A Multimodal Radiomics and Pathomics Model for Predicting Postoperative Local Recurrence in T3–4 Non-Small Cell Lung Cancer

A Multimodal Radiomics and Pathomics Model for Predicting Postoperative Local Recurrence in T3–4 Non-Small Cell Lung Cancer | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Multimodal Radiomics and Pathomics Model for Predicting Postoperative Local Recurrence in T3–4 Non-Small Cell Lung Cancer Xinyu Li, Liying Wang, Yu Zong, Changsheng Zhou, Zhen Zhou, Jianrui Li, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8760093/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Background To develop and validate a multimodal artificial intelligence model integrating clinical, imaging, and pathological data to predict the risk of postoperative local recurrence in patients with T3-4 non-small cell lung cancer (NSCLC). Methods A total of 135 patients with pathologically confirmed T3–4 NSCLC who underwent complete surgical resection were retrospectively enrolled. Patients were stratified according to postoperative local recurrence status. Clinical variables, including preoperative contrast-enhanced CT images, and hematoxylin and eosin (H&E)–stained pathological slides were collected. Intratumoral and peritumoral radiomic features were extracted using PyRadiomics, while pathomics features were learned from whole-slide images using weakly supervised deep learning. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression. Radiomics, pathomics, and multimodal models were developed using a 70% training cohort with five-fold cross-validation and validated in the remaining 30%. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Results Among the 135 patients, 30 (22.2%) experienced postoperative local recurrence. The radiomics-based model achieved an AUC of 0.77 with a sensitivity of 0.77 and specificity of 0.65. The pathomics model alone demonstrated strong predictive performance (AUC = 0.92), while the integrated multimodal model achieved comparable discrimination and significantly outperformed the radiomics-only model (p < 0.05). Conclusions A multimodal AI model integrating conventional clinical data, CT-based radiomics, and histopathology-derived deep learning features enables accurate prediction of postoperative local recurrence in patients with T3–4 NSCLC. This approach may facilitate postoperative risk stratification, supporting individualized treatment and surveillance strategies in locally advanced NSCLC. non-small cell lung cancer radiomics pathomics multimodal imaging local recurrence Figures Figure 1 Figure 2 Introduction Lung cancer is the leading cause of cancer-related mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for approximately 85% of all cases[ 1 ]. For patients with resectable T3-4 NSCLC evaluated by a multidisciplinary team (MDT), radical surgical resection followed by platinum-based adjuvant chemotherapy remains the standard treatment. Despite advances in surgical techniques and systemic therapy, postoperative local recurrence continues to be a major determinant of long-term prognosis in this population. Previous studies have reported recurrence rates ranging from 33% to 50%[ 2 ] after radical resection in locally advanced NSCLC, with a five-year survival rate of only 33%[ 3 ]. Accurately identifying patients at high risk of postoperative local recurrence is still a critical clinical challenge in the postoperative management of locally advanced NSCLC. Conventional clinicopathological factors are currently the primary means for assessing the risk of postoperative local recurrence. These factors include tumor stage, lymph node involvement, histological subtype, margin status, and the presence of visceral pleural or vascular invasion. However, such factors are largely categorical and population-based, with limited ability to capture the continuous and heterogeneous biological behavior of individual tumors. Consequently, patients with similar clinicopathological profiles may experience markedly different risks of postoperative local recurrence, underscoring the limitations of existing risk stratification approaches for precise, individualized prediction. Therefore, it is crucial to explore more accurate prognostic biomarkers in surgically resected locally advanced NSCLC. Radiomics enables high-throughput extraction of quantitative features from medical images and has emerged as a noninvasive approach for characterizing tumor heterogeneity. Previous studies have demonstrated its value in predicting treatment response and prognosis in NSCLC across various therapeutic settings[ 4 – 8 ]. However, in surgically resected locally advanced NSCLC, particularly in T3–4 disease, the application of radiomics remains limited, and most existing models rely solely on imaging-derived features, which may be insufficient to capture the underlying biological complexity of tumors. Histopathological whole-slide images provide complementary information at the cellular and microenvironmental levels, capturing microscopic tumor characteristics closely associated with aggressiveness and recurrence risks. Pathomics, which applies computational and deep learning techniques to digital pathology images, enables quantitative characterization of these features but has not been well explored for predicting postoperative local recurrence in locally advanced NSCLC[ 9 ]. Multimodal artificial intelligence offers a powerful framework to integrate radiological and pathological information across different spatial scales, potentially providing a more comprehensive representation of tumor biology than single-modality models[ 10 ]. Nevertheless, the clinical value of integrating radiomics and pathomics for predicting postoperative local recurrence in surgically resected T3–4 NSCLC has not been well established. Therefore, the aim of this study was to develop and validate a multimodal artificial intelligence model integrating conventional clinical variables, preoperative contrast-enhanced CT–based intratumoral and peritumoral radiomic features, and postoperative histopathology-derived deep learning features to predict postoperative local recurrence in patients with T3–4 NSCLC. We further compared the predictive performance of radiomics-, pathomics-, and multimodal-based models to evaluate the incremental value of multimodal integration. Materials and Methods Study Population This retrospective study consecutively enrolled patients with pathologically confirmed T3–4 NSCLC who underwent complete surgical resection at Jinling Hospital between January 2013 and December 2020. Tumor staging was determined according to the 8th edition of the American Joint Committee on Cancer (AJCC) staging system. Inclusion criteria were as follows: (1) availability of preoperative contrast-enhanced chest CT performed within two weeks before surgery; and (2) pathologically confirmed T3 or T4 NSCLC. Exclusion criteria included inadequate CT image quality due to motion artifacts and the absence of reliable follow-up information regarding recurrence status and timing. Figure 1 Flowchart of multimodal model inclusion and exclusion criteria. CT Image Acquisition Preoperative contrast-enhanced CT images were acquired using multidetector CT scanners according to institutional protocols. Non-contrast, arterial, and portal venous phase images with thin-slice reconstruction were obtained. Arterial-phase images were used for subsequent segmentation and radiomic feature extraction. Detailed CT acquisition and reconstruction parameters for different scanner models are provided in Supplementary Table S1 . Collection of Clinical, Pathological, and Laboratory Features Preoperative clinical and laboratory variables were retrospectively collected, including neutrophil and lymphocyte percentages (N%, L%), carcinoembryonic antigen (CEA), squamous cell carcinoma antigen (SCCA), neuron-specific enolase (NSE), cytokeratin 19 fragment (CYFRA 21 − 1), carbohydrate antigen 724 (CA724), carbohydrate antigen 242 (CA242), carbohydrate antigen 199 (CA199), and carbohydrate antigen 125 (CA125), based on blood tests performed within one week before surgery. Variables with more than 20% missing values (CYFRA 21 − 1, CA724, and CA242) were excluded from analysis. Missing values for the remaining variables were imputed using mean imputation. Follow-Up Endpoints and Individualized Labeling The primary endpoint of this study was postoperative local recurrence, defined as tumor recurrence at the surgical margin or within the ipsilateral hemithorax after R0 resection during follow-up. Patients without local recurrence, including those with distant metastasis or no evidence of disease, were classified as the non-local recurrence group. Disease-free survival (DFS) was defined as the interval from surgery to recurrence, metastasis, death, or last follow-up and was used as a secondary outcome. Follow-up evaluations included serum tests and chest CT every 3–6 months during the first two years after surgery and every 6 months thereafter until death or December 31, 2023. Tumor and Peritumoral Tissue Segmentation All images were in DICOM format. We used arterial-phase mediastinal window images for segmentation and feature extraction. After anonymization, thin-slice DICOM images were imported into the Deepwise multimodal research platform ( https://keyan.deepwise.com ), an integrated machine learning platform incorporating PyRadiomics (version 3.0.1) and scikit-learn (version 0.22). Semi-automatic segmentation was performed: AI automatically segmented pulmonary lesions, followed by blinded review and necessary manual correction by radiologists, who were blinded to clinical outcomes. Reader 1 (with six years of thoracic imaging experience) independently segmented all lesions. A random selection of 20% of cases was segmented again by Reader 1 after one month and by Reader 2 (with three years of experience). Peritumoral tissue segmentation was achieved by expanding the tumor segmentation outward by 3 mm using platform software to form a ring-shaped region. Radiomics Feature Extraction and Selection Radiomic features were extracted from intratumoral and peritumoral regions using PyRadiomics (version 3.0.1) on the Deepwise multimodal research platform, in accordance with the Image Biomarker Standardization Initiative (IBSI) guidelines. Extracted features included first-order statistics, shape features, and texture features derived from gray-level co-occurrence, run-length, size-zone, and dependence matrices. Wavelet and Laplacian-of-Gaussian filters were applied to generate higher-order features. To ensure feature robustness and reduce redundancy, a three-step feature selection strategy was applied. First, intra- and inter-reader reproducibility was assessed using intraclass correlation coefficients (ICC) based on repeated segmentations in 20% of randomly selected cases, and features with ICC 0.90). Finally, least absolute shrinkage and selection operator (LASSO) regression was used to identify the most predictive features for model construction. Radiomics Model Development and Validation Radiomics models were developed using logistic regression and support vector classifier algorithms. The dataset was randomly divided into a training cohort (70%) and a validation cohort (30%). Five-fold cross-validation was performed within the training cohort. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). H&E Staining and Slide Digitization Formalin-fixed, paraffin-embedded tumor tissue samples were sectioned at 5 µm thickness and stained with hematoxylin and eosin (H&E). All slides were digitized using a high-resolution whole-slide scanner (OLYMPUS VS200) at 20× magnification, with a spatial resolution of 0.46 µm per pixel. Whole-slide images (WSIs) were stored in multi-resolution pyramid TIFF format. Pathomics Image Preprocessing Whole-slide images were preprocessed using a standardized pipeline, including image tiling, background removal, and stain normalization. Each WSI was divided into non-overlapping patches of 512 × 512 pixels. Background regions were removed using RGB thresholding (threshold = 216). To reduce staining variability, color normalization was performed using the Vahadane method. Pathomics Feature Extraction and Model Development A weakly supervised multiple instance learning (MIL) framework was adopted for pathomics analysis, in which patch-level features were extracted using deep learning architectures, including Vision Transformer (ViT), DenseNet, and ResNet. Patch-level features were aggregated to generate slide-level representations without pixel-level annotations. For quantitative pathomics feature encoding, histogram-based features and term frequency–inverse document frequency (TF-IDF) representations were computed and used for subsequent model construction. Deep Learning Visual Analytics Gradient-weighted class activation mapping (Grad-CAM) was applied to visualize model attention regions and enhance the interpretability of the pathomics deep learning models. Multimodal Model Construction A late fusion strategy was used to integrate statistically significant clinical variables, intratumoral and peritumoral radiomic signatures, and pathomics model outputs to construct a multimodal predictive model for postoperative local recurrence. Statistical analysis Statistical analyses were performed using SPSS software (PASW Statistics 18.0, SPSS Inc, Chicago, IL, USA). Continuous variables were compared using the independent-samples t-test or Mann–Whitney U test, as appropriate. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, with calculation of the area under the curve (AUC), sensitivity, and specificity. Kaplan–Meier survival analysis was performed to compare survival outcomes between groups. All statistical tests were two-sided, and p < 0.05 was considered statistically significant. Results Clinical and Laboratory Characteristics A total of 135 patients were included in this study, with a mean age of 62.65 ± 8.74 years. Among them, 30 patients experienced postoperative local recurrence, while 105 patients did not. The median disease-free survival (DFS) was 19 months (range, 6–34 months). Comparisons of baseline clinical and pathological characteristics between the local recurrence and non-local recurrence groups are summarized in Table 1 . Smoking history was the only clinical variable that differed significantly between the two groups (p = 0.006). Table 1 Clinical and pathological characteristics of the enrolled patients Total n = 135 LR n = 30 non-LR n = 105 P value age 62.65 ± 8.74 61.93 ± 9.30 62.86 ± 8.61 0.611 sex female 25(18.52) 9(30.00) 16(15.24) 0.066 male 110(81.48) 21(70.00) 89(84.76) Smoking status 0.006* Never smoked 40(29.63) 15(50.00) 25(23.81) Ex-smoker or current smoker 95(70.37) 15(50.00) 80(76.19) Pathological T stage 0.521 T3 92(68.15) 19(63.33) 73(69.52) T4 43(31.85) 11(36.67) 32(30.48) Pathological N stage 0.865 N0 80(59.26) 19(63.33) 61(58.10) N1 29(21.48) 7(23.33) 22(20.95) N2 26(19.26) 4(13.33) 22(20.95) Pathological TNM stage 0.874 II 47(34.81) 10(33.33) 37(35.24) III 88(65.19) 20(66.67) 68(64.76) Histology 0.108 SCC 83(61.48) 14(46.67) 69(65.71) ADC 41(30.37) 13(43.33) 28(26.67) Others 11(8.15) 3(10.00) 8(7.62) Surgical procedure 0.973 Lobectomy 92(68.14) 20(66.67) 72(68.57) Bilobectomy 28(20.74) 7(23.33) 21(20.00) Pneumonectomy 15(11.11) 3(10.00) 12(11.43) * p <0.05 The results of preoperative laboratory tests and tumor markers are presented in Table 2 . Among the evaluated markers, squamous cell carcinoma antigen (SCCA) levels were significantly higher in the local recurrence group compared with the non-local recurrence group (p = 0.01). Table 2 The results of laboratory tests were compared between the orthotopic recurrence group and the non-orthotopic recurrence group Total n = 135 LR n = 30 Non-LR n = 105 P value N% 64.93 ± 11.79 66.70 ± 11.21 64.42 ± 11.96 0.352 L% 24.31 ± 10.16 23.50 ± 10.08 24.55 ± 10.22 0.621 CEA 5.84 ± 9.24 4.51 ± 5.32 6.22 ± 10.08 0.374 SCCA 1.44(0.84, 2.85) 0.52(0.30, 1.00) 1.74(0.90, 2.80) 0.01* NSE 14.08 ± 7.31 12.97 ± 4.58 14.40 ± 7.90 0.346 CA199 19.62 ± 36.63 14.11 ± 23.05 21.20 ± 39.62 0.352 CA125 33.85 ± 50.56 29.05 ± 21.93 35.22 ± 56.14 0.558 Performance of Radiomics Models Intratumoral and peritumoral radiomics models were developed to predict postoperative local recurrence. The intratumoral radiomics model demonstrated moderate predictive performance for postoperative local recurrence, whereas the peritumoral model showed slightly inferior performance. A combined radiomics model integrating intratumoral and peritumoral features achieved improved predictive performance compared with either model alone. Detailed performance metrics of the intratumoral, peritumoral, and combined radiomics models in the internal validation cohort are summarized in Table 3 . Pairwise comparisons of radiomics models using the DeLong test are presented in Supplementary Table S2. Table 3 The efficacy of three radiomics models in the internal validation set AUC Sensitivity Specificity Accuracy Tumor radiomic model 0.71 0.63 0.67 0.66 Peritumor radiomic model 0.79 0.77 0.67 0.69 Combined radiomic model 0.77 0.77 0.65 0.67 Performance of Pathomics Models Pathomics models based on three deep learning architectures—Vision Transformer (ViT), DenseNet, and ResNet—were constructed and evaluated. The predictive performance of these models in the internal validation cohort is summarized in Table 4 . Table 4 The efficacy of the pathomics models of three deep learning methods in the internal validation set AUC Sensitivity Specificity Accuracy Vit 0.74 0.66 0.69 0.33 DenseNet 0.92 0.78 0.90 0.88 ResNet 0.89 0.89 0.94 0.93 Among the evaluated architectures, the DenseNet-based pathomics model achieved the highest predictive performance for postoperative local recurrence. Performance of Multimodal Models Multimodal models integrating clinical variables, radiomic features, and pathomics model outputs were constructed to predict postoperative local recurrence. As shown in Table 5 , the multimodal fusion model demonstrated superior predictive performance compared with the clinical model and radiomics-only models. DeLong test analysis confirmed that the multimodal model achieved statistically significant improvement over the intratumoral radiomics model (p = 0.03). Detailed DeLong test results comparing the image fusion model, pathomics model, and multimodal model are summarized in Supplementary Table S3. Table 5 The efficacy of image fusion models, pathomics models and pathology-image fusion models in the internal validation set AUC Sensitivity Specificity Accuracy Combined radiomic model 0.77 0.77 0.65 0.67 Pathogenomics model 0.92 0.78 0.90 0.88 Multimodal model 0.92 0.78 0.94 0.90 Visualization of Pathomics Deep Learning Models Representative visualization results generated using gradient-weighted class activation mapping (Grad-CAM) are shown in Fig. 3. Regions with higher model attention were predominantly located in the nuclei of tumor cells, suggesting that nuclear morphology contributed substantially to the predictions of the pathomics model. Discussion In this study, we developed and validated a multimodal artificial intelligence model integrating clinical variables, preoperative CT-based radiomics, and histopathology-derived deep learning features to predict postoperative local recurrence in patients with surgically resected T3–4 NSCLC. The proposed multimodal model demonstrated superior predictive performance, with an AUC of 0.92, compared with radiomics-only models, highlighting the added value of incorporating histopathological information for individualized recurrence risk stratification. Postoperative local recurrence in locally advanced NSCLC is closely associated with aggressive tumor biology and pronounced intratumoral heterogeneity, which may not be accurately captured by radiological imaging alone. CT-based radiomics primarily reflects macroscopic tumor characteristics and peritumoral changes related to invasion patterns, vascular remodeling, and host–tumor interactions. However, its ability to characterize microscopic tumor behavior is inherently limited. In contrast, histopathological whole-slide images provide direct visualization of cellular morphology, nuclear atypia, stromal composition, and tumor–stroma interactions—features that are closely linked to tumor aggressiveness and local regrowth following surgical resection. The superior performance of the pathomics model observed in this study likely reflects its capacity to capture these fine-grained biological characteristics. The biological relevance of the pathomics model is further supported by the interpretability provided by Grad-CAM analysis, which revealed that the deep learning network predominantly focused on tumor cell nuclei and densely cellular regions. These regions are known to correlate with proliferative activity[ 11 ], nuclear pleomorphism, and genomic instability, all of which have been consistently associated with poor prognosis and increased recurrence risk in NSCLC. Importantly, the use of weakly supervised learning enabled efficient extraction of biologically meaningful patterns from large-scale histopathological data without the need for labor-intensive pixel-level annotations. Beyond single-modality modeling, the integration of radiomics and pathomics further improved predictive performance, indicating that macroscopic imaging features and microscopic pathological features provide complementary and non-redundant information[ 12 ]. Radiomics capture spatial heterogeneity at the tumor and peritumoral level, whereas pathomics characterize cellular-level heterogeneity and microenvironmental architecture. By integrating these modalities, the multimodal model provides a more comprehensive representation of tumor biology across multiple spatial scales, thereby enabling more accurate and robust prediction of postoperative local recurrence. From a clinical perspective, accurate identification of patients at high risk of postoperative local recurrence is particularly relevant in T3–4 NSCLC, a population for which postoperative management strategies remain heterogeneous and optimal treatment pathways are not well defined. The proposed multimodal model may assist clinicians in identifying patients who could benefit from intensified adjuvant therapy, closer imaging surveillance, or enrollment in clinical trials evaluating novel perioperative treatment strategies. As an objective and reproducible tool, this approach has the potential to support individualized decision-making in locally advanced NSCLC. Several limitations of this study should be acknowledged. First, the retrospective, single-center design may introduce selection bias and limit generalizability. Second, the relatively modest sample size underscores the need for external validation in larger, multicenter cohorts. Third, although the multimodal model demonstrated strong predictive performance, the biological mechanisms underlying specific radiomic and pathomic features were not completely elucidated. Future studies incorporating spatial transcriptomics, immunohistochemical, genetical and epigenetical analyses may help clarify the pathophysiological basis of recurrence-associated features. Prospective validation and integration of this model into clinical workflows will be essential steps toward real-world clinical application. Conclusion In conclusion, this study demonstrates that a multimodal artificial intelligence model integrating clinical characteristics, preoperative CT-based radiomic features, and pathology-derived deep learning features can better predict postoperative local recurrence in patients with surgically resected T3–4 NSCLC. Compared with radiomics-based models alone, the integration of pathomics significantly improved predictive performance, highlighting the complementary value of microscopic tumor heterogeneity captured from histopathological images. This multimodal approach enables more precise postoperative risk stratification and may be benificial to individualized treatment and surveillance strategies in locally advanced NSCLC. Abbreviations AJCC: American Joint Committee on Cancer AUC: Area under the curve CEA: Carcinoembryonic antigen CT: Computed tomography CTE: Contrast-enhanced computed tomography DFS: Disease-free survival Grad-CAM: Gradient-weighted class activation mapping H&E: Hematoxylin and eosin ICC: Intraclass correlation coefficient LASSO: Least absolute shrinkage and selection operator MDT: Multidisciplinary team NSCLC: Non-small cell lung cancer ROC: Receiver operating characteristic SCCA: Squamous cell carcinoma antigen ViT: Vision Transformer WSI: Whole-slide image Declarations • Ethics approval and consent to participate The review boards of Jinling hospital approved this retrospective study (2023DZKY-089-01). Informed consent was waived because of the retrospective nature of the study. • Consent for publication • Availability of data and materials Data generated or analyzed during this study are available from the corresponding author on reasonable request. • Competing interests All authors have no conflicts of interest. • Funding This study has received funding by the Science and Technology Innovation 2030-Major Projects (2020AAA0109500) • Authors' contributions Study concept or design: Xinyu Li, Guangming Lu; Data collection: Changsheng Zhou, Yu Zong; Formal analysis: Xinyu Li, Liying Wang; Investigation: Jianrui Li; Zhen Zhou; Methodology: Xinyu Li; Xiaoqing Cheng; Project administration: Guangming Lu; Writing – original draft: Xinyu Li; Writing – review & editing: Liying Wang and Guangming Lu • Acknowledgements The authors would like to thank the radiologists and pathologists at Jinling Hospital for their assistance in image acquisition and pathological slide preparation. We also acknowledge the technical support provided by the Deepwise multimodal research platform. And we thank the OnekeyAI company for the code consultation of this study. References Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F: Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021, 71(3):209-249. Boyd JA, Hubbs JL, Kim DW, Hollis D, Marks LB, Kelsey CR: Timing of local and distant failure in resected lung cancer: implications for reported rates of local failure. J Thorac Oncol 2010, 5(2):211-214. Siegel RL, Miller KD, Fuchs HE, Jemal A: Cancer statistics, 2022. CA Cancer J Clin 2022, 72(1):7-33. Huang Y, Liu Z, He L, Chen X, Pan D, Ma Z, Liang C, Tian J, Liang C: Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016, 281(3):947-957. Yang H, Wang L, Shao G, Dong B, Wang F, Wei Y, Li P, Chen H, Chen W, Zheng Y et al: A combined predictive model based on radiomics features and clinical factors for disease progression in early-stage non-small cell lung cancer treated with stereotactic ablative radiotherapy. Front Oncol 2022, 12:967360. She Y, He B, Wang F, Zhong Y, Wang T, Liu Z, Yang M, Yu B, Deng J, Sun X et al: Deep learning for predicting major pathological response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer: A multicentre study. EBioMedicine 2022, 86:104364. Wang S, Yu H, Gan Y, Wu Z, Li E, Li X, Cao J, Zhu Y, Wang L, Deng H et al: Mining whole-lung information by artificial intelligence for predicting EGFR genotype and targeted therapy response in lung cancer: a multicohort study. Lancet Digit Health 2022, 4(5):e309-e319. Sun R, Limkin EJ, Vakalopoulou M, Dercle L, Champiat S, Han SR, Verlingue L, Brandao D, Lancia A, Ammari S et al: A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. The Lancet Oncology 2018, 19(9):1180-1191. Lu C, Shiradkar R, Liu Z: Integrating pathomics with radiomics and genomics for cancer prognosis: A brief review. Chin J Cancer Res 2021, 33(5):563-573. Vaidya P, Khorrami M, Bera K, Fu P, Delasos L, Gupta A, Barrera C, Pennell NA, Velcheti V, Madabhushi A: Computationally integrating radiology and pathology image features for predicting treatment benefit and outcome in lung cancer. NPJ Precis Oncol 2025, 9(1):161. Saltz J, Gupta R, Hou L, Kurc T, Singh P, Nguyen V, Samaras D, Shroyer KR, Zhao T, Batiste R et al: Spatial Organization and Molecular Correlation of Tumor-Infiltrating Lymphocytes Using Deep Learning on Pathology Images. Cell Rep 2018, 23(1):181-193.e187. Li Z, Jiang Y, Lu M, Li R, Xia Y: Survival Prediction via Hierarchical Multimodal Co-Attention Transformer: A Computational Histology-Radiology Solution. IEEE Trans Med Imaging 2023, 42(9):2678-2689. Additional Declarations No competing interests reported. Supplementary Files Supplementary.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 22 Apr, 2026 Editor invited by journal 20 Apr, 2026 Editor assigned by journal 11 Feb, 2026 Submission checks completed at journal 11 Feb, 2026 First submitted to journal 01 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8760093","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":627857850,"identity":"fa1d9dd9-580c-4d3e-ac07-726d748dc0a0","order_by":0,"name":"Xinyu Li","email":"","orcid":"","institution":"Department of Radiology, Jinling Hospital, School of Medical Imaging, Nanjing Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xinyu","middleName":"","lastName":"Li","suffix":""},{"id":627857851,"identity":"695a1647-882a-438c-a982-405a87c8c9ec","order_by":1,"name":"Liying Wang","email":"","orcid":"","institution":"The Affiliated Changsha Central Hospital, Hengyang Medical School, University of South China","correspondingAuthor":false,"prefix":"","firstName":"Liying","middleName":"","lastName":"Wang","suffix":""},{"id":627857852,"identity":"90c84c74-9d28-47ea-bc92-66f747e496a2","order_by":2,"name":"Yu Zong","email":"","orcid":"","institution":"Department of Radiology, Jinling Clinical Medical College, Nanjing Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Zong","suffix":""},{"id":627857853,"identity":"843fd337-df92-46f2-92ae-318a5e6cc96e","order_by":3,"name":"Changsheng Zhou","email":"","orcid":"","institution":"Department of Radiology, Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University","correspondingAuthor":false,"prefix":"","firstName":"Changsheng","middleName":"","lastName":"Zhou","suffix":""},{"id":627857854,"identity":"5c315620-5043-4653-834b-04467252c6c3","order_by":4,"name":"Zhen Zhou","email":"","orcid":"","institution":"Deepwise AI Lab, Deepwise Inc","correspondingAuthor":false,"prefix":"","firstName":"Zhen","middleName":"","lastName":"Zhou","suffix":""},{"id":627857855,"identity":"16cb15ed-a1cf-4443-85b6-be41a962dab4","order_by":5,"name":"Jianrui Li","email":"","orcid":"","institution":"Department of Radiology, Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University","correspondingAuthor":false,"prefix":"","firstName":"Jianrui","middleName":"","lastName":"Li","suffix":""},{"id":627857856,"identity":"be23941b-79dc-4d01-88de-ba2f36604294","order_by":6,"name":"Xiaoqing Cheng","email":"","orcid":"","institution":"Department of Radiology, Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoqing","middleName":"","lastName":"Cheng","suffix":""},{"id":627857857,"identity":"b43a08d5-ab31-47e8-8882-58ffef4a4270","order_by":7,"name":"Guangming Lu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABFklEQVRIie2RMUvEMBTH3xGoy3lZK+j5FSr3hZLFLtf1uKFXKko7iNx6IuhXiNuNkYN0Sd2ECA7p4uzobb4eFW5oe6tgfpDHn5Bf3gsBcDj+JBLA/mYGMAZSh3ncr7A9ZbJTrFb9jdhe5imWQXVDOk/Tq9ICzxZ8+ZC/iWq9CEVxLFH0gOa3rE3xZRGgUvDVh54ZrotIbEYM+PoUfF2KNiUAHJpniqdmemkwRPfXQ7xEexD4Ub/y1CjhyU7JyCEl5sKECpWYUXJA8aUCyV7l5NlMPVTkxZKMmGRaDbveQldqYL9mydmjCT/ft1ly7tHypfqex2Oa37Uq2Kb+zA0mnKcJ9Q7WLmha1wTXkW1C92GHw+H4n/wAvLdwZb6beZMAAAAASUVORK5CYII=","orcid":"","institution":"Department of Radiology, Jinling Hospital, School of Medical Imaging, Nanjing Medical University","correspondingAuthor":true,"prefix":"","firstName":"Guangming","middleName":"","lastName":"Lu","suffix":""}],"badges":[],"createdAt":"2026-02-02 03:39:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8760093/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8760093/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108408618,"identity":"611de4aa-18e7-41ca-9c7e-eea15633daf8","added_by":"auto","created_at":"2026-05-04 09:55:56","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":80309,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of multimodal model inclusion and exclusion criteria.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8760093/v1/2a951e502ccea2790bde1bcc.jpg"},{"id":108408667,"identity":"bb5148c4-2e49-4117-989e-08e1213610a5","added_by":"auto","created_at":"2026-05-04 09:55:56","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":332242,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8760093/v1/af119195b7e8f0317564b2a3.png"},{"id":108492807,"identity":"f3795cf5-dcc1-4307-93ea-95b962fd1c91","added_by":"auto","created_at":"2026-05-05 09:58:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":764776,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8760093/v1/6ba93a83-96a0-4a94-af44-9824f25ca0ec.pdf"},{"id":108408797,"identity":"b650718e-d236-4019-9e4a-55cd63b3e48a","added_by":"auto","created_at":"2026-05-04 09:56:10","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":19930,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-8760093/v1/14cfdbcbbdda5459f1f48f1f.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Multimodal Radiomics and Pathomics Model for Predicting Postoperative Local Recurrence in T3–4 Non-Small Cell Lung Cancer","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLung cancer is the leading cause of cancer-related mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for approximately 85% of all cases[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. For patients with resectable T3-4 NSCLC evaluated by a multidisciplinary team (MDT), radical surgical resection followed by platinum-based adjuvant chemotherapy remains the standard treatment. Despite advances in surgical techniques and systemic therapy, postoperative local recurrence continues to be a major determinant of long-term prognosis in this population. Previous studies have reported recurrence rates ranging from 33% to 50%[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] after radical resection in locally advanced NSCLC, with a five-year survival rate of only 33%[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAccurately identifying patients at high risk of postoperative local recurrence is still a critical clinical challenge in the postoperative management of locally advanced NSCLC. Conventional clinicopathological factors are currently the primary means for assessing the risk of postoperative local recurrence. These factors include tumor stage, lymph node involvement, histological subtype, margin status, and the presence of visceral pleural or vascular invasion. However, such factors are largely categorical and population-based, with limited ability to capture the continuous and heterogeneous biological behavior of individual tumors. Consequently, patients with similar clinicopathological profiles may experience markedly different risks of postoperative local recurrence, underscoring the limitations of existing risk stratification approaches for precise, individualized prediction. Therefore, it is crucial to explore more accurate prognostic biomarkers in surgically resected locally advanced NSCLC.\u003c/p\u003e \u003cp\u003eRadiomics enables high-throughput extraction of quantitative features from medical images and has emerged as a noninvasive approach for characterizing tumor heterogeneity. Previous studies have demonstrated its value in predicting treatment response and prognosis in NSCLC across various therapeutic settings[\u003cspan additionalcitationids=\"CR5 CR6 CR7\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. However, in surgically resected locally advanced NSCLC, particularly in T3\u0026ndash;4 disease, the application of radiomics remains limited, and most existing models rely solely on imaging-derived features, which may be insufficient to capture the underlying biological complexity of tumors.\u003c/p\u003e \u003cp\u003eHistopathological whole-slide images provide complementary information at the cellular and microenvironmental levels, capturing microscopic tumor characteristics closely associated with aggressiveness and recurrence risks. Pathomics, which applies computational and deep learning techniques to digital pathology images, enables quantitative characterization of these features but has not been well explored for predicting postoperative local recurrence in locally advanced NSCLC[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMultimodal artificial intelligence offers a powerful framework to integrate radiological and pathological information across different spatial scales, potentially providing a more comprehensive representation of tumor biology than single-modality models[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Nevertheless, the clinical value of integrating radiomics and pathomics for predicting postoperative local recurrence in surgically resected T3\u0026ndash;4 NSCLC has not been well established.\u003c/p\u003e \u003cp\u003eTherefore, the aim of this study was to develop and validate a multimodal artificial intelligence model integrating conventional clinical variables, preoperative contrast-enhanced CT\u0026ndash;based intratumoral and peritumoral radiomic features, and postoperative histopathology-derived deep learning features to predict postoperative local recurrence in patients with T3\u0026ndash;4 NSCLC. We further compared the predictive performance of radiomics-, pathomics-, and multimodal-based models to evaluate the incremental value of multimodal integration.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eStudy Population\u003c/p\u003e \u003cp\u003eThis retrospective study consecutively enrolled patients with pathologically confirmed T3\u0026ndash;4 NSCLC who underwent complete surgical resection at Jinling Hospital between January 2013 and December 2020. Tumor staging was determined according to the 8th edition of the American Joint Committee on Cancer (AJCC) staging system.\u003c/p\u003e \u003cp\u003eInclusion criteria were as follows: (1) availability of preoperative contrast-enhanced chest CT performed within two weeks before surgery; and (2) pathologically confirmed T3 or T4 NSCLC. Exclusion criteria included inadequate CT image quality due to motion artifacts and the absence of reliable follow-up information regarding recurrence status and timing.\u003c/p\u003e \u003cp\u003eFigure 1 Flowchart of multimodal model inclusion and exclusion criteria.\u003c/p\u003e \u003cp\u003eCT Image Acquisition\u003c/p\u003e \u003cp\u003ePreoperative contrast-enhanced CT images were acquired using multidetector CT scanners according to institutional protocols. Non-contrast, arterial, and portal venous phase images with thin-slice reconstruction were obtained. Arterial-phase images were used for subsequent segmentation and radiomic feature extraction. Detailed CT acquisition and reconstruction parameters for different scanner models are provided in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eCollection of Clinical, Pathological, and Laboratory Features\u003c/p\u003e \u003cp\u003ePreoperative clinical and laboratory variables were retrospectively collected, including neutrophil and lymphocyte percentages (N%, L%), carcinoembryonic antigen (CEA), squamous cell carcinoma antigen (SCCA), neuron-specific enolase (NSE), cytokeratin 19 fragment (CYFRA 21\u0026thinsp;\u0026minus;\u0026thinsp;1), carbohydrate antigen 724 (CA724), carbohydrate antigen 242 (CA242), carbohydrate antigen 199 (CA199), and carbohydrate antigen 125 (CA125), based on blood tests performed within one week before surgery.\u003c/p\u003e \u003cp\u003eVariables with more than 20% missing values (CYFRA 21\u0026thinsp;\u0026minus;\u0026thinsp;1, CA724, and CA242) were excluded from analysis. Missing values for the remaining variables were imputed using mean imputation.\u003c/p\u003e \u003cp\u003eFollow-Up Endpoints and Individualized Labeling\u003c/p\u003e \u003cp\u003eThe primary endpoint of this study was postoperative local recurrence, defined as tumor recurrence at the surgical margin or within the ipsilateral hemithorax after R0 resection during follow-up. Patients without local recurrence, including those with distant metastasis or no evidence of disease, were classified as the non-local recurrence group.\u003c/p\u003e \u003cp\u003eDisease-free survival (DFS) was defined as the interval from surgery to recurrence, metastasis, death, or last follow-up and was used as a secondary outcome. Follow-up evaluations included serum tests and chest CT every 3\u0026ndash;6 months during the first two years after surgery and every 6 months thereafter until death or December 31, 2023.\u003c/p\u003e \u003cp\u003eTumor and Peritumoral Tissue Segmentation\u003c/p\u003e \u003cp\u003eAll images were in DICOM format. We used arterial-phase mediastinal window images for segmentation and feature extraction. After anonymization, thin-slice DICOM images were imported into the Deepwise multimodal research platform (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://keyan.deepwise.com\u003c/span\u003e\u003cspan address=\"https://keyan.deepwise.com\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), an integrated machine learning platform incorporating PyRadiomics (version 3.0.1) and scikit-learn (version 0.22). Semi-automatic segmentation was performed: AI automatically segmented pulmonary lesions, followed by blinded review and necessary manual correction by radiologists, who were blinded to clinical outcomes. Reader 1 (with six years of thoracic imaging experience) independently segmented all lesions. A random selection of 20% of cases was segmented again by Reader 1 after one month and by Reader 2 (with three years of experience). Peritumoral tissue segmentation was achieved by expanding the tumor segmentation outward by 3 mm using platform software to form a ring-shaped region.\u003c/p\u003e \u003cp\u003eRadiomics Feature Extraction and Selection\u003c/p\u003e \u003cp\u003e Radiomic features were extracted from intratumoral and peritumoral regions using PyRadiomics (version 3.0.1) on the Deepwise multimodal research platform, in accordance with the Image Biomarker Standardization Initiative (IBSI) guidelines. Extracted features included first-order statistics, shape features, and texture features derived from gray-level co-occurrence, run-length, size-zone, and dependence matrices. Wavelet and Laplacian-of-Gaussian filters were applied to generate higher-order features.\u003c/p\u003e \u003cp\u003eTo ensure feature robustness and reduce redundancy, a three-step feature selection strategy was applied. First, intra- and inter-reader reproducibility was assessed using intraclass correlation coefficients (ICC) based on repeated segmentations in 20% of randomly selected cases, and features with ICC\u0026thinsp;\u0026lt;\u0026thinsp;0.80 were excluded. Second, highly correlated features were removed using Pearson correlation analysis (correlation coefficient\u0026thinsp;\u0026gt;\u0026thinsp;0.90). Finally, least absolute shrinkage and selection operator (LASSO) regression was used to identify the most predictive features for model construction.\u003c/p\u003e \u003cp\u003eRadiomics Model Development and Validation\u003c/p\u003e \u003cp\u003eRadiomics models were developed using logistic regression and support vector classifier algorithms. The dataset was randomly divided into a training cohort (70%) and a validation cohort (30%). Five-fold cross-validation was performed within the training cohort. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC).\u003c/p\u003e \u003cp\u003eH\u0026amp;E Staining and Slide Digitization\u003c/p\u003e \u003cp\u003eFormalin-fixed, paraffin-embedded tumor tissue samples were sectioned at 5 \u0026micro;m thickness and stained with hematoxylin and eosin (H\u0026amp;E). All slides were digitized using a high-resolution whole-slide scanner (OLYMPUS VS200) at 20\u0026times; magnification, with a spatial resolution of 0.46 \u0026micro;m per pixel. Whole-slide images (WSIs) were stored in multi-resolution pyramid TIFF format.\u003c/p\u003e \u003cp\u003ePathomics Image Preprocessing\u003c/p\u003e \u003cp\u003eWhole-slide images were preprocessed using a standardized pipeline, including image tiling, background removal, and stain normalization. Each WSI was divided into non-overlapping patches of 512 \u0026times; 512 pixels. Background regions were removed using RGB thresholding (threshold\u0026thinsp;=\u0026thinsp;216). To reduce staining variability, color normalization was performed using the Vahadane method.\u003c/p\u003e \u003cp\u003ePathomics Feature Extraction and Model Development\u003c/p\u003e \u003cp\u003eA weakly supervised multiple instance learning (MIL) framework was adopted for pathomics analysis, in which patch-level features were extracted using deep learning architectures, including Vision Transformer (ViT), DenseNet, and ResNet. Patch-level features were aggregated to generate slide-level representations without pixel-level annotations.\u003c/p\u003e \u003cp\u003eFor quantitative pathomics feature encoding, histogram-based features and term frequency\u0026ndash;inverse document frequency (TF-IDF) representations were computed and used for subsequent model construction.\u003c/p\u003e \u003cp\u003eDeep Learning Visual Analytics\u003c/p\u003e \u003cp\u003eGradient-weighted class activation mapping (Grad-CAM) was applied to visualize model attention regions and enhance the interpretability of the pathomics deep learning models.\u003c/p\u003e \u003cp\u003eMultimodal Model Construction\u003c/p\u003e \u003cp\u003eA late fusion strategy was used to integrate statistically significant clinical variables, intratumoral and peritumoral radiomic signatures, and pathomics model outputs to construct a multimodal predictive model for postoperative local recurrence.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eStatistical analyses were performed using SPSS software (PASW Statistics 18.0, SPSS Inc, Chicago, IL, USA). Continuous variables were compared using the independent-samples t-test or Mann\u0026ndash;Whitney U test, as appropriate. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis, with calculation of the area under the curve (AUC), sensitivity, and specificity. Kaplan\u0026ndash;Meier survival analysis was performed to compare survival outcomes between groups. All statistical tests were two-sided, and p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eClinical and Laboratory Characteristics\u003c/p\u003e \u003cp\u003eA total of 135 patients were included in this study, with a mean age of 62.65\u0026thinsp;\u0026plusmn;\u0026thinsp;8.74 years. Among them, 30 patients experienced postoperative local recurrence, while 105 patients did not. The median disease-free survival (DFS) was 19 months (range, 6\u0026ndash;34 months).\u003c/p\u003e \u003cp\u003eComparisons of baseline clinical and pathological characteristics between the local recurrence and non-local recurrence groups are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Smoking history was the only clinical variable that differed significantly between the two groups (p\u0026thinsp;=\u0026thinsp;0.006).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClinical and pathological characteristics of the enrolled patients\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal n\u0026thinsp;=\u0026thinsp;135\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLR n\u0026thinsp;=\u0026thinsp;30\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003enon-LR n\u0026thinsp;=\u0026thinsp;105\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e62.65\u0026thinsp;\u0026plusmn;\u0026thinsp;8.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e61.93\u0026thinsp;\u0026plusmn;\u0026thinsp;9.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e62.86\u0026thinsp;\u0026plusmn;\u0026thinsp;8.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.611\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esex\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003efemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25(18.52)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9(30.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16(15.24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.066\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e110(81.48)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e21(70.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89(84.76)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSmoking status\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.006*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNever smoked\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e40(29.63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15(50.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e25(23.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEx-smoker or current smoker\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e95(70.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15(50.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e80(76.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePathological T stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.521\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e92(68.15)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19(63.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e73(69.52)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e43(31.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11(36.67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e32(30.48)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePathological N stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.865\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e80(59.26)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19(63.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e61(58.10)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29(21.48)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7(23.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e22(20.95)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e26(19.26)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4(13.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e22(20.95)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePathological TNM stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.874\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e47(34.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10(33.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e37(35.24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e88(65.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20(66.67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e68(64.76)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHistology\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.108\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e83(61.48)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14(46.67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e69(65.71)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eADC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e41(30.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13(43.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e28(26.67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOthers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11(8.15)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3(10.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8(7.62)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSurgical procedure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLobectomy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e92(68.14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20(66.67)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e72(68.57)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBilobectomy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28(20.74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7(23.33)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21(20.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePneumonectomy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15(11.11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3(10.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12(11.43)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e*\u003cem\u003ep\u003c/em\u003e \u0026lt;0.05\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe results of preoperative laboratory tests and tumor markers are presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Among the evaluated markers, squamous cell carcinoma antigen (SCCA) levels were significantly higher in the local recurrence group compared with the non-local recurrence group (p\u0026thinsp;=\u0026thinsp;0.01).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe results of laboratory tests were compared between the orthotopic recurrence group and the non-orthotopic recurrence group\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003cp\u003en\u0026thinsp;=\u0026thinsp;135\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLR\u003c/p\u003e \u003cp\u003en\u0026thinsp;=\u0026thinsp;30\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNon-LR\u003c/p\u003e \u003cp\u003en\u0026thinsp;=\u0026thinsp;105\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e64.93\u0026thinsp;\u0026plusmn;\u0026thinsp;11.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e66.70\u0026thinsp;\u0026plusmn;\u0026thinsp;11.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e64.42\u0026thinsp;\u0026plusmn;\u0026thinsp;11.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.352\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eL%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24.31\u0026thinsp;\u0026plusmn;\u0026thinsp;10.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23.50\u0026thinsp;\u0026plusmn;\u0026thinsp;10.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.55\u0026thinsp;\u0026plusmn;\u0026thinsp;10.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.621\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCEA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.84\u0026thinsp;\u0026plusmn;\u0026thinsp;9.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.51\u0026thinsp;\u0026plusmn;\u0026thinsp;5.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.22\u0026thinsp;\u0026plusmn;\u0026thinsp;10.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.374\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSCCA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.44(0.84, 2.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.52(0.30, 1.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.74(0.90, 2.80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.01*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.08\u0026thinsp;\u0026plusmn;\u0026thinsp;7.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.97\u0026thinsp;\u0026plusmn;\u0026thinsp;4.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.40\u0026thinsp;\u0026plusmn;\u0026thinsp;7.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.346\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA199\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e19.62\u0026thinsp;\u0026plusmn;\u0026thinsp;36.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.11\u0026thinsp;\u0026plusmn;\u0026thinsp;23.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21.20\u0026thinsp;\u0026plusmn;\u0026thinsp;39.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.352\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA125\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33.85\u0026thinsp;\u0026plusmn;\u0026thinsp;50.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29.05\u0026thinsp;\u0026plusmn;\u0026thinsp;21.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e35.22\u0026thinsp;\u0026plusmn;\u0026thinsp;56.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.558\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003ePerformance of Radiomics Models\u003c/p\u003e \u003cp\u003eIntratumoral and peritumoral radiomics models were developed to predict postoperative local recurrence. The intratumoral radiomics model demonstrated moderate predictive performance for postoperative local recurrence, whereas the peritumoral model showed slightly inferior performance.\u003c/p\u003e \u003cp\u003eA combined radiomics model integrating intratumoral and peritumoral features achieved improved predictive performance compared with either model alone. Detailed performance metrics of the intratumoral, peritumoral, and combined radiomics models in the internal validation cohort are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Pairwise comparisons of radiomics models using the DeLong test are presented in Supplementary Table S2.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe efficacy of three radiomics models in the internal validation set\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTumor radiomic model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeritumor radiomic model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCombined radiomic model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003ePerformance of Pathomics Models\u003c/p\u003e \u003cp\u003ePathomics models based on three deep learning architectures\u0026mdash;Vision Transformer (ViT), DenseNet, and ResNet\u0026mdash;were constructed and evaluated. The predictive performance of these models in the internal validation cohort is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe efficacy of the pathomics models of three deep learning methods in the internal validation set\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.33\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDenseNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e Among the evaluated architectures, the DenseNet-based pathomics model achieved the highest predictive performance for postoperative local recurrence.\u003c/p\u003e \u003cp\u003ePerformance of Multimodal Models\u003c/p\u003e \u003cp\u003eMultimodal models integrating clinical variables, radiomic features, and pathomics model outputs were constructed to predict postoperative local recurrence.\u003c/p\u003e \u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, the multimodal fusion model demonstrated superior predictive performance compared with the clinical model and radiomics-only models. DeLong test analysis confirmed that the multimodal model achieved statistically significant improvement over the intratumoral radiomics model (p\u0026thinsp;=\u0026thinsp;0.03). Detailed DeLong test results comparing the image fusion model, pathomics model, and multimodal model are summarized in Supplementary Table S3.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe efficacy of image fusion models, pathomics models and pathology-image fusion models in the internal validation set\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCombined radiomic model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePathogenomics model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMultimodal model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eVisualization of Pathomics Deep Learning Models\u003c/p\u003e \u003cp\u003eRepresentative visualization results generated using gradient-weighted class activation mapping (Grad-CAM) are shown in Fig.\u0026nbsp;3. Regions with higher model attention were predominantly located in the nuclei of tumor cells, suggesting that nuclear morphology contributed substantially to the predictions of the pathomics model.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed and validated a multimodal artificial intelligence model integrating clinical variables, preoperative CT-based radiomics, and histopathology-derived deep learning features to predict postoperative local recurrence in patients with surgically resected T3\u0026ndash;4 NSCLC. The proposed multimodal model demonstrated superior predictive performance, with an AUC of 0.92, compared with radiomics-only models, highlighting the added value of incorporating histopathological information for individualized recurrence risk stratification.\u003c/p\u003e \u003cp\u003ePostoperative local recurrence in locally advanced NSCLC is closely associated with aggressive tumor biology and pronounced intratumoral heterogeneity, which may not be accurately captured by radiological imaging alone. CT-based radiomics primarily reflects macroscopic tumor characteristics and peritumoral changes related to invasion patterns, vascular remodeling, and host\u0026ndash;tumor interactions. However, its ability to characterize microscopic tumor behavior is inherently limited. In contrast, histopathological whole-slide images provide direct visualization of cellular morphology, nuclear atypia, stromal composition, and tumor\u0026ndash;stroma interactions\u0026mdash;features that are closely linked to tumor aggressiveness and local regrowth following surgical resection. The superior performance of the pathomics model observed in this study likely reflects its capacity to capture these fine-grained biological characteristics.\u003c/p\u003e \u003cp\u003eThe biological relevance of the pathomics model is further supported by the interpretability provided by Grad-CAM analysis, which revealed that the deep learning network predominantly focused on tumor cell nuclei and densely cellular regions. These regions are known to correlate with proliferative activity[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], nuclear pleomorphism, and genomic instability, all of which have been consistently associated with poor prognosis and increased recurrence risk in NSCLC. Importantly, the use of weakly supervised learning enabled efficient extraction of biologically meaningful patterns from large-scale histopathological data without the need for labor-intensive pixel-level annotations.\u003c/p\u003e \u003cp\u003eBeyond single-modality modeling, the integration of radiomics and pathomics further improved predictive performance, indicating that macroscopic imaging features and microscopic pathological features provide complementary and non-redundant information[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Radiomics capture spatial heterogeneity at the tumor and peritumoral level, whereas pathomics characterize cellular-level heterogeneity and microenvironmental architecture. By integrating these modalities, the multimodal model provides a more comprehensive representation of tumor biology across multiple spatial scales, thereby enabling more accurate and robust prediction of postoperative local recurrence.\u003c/p\u003e \u003cp\u003eFrom a clinical perspective, accurate identification of patients at high risk of postoperative local recurrence is particularly relevant in T3\u0026ndash;4 NSCLC, a population for which postoperative management strategies remain heterogeneous and optimal treatment pathways are not well defined. The proposed multimodal model may assist clinicians in identifying patients who could benefit from intensified adjuvant therapy, closer imaging surveillance, or enrollment in clinical trials evaluating novel perioperative treatment strategies. As an objective and reproducible tool, this approach has the potential to support individualized decision-making in locally advanced NSCLC.\u003c/p\u003e \u003cp\u003eSeveral limitations of this study should be acknowledged. First, the retrospective, single-center design may introduce selection bias and limit generalizability. Second, the relatively modest sample size underscores the need for external validation in larger, multicenter cohorts. Third, although the multimodal model demonstrated strong predictive performance, the biological mechanisms underlying specific radiomic and pathomic features were not completely elucidated. Future studies incorporating spatial transcriptomics, immunohistochemical, genetical and epigenetical analyses may help clarify the pathophysiological basis of recurrence-associated features. Prospective validation and integration of this model into clinical workflows will be essential steps toward real-world clinical application.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, this study demonstrates that a multimodal artificial intelligence model integrating clinical characteristics, preoperative CT-based radiomic features, and pathology-derived deep learning features can better predict postoperative local recurrence in patients with surgically resected T3\u0026ndash;4 NSCLC.\u003c/p\u003e \u003cp\u003eCompared with radiomics-based models alone, the integration of pathomics significantly improved predictive performance, highlighting the complementary value of microscopic tumor heterogeneity captured from histopathological images. This multimodal approach enables more precise postoperative risk stratification and may be benificial to individualized treatment and surveillance strategies in locally advanced NSCLC.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAJCC: American Joint Committee on Cancer\u003c/p\u003e\n\u003cp\u003eAUC: Area under the curve\u003c/p\u003e\n\u003cp\u003eCEA: Carcinoembryonic antigen\u003c/p\u003e\n\u003cp\u003eCT: Computed tomography\u003c/p\u003e\n\u003cp\u003eCTE: Contrast-enhanced computed tomography\u003c/p\u003e\n\u003cp\u003eDFS: Disease-free survival\u003c/p\u003e\n\u003cp\u003eGrad-CAM: Gradient-weighted class activation mapping\u003c/p\u003e\n\u003cp\u003eH\u0026amp;E: Hematoxylin and eosin\u003c/p\u003e\n\u003cp\u003eICC: Intraclass correlation coefficient\u003c/p\u003e\n\u003cp\u003eLASSO: Least absolute shrinkage and selection operator\u003c/p\u003e\n\u003cp\u003eMDT: Multidisciplinary team\u003c/p\u003e\n\u003cp\u003eNSCLC: Non-small cell lung cancer\u003c/p\u003e\n\u003cp\u003eROC: Receiver operating characteristic\u003c/p\u003e\n\u003cp\u003eSCCA: Squamous cell carcinoma antigen\u003c/p\u003e\n\u003cp\u003eViT: Vision Transformer\u003c/p\u003e\n\u003cp\u003eWSI: Whole-slide image\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Ethics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThe review boards of Jinling hospital approved this retrospective study (2023DZKY-089-01). Informed consent was waived because of the retrospective nature of the study.\u003c/p\u003e\n\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Consent for publication\u003c/p\u003e\n\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Availability of data and materials\u003c/p\u003e\n\u003cp\u003eData generated or analyzed during this study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Competing interests\u003c/p\u003e\n\u003cp\u003eAll authors have no conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Funding\u003c/p\u003e\n\u003cp\u003eThis study has received funding by the Science and Technology Innovation 2030-Major Projects (2020AAA0109500)\u003c/p\u003e\n\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Authors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003eStudy concept or design: Xinyu Li, Guangming Lu;\u003c/p\u003e\n\u003cp\u003eData collection: Changsheng Zhou, Yu Zong;\u003c/p\u003e\n\u003cp\u003eFormal analysis: Xinyu Li, Liying Wang;\u003c/p\u003e\n\u003cp\u003eInvestigation: Jianrui Li; Zhen Zhou;\u003c/p\u003e\n\u003cp\u003eMethodology: Xinyu Li; Xiaoqing Cheng;\u003c/p\u003e\n\u003cp\u003eProject administration: Guangming Lu;\u003c/p\u003e\n\u003cp\u003eWriting \u0026ndash; original draft: Xinyu Li;\u003c/p\u003e\n\u003cp\u003eWriting \u0026ndash; review \u0026amp; editing: Liying Wang and Guangming Lu\u003c/p\u003e\n\u003cp\u003e\u0026bull;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Acknowledgements\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank the radiologists and pathologists at Jinling Hospital for their assistance in image acquisition and pathological slide preparation. We also acknowledge the technical support provided by the Deepwise multimodal research platform. And we thank the OnekeyAI company for the code consultation of this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F: Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021, 71(3):209-249.\u003c/li\u003e\n\u003cli\u003eBoyd JA, Hubbs JL, Kim DW, Hollis D, Marks LB, Kelsey CR: Timing of local and distant failure in resected lung cancer: implications for reported rates of local failure. J Thorac Oncol 2010, 5(2):211-214.\u003c/li\u003e\n\u003cli\u003eSiegel RL, Miller KD, Fuchs HE, Jemal A: Cancer statistics, 2022. CA Cancer J Clin 2022, 72(1):7-33.\u003c/li\u003e\n\u003cli\u003eHuang Y, Liu Z, He L, Chen X, Pan D, Ma Z, Liang C, Tian J, Liang C: Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016, 281(3):947-957.\u003c/li\u003e\n\u003cli\u003eYang H, Wang L, Shao G, Dong B, Wang F, Wei Y, Li P, Chen H, Chen W, Zheng Y et al: A combined predictive model based on radiomics features and clinical factors for disease progression in early-stage non-small cell lung cancer treated with stereotactic ablative radiotherapy. Front Oncol 2022, 12:967360.\u003c/li\u003e\n\u003cli\u003eShe Y, He B, Wang F, Zhong Y, Wang T, Liu Z, Yang M, Yu B, Deng J, Sun X et al: Deep learning for predicting major pathological response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer: A multicentre study. EBioMedicine 2022, 86:104364.\u003c/li\u003e\n\u003cli\u003eWang S, Yu H, Gan Y, Wu Z, Li E, Li X, Cao J, Zhu Y, Wang L, Deng H et al: Mining whole-lung information by artificial intelligence for predicting EGFR genotype and targeted therapy response in lung cancer: a multicohort study. Lancet Digit Health 2022, 4(5):e309-e319.\u003c/li\u003e\n\u003cli\u003eSun R, Limkin EJ, Vakalopoulou M, Dercle L, Champiat S, Han SR, Verlingue L, Brandao D, Lancia A, Ammari S et al: A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. The Lancet Oncology 2018, 19(9):1180-1191.\u003c/li\u003e\n\u003cli\u003eLu C, Shiradkar R, Liu Z: Integrating pathomics with radiomics and genomics for cancer prognosis: A brief review. Chin J Cancer Res 2021, 33(5):563-573.\u003c/li\u003e\n\u003cli\u003eVaidya P, Khorrami M, Bera K, Fu P, Delasos L, Gupta A, Barrera C, Pennell NA, Velcheti V, Madabhushi A: Computationally integrating radiology and pathology image features for predicting treatment benefit and outcome in lung cancer. NPJ Precis Oncol 2025, 9(1):161.\u003c/li\u003e\n\u003cli\u003eSaltz J, Gupta R, Hou L, Kurc T, Singh P, Nguyen V, Samaras D, Shroyer KR, Zhao T, Batiste R et al: Spatial Organization and Molecular Correlation of Tumor-Infiltrating Lymphocytes Using Deep Learning on Pathology Images. Cell Rep 2018, 23(1):181-193.e187.\u003c/li\u003e\n\u003cli\u003eLi Z, Jiang Y, Lu M, Li R, Xia Y: Survival Prediction via Hierarchical Multimodal Co-Attention Transformer: A Computational Histology-Radiology Solution. IEEE Trans Med Imaging 2023, 42(9):2678-2689.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"non-small cell lung cancer, radiomics, pathomics, multimodal imaging, local recurrence","lastPublishedDoi":"10.21203/rs.3.rs-8760093/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8760093/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBackground\u003c/p\u003e\n\u003cp\u003eTo develop and validate a multimodal artificial intelligence model integrating clinical, imaging, and pathological data to predict the risk of postoperative local recurrence in patients with T3-4 non-small cell lung cancer (NSCLC).\u003c/p\u003e\n\u003cp\u003eMethods\u003c/p\u003e\n\u003cp\u003eA total of 135 patients with pathologically confirmed T3–4 NSCLC who underwent complete surgical resection were retrospectively enrolled. Patients were stratified according to postoperative local recurrence status. Clinical variables, including preoperative contrast-enhanced CT images, and hematoxylin and eosin (H\u0026amp;E)–stained pathological slides were collected. Intratumoral and peritumoral radiomic features were extracted using PyRadiomics, while pathomics features were learned from whole-slide images using weakly supervised deep learning. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression. Radiomics, pathomics, and multimodal models were developed using a 70% training cohort with five-fold cross-validation and validated in the remaining 30%. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.\u003c/p\u003e\n\u003cp\u003eResults\u003c/p\u003e\n\u003cp\u003eAmong the 135 patients, 30 (22.2%) experienced postoperative local recurrence. The radiomics-based model achieved an AUC of 0.77 with a sensitivity of 0.77 and specificity of 0.65. The pathomics model alone demonstrated strong predictive performance (AUC = 0.92), while the integrated multimodal model achieved comparable discrimination and significantly outperformed the radiomics-only model (p \u0026lt; 0.05).\u003c/p\u003e\n\u003cp\u003eConclusions\u003c/p\u003e\n\u003cp\u003eA multimodal AI model integrating conventional clinical data, CT-based radiomics, and histopathology-derived deep learning features enables accurate prediction of postoperative local recurrence in patients with T3–4 NSCLC. This approach may facilitate postoperative risk stratification, supporting individualized treatment and surveillance strategies in locally advanced NSCLC.\u003c/p\u003e","manuscriptTitle":"A Multimodal Radiomics and Pathomics Model for Predicting Postoperative Local Recurrence in T3–4 Non-Small Cell Lung Cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-04 09:55:19","doi":"10.21203/rs.3.rs-8760093/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-22T14:33:33+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-20T04:56:02+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-11T10:56:31+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-11T10:51:50+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-02-02T03:28:25+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"dbcefbbb-7ab3-4c9b-a650-c28537b3c8f0","owner":[],"postedDate":"May 4th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T09:55:22+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-04 09:55:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8760093","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8760093","identity":"rs-8760093","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}