therefore, the survival rate is lower among patients with the solid/micropapillary subtype, even at early disease stages (9). The preoperative identification of these subtypes helps guide surgery, and pure lepidic or mixed tumors with a dominant lepidic component may be treated with sublobar resection. Tumors with solid or micropapillary components often require lobectomy with lymph node dissection, and information regarding subtype may inform prognosis. Patients with solid or micropapillary patterns may benefit from closer follow-up and are more likely to be considered for adjuvant chemotherapy after surgery (10). Although histology itself does not directly guide targeted therapy or immunotherapy, aggressive subtypes are often accompanied by molecular alterations (11). However, treatment decisions remain primarily based on stage, molecular profiling, and patient factors. Therefore, precise preoperative differentiation of these subtypes is essential for guiding the selection of surgical strategy (sublobar resection or lobectomy) and determining the need for adjuvant therapy. Conventional radiological assessment, which is predominantly based on subjective visual interpretation of CT features such as tumor size, density, and growth rate, often lacks sufficient accuracy in distinguishing these histological patterns, particularly in small or mixed GGNs. In this context, radiomics has emerged as a promising tool, as it can facilitate the high-throughput extraction of quantitative imaging features from medical images, transforming them into mineable data for decision support (6). By capturing subtle intratumoral heterogeneity, radiomics provides insights into tumor biology that are not discernible to the human eye (12-14). Recent evidence suggests that the peritumoral region—the tissue immediately surrounding the tumor—also plays a critical role in tumor progression through mechanisms such as inflammation, angiogenesis, and immune modulation (15,16). Peritumoral radiomics extends the analysis beyond the tumor boundary, offering a more comprehensive characterization of the tumor microenvironment and improving diagnostic performance (17). Despite its strengths, traditional radiomics depends heavily on handcrafted features and may fail to capture complex spatial patterns and high-level semantic information. In contrast, deep learning models, particularly convolutional neural networks (CNNs) such as ResNet, automatically learn hierarchical feature representations directly from raw images through end-to-end training. These models can extract high-dimensional, abstract features that reflect intricate textural and structural patterns, often surpassing conventional radiomic features in discriminative power. Moreover, deep learning is less sensitive to image preprocessing and segmentation variability, enhancing reproducibility. In this study, we developed a multimodal machine learning framework that integrates clinical variables, intratumoral and peritumoral radiomics, and deep learning features extracted via ResNet to improve the preoperative differentiation of the lepidic, acinar/papillary, and solid/micropapillary lung adenocarcinoma subtypes. By leveraging complementary information from multiple data sources, our approach aims to provide a more accurate, noninvasive, and clinically applicable tool for precision diagnosis and individualized treatment planning. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0779/rc). Methods Study design and patient cohort A total of 3,038 patients with pathologically confirmed lung adenocarcinoma were retrospectively enrolled from four medical centers (Sir Run Run Shaw Hospital, Zhejiang University School of Medicine; Tongde Hospital of Zhejiang Province Affiliated to Zhejiang Chinese Medical University; Anqing Municipal Hospital; and Taizhou Municipal Hospital) across two provinces in China between January 2015 and December 2024. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and by the institutional ethics committees of each participating hospital (Sir Run Run Shaw Hospital, Zhejiang University School of Medicine: approval No. 2025-0642; Tongde Hospital of Zhejiang Province Affiliated to Zhejiang Chinese Medical University: approval No. MR-33-24-041194; Anqing Municipal Hospital: approval No. 83230471; Taizhou Municipal Hospital: approval No. LWYJ202300055). The requirement for informed consent was waived due the retrospective nature of the analysis. All patients underwent preoperative noncontrast chest CT scans prior to histopathological confirmation via surgical resection. Tumor subtypes were classified as lepidic, acinar/papillary, or solid/micropapillary based on the 2021 WHO classification of lung adenocarcinoma (Figure 1). Because invasive lung adenocarcinomas often contain mixed histological patterns, the percentage of each component was assessed, and tumors were assigned to one of three clinically relevant categories according to the predominant and/or highest-grade component. Lepidic-predominant tumors placed into the lepidic group, acinar- or papillary-predominant tumors into the acinar/papillary group, and tumors with solid or micropapillary predominant/high-grade components into the solid/micropapillary group. Surgical procedures were selected according to tumor characteristics, suspected histological aggressiveness, patient condition, and surgeon discretion. In general, lepidic tumors are often treated with wedge resection with sentinel lymph node assessment or sampling; acinar/papillary tumors are treated with wedge resection or segmentectomy; and solid/micropapillary tumors are treated with lobectomy or segmentectomy with systematic mediastinal lymph node dissection. Pathological lymph node status was recorded according to the final surgical pathology report. The initial cohort consisted of 4,138 patients who underwent surgery during the study period. Patients were excluded if they had incomplete thin-slice (0.625–2.0 mm) CT images (n=512), significant motion artifacts (n=90), chemoradiotherapy administered before CT scanning (n=282), or a diagnosis of metastatic tumors or small-cell lung cancer (n=216). After these inclusion and exclusion criteria were applied, a final cohort of 3,038 patients was included in the study. Flowchart of patient inclusion. CT, computed tomography. The cohort was randomly divided into a training (n=1,822), test (n=608), and validation set (n=608) at a 6:2:2 ratio. The distribution of histological subtypes was as follows: lepidic, 1,315 tumors, acinar/papillary, 1,213 tumors; and solid/micropapillary, 510 tumors. Region-of-interest segmentation Two experienced radiologists (with 8 and 12 years of thoracic imaging experience, respectively) independently performed manual segmentation of the tumor region on axial noncontrast CT images using ITK-SNAP v. 3.8.0. A two-dimensional (2D) region of interest (ROI) was delineated on the slice showing the maximum cross-sectional area of the nodule. To incorporate peritumoral information, four peritumoral annular zones were automatically generated with a Python-based dilation algorithm, including 5-, 10-, 15-, and 20-mm beyond the tumor boundary. This approach ensured consistent and objective definition of the peritumoral region. Interobserver agreement for intratumoral segmentation was assessed via intraclass correlation coefficients (ICCs), and discrepancies were resolved through consensus reading. Radiomic feature extraction Radiomic features were extracted from both intratumoral and peritumoral ROIs with PyRadiomics package v. 3.1.0 in Python v. 3.9 (Python Software Foundation, Wilmington, DE, USA). A total of 107 first-order, shape, and texture features were calculated after image resampling to 1×1×1 mm3 and intensity discretization. To extract high-level semantic features, a ResNet-101 model pretrained on ImageNet was fine-tuned via transfer learning. The central axial slice of each nodule was resized to 224×224 and normalized. Features from the penultimate fully connected layer (2,048 dimensions) were extracted with PyTorch 2.0. No additional training data augmentation was applied during feature extraction (Figure 2). Overview of the study design and analytical pipeline. The workflow encompassed data collection, feature extraction from multimodal sources (clinical, radiomic, and deep learning), feature selection, model development, and performance evaluation across independent cohorts. DCA, decision curve analysis; FPR, false positive rate; ICE, individual conditional expectation; KNN, K-nearest neighbors; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; MLP, multilayer perceptron; RF, random forest; ROC, receiver operating characteristic; ROI, region of interest; SHAP, Shapley additive explanations; SMOTE, synthetic minority oversampling technique; SVM, support vector machine; SVMSMOTE, support vector machine-SMOTE; TPR, true positive rate; XGBoost, extreme gradient boosting. Feature preprocessing and dimensionality reduction All radiomic and deep learning features were subjected to rigorous preprocessing: Pearson correlation analysis (|r| >0.8) was used to remove highly correlated radiomic features. Least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was applied to select the most predictive features for each modality. Two linear combination scores—radiomics score (Rad-score; from selected radiomic features) and deep learning score (DL-score; from selected deep learning features)—were constructed via regression coefficients (see Appendix 1 for the relevant formulae) (Figure 3). Model performance and feature selection for peritumoral radiomics and deep learning (ResNet-101) features. (A-C) Radiomics-based model: ROC curve, LASSO coefficient path, and LASSO regression weight plot. (D-F) ResNet-101-based model: ROC curve, LASSO coefficient path, and LASSO regression weight plot. The LASSO procedure was used to select the most predictive features while minimizing overfitting. AUC, area under the receiver operating characteristic curve; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic. Clinical feature selection Data on 20 clinical variables were initially collected, including age, sex, smoking status, tumor size, and carcinoembryonic antigen (CEA) level. Univariate analysis was performed to screen for potential predictors. Continuous variables were compared via one-way analysis of variance (ANOVA) or the Kruskal-Wallis test (for nonnormally distributed data), while categorical variables were compared via the chi-squared test or Fisher exact test (when ≥20% of expected cell counts were less than 5). Significant omnibus results were followed by pairwise post hoc comparisons with Bonferroni correction. Variables with a P value <0.05 were retained for further analysis. The Boruta algorithm (20 iterations) was then applied to identify clinically relevant features, resulting in 11 selected clinical variables. Radiomic and deep learning features—Rad-score and DL-score—were independently constructed through LASSO Cox regression and integrated into the feature pool. Subsequently, recursive feature elimination (RFE) was implemented with fivefold cross-validation and F1-score optimization to jointly evaluate all candidate features (clinicals, radiomics, and deep learning). This final step led to the parsimonious selection of 13 features for inclusion in the fused model (18) (Figure 4). Feature selection results using the Boruta algorithm (A) and RFE (B). The scores shown in this figure represent feature-selection ranking during model development and should not be interpreted as the final predictive importance of variables in the fused model. CTmax, maximum computed tomography attenuation value; CTmean, mean computed tomography attenuation value; CTmin, minimum computed tomography attenuation value; CTsd, standard deviation of computed tomography attenuation; CV, cross-validation; HU, Hounsfield unit; LD, long diameter; mGGO, mixed ground-glass opacity; RFE, recursive feature elimination; SD, short diameter; UPPV, uneven perfusion of pulmonary ventilation. Model development and fusion To evaluate the impact of peritumoral extent, multiple radiomic models were developed: intratumoral-only, peritumoral-only (5-, 10-, 15-, and 20-mm annular zones), and combined intratumoral-peritumoral models. For each model, radiomic features were extracted and a Rad-score was constructed via LASSO logistic regression to select the optimal peritumoral extent. The performance of each Rad-score was assessed via logistic regression in the training set. Separately, a deep learning model was built through construction of a DL-score from features extracted by ResNet-101. The radiomic model with the best cross-validated performance—specifically, the combined intratumoral and 10-mm peritumoral model—was selected. Its Rad-score was then combined with the ResNet-101 DL-score and the selected clinical features to form a multimodal feature set for subsequent fusion modeling (Table S1). Machine learning To address class imbalance in the trichotomous classification task, six oversampling techniques—including synthetic minority oversampling technique (SMOTE), adaptive synthetic sampling (ADASYN), support vector machine-SMOTE (SVMSMOTE), and others—were tested in combination with six machine learning classifiers [logistic regression, random forest, SVM, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), categorical boosting (CatBoost)] under default hyperparameters. Model performance was validated via fivefold cross-validation. The combination of SVMSMOTE oversampling and XGBoost classifier yielded the highest average F1-score and was selected as the final fusion model. Model evaluation and interpretability The final model was evaluated according to multiple performance metrics. Multiclass receiver operating characteristic (ROC) curves and macro-averaged area under the curve (AUC) were computed to assess the overall discriminative ability across the three histological subtypes. In addition, the Obuchowski index was calculated as a multiclass extension of the binary AUC to summarize the average pairwise discriminatory ability among the three subtypes. This metric provides a single overall estimate of model discrimination in the multiclass setting, complementing the macro-averaged AUC. Calibration curves were generated to evaluate the agreement between the predicted probabilities and the observed outcomes, with a diagonal line representing perfect calibration. Decision curve analysis (DCA) was performed to assess the clinical utility of the model across a range of threshold probabilities (Figure 5). To enhance interpretability, Shapley additive explanations (SHAP) values were calculated to visualize the relative importance of each feature in the model’s predictions. It should be noted that the feature-selection scores shown in Figure 4 and the SHAP values shown in Figure 6A represent different analytical stages. Figure 4 reflects the ranking used for feature selection, whereas Figure 6A reflects the contribution of each feature to predictions made by the final fused model. Additionally, individual conditional expectation (ICE) plots were constructed to illustrate how variations in individual features influence the predicted probabilities across different patient subgroups, providing insight into model behavior at the instance level. Performance evaluation of the fused multimodal model in the training, test, and validation cohorts. (A-C) ROC curves. (D-F) Calibration plots assessing the agreement between predicted and observed outcomes. (G-I) Confusion matrices at an optimal decision threshold. (J) DCA evaluating the clinical utility of the model across a range of threshold probabilities. AUC, area under the curve; DCA, decision curve analysis; ROC, receiver operating characteristic; XGBoost, extreme gradient boosting. Model interpretability. (A) SHAP summary plot showing global feature importance in the fused model. (B-D) For representative cases predicted as lepidic (B), acinar/papillary (C), and solid/micropapillary (D) subtypes, each panel shows a SHAP waterfall plot depicting the contribution of individual features to the prediction (positive values support the predicted class, while negative values oppose it) (top) and the corresponding Grad-CAM heatmap overlaid on the CT image, highlighting the regions most influential for the model’s decision (bottom). Grad-CAM, Gradient-weighted Class Activation Mapping; CT, computed tomography; SHAP, Shapley additive explanations; UPPV, uneven perfusion of pulmonary ventilation. Results Patients A total of 3,038 patients were included in this study, with 1,822 in the training set, 608 in the internal test set, and 608 in the validation set. The mean age was 60.02±11.31 years, with 1,158 (38.12%) male and 1,880 (61.88%) female patients. Histopathological subtypes, including lepidic (n=1,315), acinar/papillary (n=1,213), and solid/micropapillary (n=510), were relatively balanced across the three cohorts. No significant differences in baseline clinical characteristics were observed between the training, internal test, and validation sets (all P values >0.05), indicating good comparability among the cohorts (Tables 1,2). Table 1Clinical characteristics of patients according to histological subtype (lepidic, acinar/papillary, and solid/micropapillary)VariableTraining set (n=1,822)Test set (n=608)Validation set (n=608)Lepidic (n=789)Acinar/papillary (n=727)Solid/micropapillary (n=306)PLepidic (n=263)Acinar/papillary (n=243)Solid/micropapillary (n=102)Lepidic (n=263)Acinar/papillary (n=243)Solid/micropapillary (n=102)Age (years)57.24±12.3160.16±10.9462.56±9.80<0.001*58.95±11.5960.79±11.1664.40±9.9457.43±11.4860.48±11.1463.50±9.81LD (mm)12.52±5.6116.59±6.1619.34±5.97<0.001*12.85±5.8316.61±6.0919.45±6.0012.21±5.1816.19±6.2819.08±5.70SD (mm)9.53±4.2712.26±4.8414.92±5.00<0.001*9.73±4.4812.22±4.8714.86±5.739.35±4.2612.24±5.1414.62±5.15CTmax (HU)−193.87±238.2429.13±230.39150.76±132.08<0.001*−181.51±235.4737.07±221.73170.29±96.71−189.63±244.3738.39±220.07153.77±130.36CTmin (HU)−480.38±232.10−271.79±260.53−117.25±167.64<0.001*−447.86±232.93−260.94±254.08−102.95±127.27−476.83±231.46−256.25±256.11−135.46±185.09CTmean (HU)−334.72±220.42−115.81±227.1118.24±114.98<0.001*−314.64±216.71−107.41±219.7633.23±65.81−328.28±229.04−104.26±219.2412.46±131.53CTsd (HU)91.11±58.5489.73±56.9568.36±42.41<0.001*86.93±47.1389.39±47.1770.08±39.1592.30±51.8189.94±50.7273.23±41.76Data are presented as mean ± standard deviation. *, significant differences among the lepidic, acinar/papillary, and solid/micropapillary subtypes. Statistical comparisons across these three groups were performed via one-way analysis of variance (or the Kruskal-Wallis test for nonnormally distributed variables). CTmax, maximum computed tomography attenuation value; CTmean, mean computed tomography attenuation value; CTmin, minimum computed tomography attenuation value; CTsd, standard deviation of computed tomography attenuation; HU, Hounsfield unit; LD, long diameter; SD, short diameter. Clinical characteristics of patients according to histological subtype (lepidic, acinar/papillary, and solid/micropapillary) Data are presented as mean ± standard deviation. *, significant differences among the lepidic, acinar/papillary, and solid/micropapillary subtypes. Statistical comparisons across these three groups were performed via one-way analysis of variance (or the Kruskal-Wallis test for nonnormally distributed variables). CTmax, maximum computed tomography attenuation value; CTmean, mean computed tomography attenuation value; CTmin, minimum computed tomography attenuation value; CTsd, standard deviation of computed tomography attenuation; HU, Hounsfield unit; LD, long diameter; SD, short diameter. Table 2Clinical characteristics of patientsVariableTraining set (n=1,822)Test set (n=608)Validation set (n=608)Lepidic (n=789)Acinar/papillary (n=727)Solid/micropapillary (n=306)PLepidic (n=263)Acinar/papillary (n=243)Solid/micropapillary (n=102)Lepidic (n=263)Acinar/papillary (n=243)Solid/micropapillary (n=102)Visceral pleural invasion‡<0.001 Yes10 (1.27)52 (7.15)92 (30.07)2 (0.76)13 (5.35)35 (34.31)1 (0.38)18 (7.41)27 (26.47)Lymphovascular invasion‡<0.001 Yes1 (0.13)29 (3.99)73 (23.86)3 (1.14)4 (1.65)27 (26.47)1 (0.38)9 (3.70)22 (21.57)Positive LN‡<0.001 Yes2 (0.25)34 (4.68)92 (30.07)1 (0.38)9 (3.70)25 (24.51)2 (0.76)17 (7.00)21 (20.59)Sex†<0.001 Female540 (68.44)458 (63.00)156 (50.98)162 (61.60)138 (56.79)51 (50.00)185 (70.34)147 (60.49)43 (42.16) Male249 (31.56)269 (37.00)150 (49.02)101 (38.40)105 (43.21)51 (50.00)78 (29.66)96 (39.51)59 (57.84)Smoking†0.010 Never712 (90.24)641 (88.17)254 (83.01)235 (89.35)208 (85.60)86 (84.31)242 (92.02)217 (89.30)84 (82.35) Current45 (5.70)52 (7.15)36 (11.76)15 (5.70)18 (7.41)7 (6.86)12 (4.56)14 (5.76)11 (10.78) Former32 (4.06)34 (4.68)16 (5.23)13 (4.94)17 (7.00)9 (8.82)9 (3.42)12 (4.94)7 (6.86)High blood pressure†0.656 Yes99 (12.55)90 (12.38)44 (14.38)35 (13.31)32 (13.17)19 (18.63)34 (12.93)28 (11.52)27 (26.47)Diabetes mellitus‡0.180 Yes31 (3.93)21 (2.89)16 (5.23)14 (5.32)10 (4.12)7 (6.86)11 (4.18)9 (3.70)3 (2.94)Emphysema or bullae of lung†<0.001 Yes100 (12.67)103 (14.17)79 (25.82)23 (8.75)34 (13.99)21 (20.59)31 (11.79)41 (16.87)24 (23.53)Uneven perfusion of pulmonary ventilation†<0.001 Yes73 (9.25)121 (16.64)49 (16.01)32 (12.17)39 (16.05)17 (16.67)29 (11.03)28 (11.52)13 (12.75)Tumor indicator†<0.001 Yes193 (24.46)225 (30.99)139 (45.42)60 (22.81)74 (30.45)45 (44.12)61 (23.28)68 (27.98)54 (52.94)Location†0.021 Right upper lobe293 (37.14)224 (30.81)89 (29.08)98 (37.26)65 (26.75)24 (23.53)80 (30.42)78 (32.10)24 (23.53) Right middle lobe67 (8.49)56 (7.70)22 (7.19)28 (10.65)25 (10.29)12 (11.76)20 (7.60)21 (8.64)8 (7.84) Right lower lobe135 (17.11)156 (21.46)51 (16.67)41 (15.59)54 (22.22)26 (25.49)40 (15.21)51 (20.99)28 (27.45) Left upper lobe192 (24.33)190 (26.13)88 (28.76)68 (25.86)70 (28.81)27 (26.47)81 (30.80)64 (26.34)27 (26.47) Left lower lobe102 (12.93)101 (13.89)56 (18.30)28 (10.65)29 (11.93)13 (12.75)42 (15.97)29 (11.93)15 (14.71)Lobulation†<0.001 Yes388 (49.18)565 (77.72)269 (87.91)132 (50.19)201 (82.72)95 (93.14)127 (48.29)192 (79.01)94 (92.16)Spiculation†<0.001 None644 (81.62)430 (59.15)123 (40.20)221 (84.03)151 (62.14)38 (37.25)219 (83.27)153 (62.96)39 (38.24) Short119 (15.08)221 (30.40)132 (43.14)39 (14.83)66 (27.16)41 (40.20)36 (13.69)69 (28.40)41 (40.20) Long26 (3.30)76 (10.45)51 (16.67)3 (1.14)26 (10.70)23 (22.55)8 (3.04)21 (8.64)22 (21.57)Airspace†<0.001 Yes104 (13.18)118 (16.23)69 (22.55)22 (8.37)51 (20.99)18 (17.65)33 (12.55)39 (16.05)27 (26.47)Air bronchogram†<0.001 Yes90 (11.41)165 (22.70)78 (25.49)22 (8.37)60 (24.69)35 (34.31)37 (14.07)70 (28.81)32 (31.37)mGGO†<0.001 0537 (68.06)205 (28.20)10 (3.27)175 (66.54)59 (24.28)0173 (65.78)60 (24.69)3 (2.94) ≤25%178 (22.56)223 (30.67)26 (8.50)63 (23.95)103 (42.39)10 (9.80)65 (24.71)81 (33.33)10 (9.80) ≤50%24 (3.04)67 (9.22)22 (7.19)12 (4.56)15 (6.17)4 (3.92)11 (4.18)26 (10.70)6 (5.88) ≤75%16 (2.03)56 (7.70)23 (7.52)7 (2.66)20 (8.23)8 (7.84)5 (1.90)13 (5.35)9 (8.82) <100%15 (1.90)68 (9.35)41 (13.40)2 (0.76)14 (5.76)15 (14.71)4 (1.52)23 (9.47)10 (9.80) 100%19 (2.41)108 (14.86)184 (60.13)4 (1.52)32 (13.17)65 (63.73)5 (1.90)40 (16.46)64 (62.75)Data are presented as n (%). †, categorical variables were compared across the three groups via the omnibus Chi-squared test; ‡, the Fisher’s exact test was used when ≥20% of expected cell counts were <5. For variables showing significant overall differences, pairwise post hoc comparisons were performed with the Chi-squared or Fisher’s exact test with Bonferroni correction (adjusted α=0.05/3=0.0167). LN, lymph node; mGGO, mixed ground-glass opacity. Clinical characteristics of patients Data are presented as n (%). †, categorical variables were compared across the three groups via the omnibus Chi-squared test; ‡, the Fisher’s exact test was used when ≥20% of expected cell counts were <5. For variables showing significant overall differences, pairwise post hoc comparisons were performed with the Chi-squared or Fisher’s exact test with Bonferroni correction (adjusted α=0.05/3=0.0167). LN, lymph node; mGGO, mixed ground-glass opacity. Model performance The performance of the fused model was evaluated with the Obuchowski index—a multiclass extension of the AUC suitable for correlated outcomes—along with accuracy and F1-score. As shown in Table 3, the fused model (combining clinical, radiomic, and deep learning features) achieved excellent diagnostic performance in the training set, with an Obuchowski index of 0.85 [95% confidence interval (CI): 0.84–0.87]. The model also demonstrated good generalization capability, with Obuchowski indices of 0.81 (95% CI: 0.78–0.83) in the internal test set and 0.79 (95% CI: 0.76–0.81) in the validation set. Calibration curves showed good agreement between predicted probabilities and observed outcomes for all three histological subtypes, particularly for the lepidic and solid/micropapillary patterns. DCA further indicated that the fused model provided the highest net benefit across a broad range of threshold probabilities, demonstrating favorable clinical utility. Table 3Model comparisonVariableTraining set (n=1,822)Test set (n=608)Validation set (n=608)PrecisionRecallF1-scorePrecisionRecallF1-scorePrecisionRecallF1-scoreLepidic0.720.820.770.690.760.720.700.790.74Acinar/papillary0.640.580.610.570.570.570.550.520.54Solid/micropapillary0.690.580.630.680.520.590.590.460.52Accuracy0.680.680.680.640.640.640.630.630.63Macro-average0.680.660.670.650.620.630.610.590.60Weighted average0.680.680.680.640.640.640.620.630.62Obuchowski index0.85370.80580.786095% CI0.8401–0.86670.7773–0.83120.7604–0.8120CI, confidence interval. Model comparison CI, confidence interval. Feature interpretability For interpretation of the fused model, SHAP analysis was performed to assess feature importance. The most influential features were the ResNet-101-based deep learning feature, the peritumoral radiomic Rad-score, tumor size, spiculation, and age. Among these the ResNet-101 feature ranked highest, suggesting that high-level imaging patterns captured by the deep neural network played a dominant role in subtype classification (Figure 6A). Figure 6B-6D show SHAP the waterfall plots of representative cases predicted as lepidic, acinar/papillary, and solid/micropapillary, respectively. Each figure shows the influence of each feature in each case on the prediction of the model and visually presents the key features that support the diagnosis of this subtype and reduce the likelihood of other subtypes. ICE plots To evaluate how individual features influence predictions for each histological subtype, ICE plots were generated. For the lepidic subtype, a lower peritumoral radiomic Rad-score and lower ResNet-101 output were associated with higher predicted probability, whereas spiculation and higher ResNet-101 output were linked to reduced likelihood (Figure S1). In contrast, for the solid/micropapillary subtype, older age, spiculation, higher peritumoral Rad-score, and higher ResNet-101 output consistently increased the predicted probability (Figures S2,S3). Discussion Accurate preoperative differentiation of lung adenocarcinoma histological subtypes is critical for clinical practice, as it directly informs surgical strategy, prognostic counseling, and postoperative management. Although specimens from resection provide definitive histopathological classification, preoperative modalities—such as biopsy and cytology—are limited by sampling error, low yield, and the inability to capture spatial heterogeneity, particularly in mixed-pattern tumors (19,20). As a result, many patients undergo suboptimal surgical procedures or lack appropriate risk stratification prior to surgery (9,21). Noninvasive imaging-based prediction offers a promising solution to this issue. In this study, we developed and validated a multimodal machine learning model that integrates clinical features, peritumoral radiomics, and deep learning (ResNet) for the preoperative differentiation of lung adenocarcinoma subtypes, including lepidic, acinar/papillary, and solid/micropapillary. The fused model demonstrated robust performance across the training, test, and validation sets, with Obuchowski indices of 0.85, 0.81, and 0.79, respectively, outperforming individual modality models. Notably, the incorporation of SHAP and ICE analyses provided transparent, interpretable insights into model decision-making, enhancing its potential for clinical adoption. The superior performance of the fused model highlights the complementary strengths of integrating multimodal data. Clinical features provide interpretable and biologically relevant information, such as patient age and smoking history, which are well-established determinants of disease progression. We intentionally retained selected clinical variables as separate predictors rather than constructing a single clinical score because these variables are readily interpretable and may influence histological subtype through different biological or clinical pathways. Their separation allowed the model to capture their distinct contributions and enabled the transparent interpretation of feature effects. Radiomic features, in contrast, offer a reproducible means of quantifying heterogeneity within and around the tumor (22,23). Deep learning models, particularly ResNet-101, can further automatically extract high-dimensional representations that capture complex spatial patterns not easily discernible through handcrafted radiomic approaches (24). SHAP analysis revealed that features derived from ResNet-101 were the most influential in the final prediction, suggesting that the model can effectively identify subtle textural patterns (25). Importantly, the inclusion of peritumoral radiomics significantly enhanced model performance (26,27). By extending analysis beyond the tumor boundary, our model captures biologically relevant changes in the host-tumor interface, which may explain its improved discrimination between indolent (lepidic) and aggressive (solid/micropapillary) subtypes. Furthermore, the integration of model interpretability tools (SHAP and ICE) addresses a key barrier to the adoption of artificial intelligence in clinical practice: the “black box” problem. The SHAP waterfall plot illustrated how individual features collectively drive a prediction—for example, high ResNet activation and spiculation indicated a solid/micropapillary diagnosis, while ICE plots revealed consistent, biologically plausible patterns across patient populations. These visual explanations not only validate the model’s reasoning but also enhance clinician trust, thus promoting real-world implementation. Several limitations to this study should be acknowledged. First, 2D ROI segmentation was performed on the maximum cross-sectional slice, which, although efficient and clinically feasible, may not fully capture the 3D spatial heterogeneity of the tumor and peritumoral region. Future work will examine 3D deep learning and 2.5D multiple instance learning frameworks that leverage volumetric data for more comprehensive feature extraction. Second, manual segmentation was used for ROI delineation, which is time-consuming and subject to interobserver variability. Although high ICC values were achieved, the process limits scalability. Automated segmentation from deep learning-based tools (e.g., nnU-Net) will be investigated to improve efficiency and reproducibility. Third, the extent of lymph node assessment varied according to the type of surgical procedure. Patients undergoing wedge resection or segmentectomy received less extensive lymph node evaluation compared to those undergoing lobectomy, which may theoretically lead to an underestimation of pathological nodal involvement, particularly in less aggressive subtypes. However, given the low rates of nodal positivity observed in lepidic and acinar/papillary tumors and the fact that our model relies solely on preoperative imaging features, it is unlikely that the variability materially affected the study’s conclusions. Conclusions Our multimodal model combining clinical, radiomic, and deep learning features enables accurate and interpretable preoperative classification of lung adenocarcinoma subtypes (28). The fusion of high-dimensional deep features with peritumoral context and clinical data represents a significant advance in precision oncology. With further refinement and external validation, this approach has the potential to guide personalized surgical and adjuvant strategies in clinical practice. Supplementary The article’s supplementary files as