Prediction of Tumor Spread Through Air Spaces with an Automatic Segmentation Deep Learning Model in Peripheral Stage I Lung Adenocarcinoma

Prediction of Tumor Spread Through Air Spaces with an Automatic Segmentation Deep Learning Model in Peripheral Stage I Lung Adenocarcinoma | 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 Prediction of Tumor Spread Through Air Spaces with an Automatic Segmentation Deep Learning Model in Peripheral Stage I Lung Adenocarcinoma Cong Liu, Yu-feng Wang, Ping Gong, Xiu-Qing Xue, Hong-Ying Zhao, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4768392/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 08 Mar, 2025 Read the published version in Respiratory Research → Version 1 posted 9 You are reading this latest preprint version Abstract Purpose: To evaluate the clinical applicability of deep learning (DL) models based on automatic segmentation in preoperatively predicting tumor spread through air spaces (STAS) in peripheral stage I lung adenocarcinoma (LUAD). Methods: This retrospective study analyzed data from patients who underwent surgical treatment for lung tumors from January 2022 to December 2023. An external validation set was introduced to assess the model's generalizability. The study utilized conventional radiomic features and DL models for comparison. ROI segmentation was performed using the VNet architecture, and DL models were developed with transfer learning and optimization techniques. We assessed the diagnostic accuracy of our models via calibration curves, decision curve analysis, and ROC curves. Results: The DL model based on automatic segmentation achieved an AUC of 0.880 (95% CI 0.780-0.979), outperforming the conventional radiomics model with an AUC of 0.833 (95% CI 0.707-0.960). The DL model demonstrated superior performance in both internal validation and external testing cohorts. Calibration curves, decision curve analysis, and ROC curves confirmed the enhanced diagnostic accuracy and clinical utility of the DL approach. Conclusion: The DL model based on automatic segmentation technology shows significant promise in preoperatively predicting STAS in peripheral stage I LUAD, surpassing traditional radiomics models in diagnostic accuracy and clinical applicability. Radiomics Deep Learning Lung Adenocarcinoma Tumor Spread Through Air Spaces Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Introduction In 2015, the World Health Organization introduced the concept of tumor spread through air spaces (STAS) in its lung cancer classification[1]. Subsequent studies have confirmed that STAS is an independent risk factor for recurrence in patients with stage I lung adenocarcinoma (LUAD) who undergo sublobar resection[2]. Eguchi[3] suggested that for patients with T1 stage LUAD who are STAS-positive, lobectomy offers greater survival benefits compared to sublobar resection. Furthermore, STAS is also an independent adverse prognostic factor for patients with stage I LUAD[4-5], significantly associated with recurrence-free survival[6] (HR=1.975, 95% CI: 1.691-2.307). Therefore, accurate preoperative identification of STAS is critical for surgical planning and prognostic evaluation in stage I LUAD. Current studies indicate that intraoperative frozen section (FS) analysis has a sensitivity of 50% and a negative predictive value of only 8%, rendering it suboptimal for diagnosing STAS[7]. The limited efficacy of intraoperative FS diagnosis of STAS can affect the extent of resection and the choice of surgical method[8-9]. Additionally, due to the difficulty in obtaining live tissue specimens for pathological diagnosis of tumor cells within alveolar or air spaces, preoperative percutaneous biopsy is also inadequate for definitive STAS diagnosis. Thus, there is an urgent need for a more accurate preoperative method to diagnose STAS. Recently, imaging-based deep learning (DL) tools in the computer vision field have gained significant attention. They have shown great promise in quantifying early-stage lung cancer heterogeneity and providing potential clinical imaging features for patient stratification. Specifically, the clinical malignancy probability assessment based on radiomic features has demonstrated considerable potential [10-11]. Accurate tumor delineation is a priority in radiomics, however, several challenges remain: the accuracy and reproducibility of early-stage lung cancer lesion delineation and the robustness of radiomic feature extraction are still debated topics[12]. Recently, deep convolutional neural networks (CNNs) have achieved significant success in medical image segmentation, and CT images, being volumetric data, require the full exploitation of volumetric information[13-14]. Additionally, two challenges remain with the increased number of parameters: (1) Label scarcity due to the cost of annotations by experienced domain experts, and (2) the higher risk of overfitting. To address these issues, we proposed a automatic segmentation and DL method to utilize volumetric spatial information. We also assess the clinical applicability of DL based on automatic segmentation to predict STAS in peripheral stage I LUAD. The primary aim of this study was to compare models constructed using conventional radiomic features with a DL model based on automatic segmentation, to evaluate their clinical applicability in preoperatively predicting STAS in peripheral stage I LUAD. Additionally, this study explores the feasibility of using an automatic segmentation algorithm to identify lesion regions of interest (ROI) in peripheral stage I LUAD. Methods Patients and Clinical Data This retrospective analysis utilized data collected from January 2022 to December 2023. Clinical and radiological data were obtained from patients who underwent surgical treatment for lung tumors at our institution, supplemented with an external validation set from another hospital. Data Sets Inclusion criteria were as follows: i) Clinical stage T1-T2aN0M0, according to the 8th edition of the American Joint Committee on Cancer cancer staging manual [15]; ii) Tumors located in the outer two-thirds of the lung field on chest CT axial images, with the tumor center within this specified area; iii) Radical resection for lung cancer and systematic lymph node dissection with a minimum of 6 lymph nodes excised; iv) Postoperative pathological diagnosis confirmed as adenocarcinoma. Exclusion criteria included: i) Multiple pulmonary neoplastic lesions diagnosed preoperatively or synchronous primary or multiple primary lung cancers (more than 2 lesions) identified postoperatively; ii) Preoperative exposure to radiotherapy, chemotherapy, immunotherapy, or targeted therapy for cancer; iii) A history of other malignant tumors within the past three years. The study received approval from the local Institutional Review Board (2023-02-027-K01) and adhered to the Declaration of Helsinki. Informed consents were waived by the Committee due to the retrospective and anonymous nature of this study. The study was registered in the Research Registry (researchregistry10213). The work has been reported in line with the STARD (Standards for the Reporting of Diagnostic accuracy studies) criteria[16]. Compliance with the CheckList for EvaluAtion of Radiomics research (CLEAR) [17] guidelines was maintained, as detailed in Supplementary Table S1. The dataset was randomly divided in a 7:3 ratio into training and internal validation sets. Additionally, an external validation set was introduced to assess the model's generalizability. The training set utilized manually annotated ROIs, whereas the test set did not involve manual ROI annotations. Automatic segmentation algorithms were employed to predict and obtain the ROI segmentation results. Image Acquisition All CT scans were performed using a GE Discovery 750HD, SIEMENS SOMATOM Definition AS and SOMATOM Definition Flash scanners, spanning from the apex to the base of the lungs. Patients were positioned supine, with scan parameters set at a tube voltage of 120 kV and an automatic tube current ranging from 80 to 350 mA. The rotation time was 0.5-0.6 seconds per rotation. The standard scanning slice thickness and interval were 5 mm, with a reconstructed slice thickness and interval of 0.6-0.625 mm, and a display field of view of 40cm x 45cm. Images were analyzed using both lung (window width 1500 HU, window level -450 HU) and mediastinal (window width 350 HU, window level 35 HU) settings. For contrast-enhanced scans, iodinated contrast agent iohexol (350 mg/ml) was administered intravenously at a rate of 3ml/s, with a dosage of 1.5-2.0ml/kg. Arterial and venous phase scans were conducted 10 seconds and 30 seconds, respectively, after the aortic threshold reached 80 HU. ROI Segmentation In our research, we focused on refining the process of automating ROI segmentation by employing the VNet architecture. These models were trained specifically to carry out ROI segmentation autonomously, minimizing the need for manual input. We implemented an early stopping mechanism during training, setting a threshold at 32 checkpoints, to preserve the most efficient configurations of our model. The detailed methodologies utilized for training are described in Supplementary Material 1A. Deep Learning Signature Data Preparation ROI Cropping : For each subject, the most significant ROI slice was selected as the primary image. To focus the analysis and diminish external noise, we extracted the smallest rectangle that encompassed the ROI, adding a margin of 10 pixels to account for the importance of the surrounding tissue, as suggested by recent research. Model Training Data Augmentation : We normalized the intensity values across the RGB spectrum via Z-score normalization. These processed images were then fed into our deep learning models. We incorporated real-time augmentation techniques like random cropping and flipping during the training phase, whereas normalization remained the sole preprocessing step for the test images. Transfer Learning : Our study assessed the utility of well-known architectures such as Resnet101, Resnet50, and DenseNet121 to enhance traditional CNN models. We carried out comparative studies to identify the most suitable algorithm for our research needs. Optimizing Hyperparameters : We leveraged transfer learning, initializing our models with weights pre-trained on the ImageNet dataset to boost adaptability to varying datasets. One pivotal aspect of our approach was optimizing the learning rate for improved generalization, utilizing a cosine decay learning rate adjustment: Radiomics Signature Feature Extraction The extraction of handcrafted features encompasses three distinct categories: (1) geometric features, which encapsulate the three-dimensional form of the tumor; (2) intensity features, which capture the primary statistical distribution of voxel intensities within the tumor; and (3) textural features, which represent the patterns or higher-order spatial distributions of intensities. The extraction of textural features is accomplished through various techniques, including the methods of gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). Feature Selection Statistical: Initially, a z-score normalization was applied to all extracted features to transform them into a normal distribution. Consequently, the t-test was utilized for statistical analysis and feature selection among all radiomic features, retaining only those with a p-value less than 0.05. Correlation: For highly repeatable features, Pearson's rank correlation coefficient was utilized to assess the correlation between features. Among features with a correlation coefficient exceeding 0.9, only one feature was preserved. To maximize the descriptive capability of the features, a greedy recursive elimination strategy was adopted for feature filtering, sequentially removing the feature with the highest redundancy in each step. LASSO: The radiomics signature was developed using the least absolute shrinkage and selection operator (LASSO) regression model. This approach uses a regularization weight λ to reduce all regression coefficients towards zero, effectively eliminating many irrelevant features. Optimal λ was determined through 10-fold cross-validation, selecting the value that minimized cross-validation error. Features with nonzero coefficients were used to fit the regression model and form the radiomics signature. The radiomics score for each patient was calculated as a linear combination of these retained features, weighted by their respective coefficients. Signature Construction Radiomic Signature: Utilizing Lasso for feature selection, we integrated the selected features into six conventional machine learning models: Logistic Regression (LR), Support Vector Machines (SVM), Random Forest (RF), Light Gradient Boosting Machine (LGBM), eXtreme Gradient Boosting (XGB)and Extra Trees. The model demonstrating superior performance on the internal validation dataset was chosen for further comparative analysis across various signatures. DL Signature: The predicted probabilities derived from our CNN model were designated as the DL Signature. To explore the potential of multi-model integration, we employed three distinct fusion methods-mean, minimum, and maximum value fusion-to amalgamate the outcomes from these models. Model Evaluation Metrics: We assessed the diagnostic accuracy of our models via Receiver Operating Characteristic (ROC) curves. Model calibration was evaluated using calibration curves and Hosmer-Lemeshow (HL) tests, which ascertain the precision of the models' predictions. Decision Curve Analysis (DCA) was also implemented to determine the clinical utility of our predictive models. Statistical Methodology Our statistical evaluations and model development were executed using Python version 3.7.12, supplemented by the statsmodels library version 0.13.2. Machine learning frameworks were developed employing the scikit-learn library version 1.0.2. DL training utilized an NVIDIA 4090 GPU, with software frameworks including MONAI version 0.8.1 and PyTorch version 1.8.1. The images in the training set segmentation for 3D regions was conducted using the 3D Slicer software (version 5.3.0-2023-08-03). Results Baseline Characteristics A total of 290 cases met the inclusion and exclusion criteria, with 65 cases (22.41%) testing positive for STAS. The cohort comprised 55% males and 45% females, with an average age of 62 years. No statistically significant differences were found between clinical and pathological variables from Center 1 and Center 2, as all p-values exceeded 0.05. Study subject enrollment is depicted in Figure 1 and Figure 2 details the study flow. Radiomics Signature Feature Statistics : A total of 6 categories and 1834 handcrafted features were extracted, including 360 first-order features, 14 shape features, and the remaining texture features. All handcrafted features were extracted using an in-house feature analysis program implemented in Pyradiomics (http://pyradiomics.readthedocs.io)[18]. Figure 3 displays all features and their corresponding p-value results. Lasso Feature Selection : After Lasso 12 nonzero coefficients were selected to establish the Rad-score using a least absolute shrinkage and selection operator (LASSO) logistic regression model. The coefficients and the mean standard error (MSE) of the 10-fold validation are shown in Figure 4. In the validation cohort, the XGBoost model exhibited an Area Under the Curve (AUC) of 0.833 (95% CI 0.707-0.960). This performance metric signifies the model's ability to distinguish between the classes effectively, albeit not the highest among the compared algorithms. The AUC value, while respectable, indicates a moderate discriminative capability in the validation setting, necessitating further evaluation and potentially recalibration before clinical implementation or extensive comparative analysis with other models. As shown in Fig5&6. Deep Learning Signature In the validation cohort, the ResNet101 model achieved an AUC of 0.880 (95% CI 0.780-0.979). This AUC value indicates a high capability of the model to discriminate between positive and negative classes effectively. Despite this strong performance, the decision to utilize ResNet101 for further model comparisons suggests a strategic focus on exploring its utility against other models under varied conditions or specific performance aspects not solely defined by AUC. As shown in Fig7&8. We also visualized the model's prediction process using Grad-CAM, with detailed information available in Supplementary Material 2A. Signature Comparison In the comparison of traditional radiomics and DL-based models, the focus on AUC across test and validation cohorts reveals a distinct advantage for the DL approach. In the test cohort, the DL model demonstrates higher efficacy with an AUC of 0.880, compared to the radiomics model's AUC of 0.803. This indicates a superior discriminative capability in the DL model for distinguishing between positive and negative classes. Similarly, in the validation cohort, the DL signature again recorded an AUC of 0.880, which surpasses the traditional radiomics signature's AUC of 0.833. This further confirms the enhanced performance of DL models in generalizing to new, unseen data. These results underscore the effectiveness of DL models, particularly when integrated with automated delineation techniques, in achieving higher accuracy and reliability in medical diagnostics over traditional radiomics approaches that rely on manual delineation. As shown in Table1&Fig9. Calibration Curve : The HL test is key for assessing a predictive model's calibration, comparing predicted probabilities with actual outcomes. A higher HL statistic indicates better calibration, showing closer alignment between model predictions and observed outcomes. In our study, the DL model demonstrated excellent calibration, evidenced by HL test statistics of 0.828, 0.911 and 0.852 in the train, test and val cohort, suggesting its high effectiveness in reflecting observed data, as shown Fig10. DCA : Figure 11 illustrates the DCA for the train and test sets. These curves reveal that our fusion model provides considerable advantages in terms of its predictive probabilities. Discussion In this study, we proposed a DL algorithm based on automatic segmentation for predicting STAS in peripheral stage I LUAD, achieving an AUC of 0.880 (95% CI 0.780–0.979). This performance surpasses that of the conventional radiomics model, which achieved an AUC of 0.833 (95% CI 0.707–0.960), in both the test set and the external validation set. The integration of automated segmentation technology significantly enhances the clinical applicability of these results. For thoracic surgeons, given the poorer prognosis of STAS-positive patients, determining the presence of STAS in early-stage lung cancer is crucial for selecting the appropriate surgical approach. Research by Suzuki et al.[ 4 ] have demonstrated that postoperative local control and long-term survival rates are comparable between sublobar resection and lobectomy in patients with peripheral stage I LUAD, thereby affirming the clinical efficacy of sublobar resection for early-stage lung cancer. Liu et al[ 19 ] have indicated that in STAS-positive peripheral stage I lung cancer, STAS is closely associated with reduced recurrence-free survival (HR = 4.318, 95% CI: 1.593–11.701) and overall survival (HR = 4.421, 95% CI: 1.273–15.354). Compared to lobectomy, sublobar resection often results in higher local recurrence rates and lung cancer-specific mortality, suggesting that sublobar resection may not be the optimal choice for patients with STAS-positive peripheral stage I lung cancer. Raj et al.[ 20 ] have previously shown that STAS is a predictor of occult lymph node metastasis in clinical stage IA lung adenocarcinoma and may be an important factor leading to poor tumor prognosis. In patients who are eligible for both lobar and sublobar resection, intraoperative identification of STAS can help to determine the most appropriate type of resection to perform. Contradictorily, the diagnosis of STAS depends on the pathological results of the resected lung tissue post-surgery, which poses the disadvantage of delayed diagnosis. Julian et al.[ 21 ] reported the value of intraoperative frozen pathology in diagnosing STAS, with FS showing low sensitivity (44%), high specificity (91%), and accuracy (71%), and an AUC of 0.67. Due to the subjective factors of pathologists and technical limitations, FS has many constraints in diagnosing STAS. Owing to the emergence of radiomics, preoperative imaging might enable earlier diagnosis of STAS. Before starting this study, we reviewed the literature and found that many scholars have made unique contributions to predicting STAS using imaging characteristics. Traditional imaging features such as CTR, pleural indentation, and vessel cancer embolus are closely related to the occurrence of STAS[ 22 – 23 ]. However, due to measurement biases and subjective imaging factors, these indicators have limited value in predicting the occurrence of STAS in lung adenocarcinoma. Radiomics has achieved significant results in diagnosing and predicting the prognosis of various diseases. In predicting LUAD STAS, Jiang et al.[ 24 ] constructed a random forest model incorporating 12 radiomics features and age, with a predictive efficacy AUC of 0.754 (sensitivity of 0.880 and specificity of 0.588). Liao[ 25 ] conducted a comparative analysis of multiple predictive models combining radiomics features and clinical characteristics, constructing a model with 18 radiomics features and 2 clinical characteristics, achieving an AUC of 0.87 (95% CI: 0.82–0.92). Our previous network meta-analysis comprehensively evaluated the predictive value of various models for model construction algorithms and whether to include peritumoral features. Current reports indicate that machine learning radiomics models incorporating peritumoral features have better predictive efficacy[ 26 ]. However, only one study on a deep learning model was reported previously and was not included in the meta analysis. The implementation of machine learning combined with radiomics features in practical clinical applications still faces challenges, primarily due to the variability in model generalizability and external application stability. Even with some scholars open-sourcing their code, the manual delineation and feature extraction step remains an obstacle to practical application. Accurate identification of ROI is crucial for advancing artificial intelligence in medical diagnostics. The inefficiency associated with manual segmentation significantly limits the clinical applicability of AI-based radiomics technology. Our study explored the role of automatic segmentation technology in improving workflow and yielded satisfactory results. The automatic segmentation model constructed using the VNet algorithm showed minimal differences in recognition accuracy, as visualized in our results. These differences are considered negligible within our DL workflow, effectively validating the feasibility of our proposed automatic delineation process. Relying solely on radiographic scan information avoids potential errors caused by inaccurate or incomplete clinical variables and subjectively determined CT signs. Therefore, radiomics combined with DL models holds promise as a preoperative diagnostic strategy for predicting STAS status, providing reliable support for clinical surgical decisions and other treatment plans. Additionally, we proposed the feasibility of implementing automatic segmentation in peripheral stage I LUAD and constructing a DL model. Our results confirmed our hypothesis, as the predictive model based on the ResNet101 algorithm achieved AUCs of 0.880 (95% CI: 0.778–0.982) and 0.880 (95% CI: 0.780–0.979) in the test set and external validation set, respectively. The clinical decision curve and calibration curve results also showed better performance compared to the machine learning model based on the XGBoost algorithm. These findings suggest that a DL model provides superior clinical predictive value compared to traditional machine learning algorithms. Indeed, existing studies have constructed deep learning predictive models. According to our literature review, Wang et al.[ 27 ] developed a DL model, SE-Resnet50, which achieved an AUC of 0.933 [95% CI: 0.917–0.945], indicating higher predictive performance than our study. Unlike our research, their study included LUAD stages I-IV, whereas our focus was on peripheral stage I LUAD. Study by Lin et al.[ 28 ], which included 581 LUAD patients [Minimally invasive adenocarcinoma (MIA)-IIIA] from two centers, proposed a STAS DL model that achieved an AUC of 0.82 and an accuracy of 74% (with a sensitivity of 79%). These reports support our findings that DL has high application value in early prediction of STAS. In clinical practice, sublobar resection is often used for patients with tumors less than 2 cm in size located in the outer two-third of the lung parenchyma. Our study focuses on peripheral stage IA LUAD, and to our knowledge, this is the first report specifically addressing this particular stage. Our study results demonstrate the unique value of the DL model in preoperatively predicting the presence of STAS in peripheral stage IA LUAD. Therefore, the DL model based on automatic segmentation technology not only surpasses traditional machine learning in diagnostic efficacy but also facilitates the clinical application of AI technology in this field. Limitations This study has several limitations. Firstly, although we included an external validation dataset, future research should focus on increasing the sample size and incorporating multi-center data to validate the generalizability of the model. Secondly, selection bias is an inherent issue in any retrospective study. Additionally, we only included patients with peripheral stage I LUAD, so our conclusions are only applicable to this patient group. Finally, exploring the integration of multi-modal data, such as combining imaging with genetic or histopathological data, could further enhance the predictive accuracy and clinical applicability of the models. Investigating the potential benefits of incorporating tumor microenvironment features into the predictive models is also a promising direction. Prospective studies are needed to validate the model's performance in a real-world clinical setting and to assess its impact on clinical decision-making and patient outcomes. Conclusion In peripheral stage I LUAD, a DL model based on automatic segmentation outperforms the conventional radiomics model in predicting STAS. Additionally, the integrated workflow with automated segmentation technology significantly enhances the clinical applicability of the predictive models. Declarations Acknowledgements Guarantor: The scientifc guarantor of this publication is Xiao-Feng Li. Statistics and Biometry: No complex statistical methods were necessary for this paper. Author contribution Conception and design:Xiao‑Feng Li, Hui Qian; (II) Administrative support:Xiao‑Feng Li, Ping Gong, Yu‑Feng Wang; (III) Provision of study materials or patients:Chao Jia, Xiu-Qing Xue; (IV) Collection and assembly of data:Hong-Ying Zhao, Cong Liu; (V) Data analysis and interpretation:Cong Liu, Xiao‑Feng Li; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors. F unding This work was supported by Xuzhou Science and Technology Bureau Project [grant number KC23229]; The study was supported by Clinical medicine science and technology development foundation of Jiangsu University, China (Pro. No. JLY2021082). Data availability The data and and analysis code used in the current study are available in the open source website github (https://github.com/liucong1994/model-data.git). Ethical approval The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013), The ethics committee of Xuzhou Cancer Hospital approved the study protocol (2023-02-027-K01); This work was supported by the Natural Science Foundation of China (No.82001987). Conflict of interest The authors declare that they have no conflict of interest. Informed consent Due to the retrospective nature of our study, informed consent from patients was waived. Author details 1 Department of Minimally Invasive Oncology, XuZhou New Health Geriatric Hospital, Xuzhou, People’s Republic of China. 2 Departments of Nuclear Medicine, The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital, Xuzhou, People’s Republic of China. 3 School of Medical Imaging, Xuzhou Medical University, Xuzhou, People’s Republic of China. 4 Department of Nuclear Medicine, The First People’s Hospital of Yancheng, Yancheng, People’s Republic of China. 5 Department of Radiotherapy, The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital, Xuzhou, People’s Republic of China. 6 Medical College of Jiangsu University, Zhenjiang, People’s Republic of China. 7 Department of Radiology, The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospita, Xuzhou, People’s Republic of China. References Travis WD, Brambilla E, Nicholson AG, et al. WHO Panel. The 2015 World Health Organization Classification of Lung Tumors: Impact of Genetic, Clinical and Radiologic Advances Since the 2004 Classification. J Thorac Oncol. 2015 Sep;10(9):1243-1260. Ren Y, Xie H, Dai C, et al. Prognostic impact of tumor spread through air spaces in sublobar resection for 1A lung adenocarcinoma patients. Ann Surg Oncol 2019;26:1901-8. Eguchi T, Kameda K, Lu S, et al. Lobectomy is associated with better outcomes than sublobar resection in spread through air spaces (STAS)-positive T1 lung adenocarcinoma: a propensity score matched analysis. J Thorac Oncol 2019;14:87–98. Suzuki K, Saji H, Aokage K, et al. Comparison of pulmonary segmentectomy and lobectomy: Safety results of a randomized trial. J Thorac Cardiovasc Surg. 2019 Sep;158(3):895-907. Chen D, Mao Y, Wen J, et al. Tumor Spread Through Air Spaces in Non-Small Cell Lung Cancer: a systematic review and meta-analysis. The Annals of Thoracic Surgery 2019 Sep;108(3):945-954. Kadota K, Nitadori JI, Sima CS, et al. Tumor Spread through Air Spaces is an Important Pattern of Invasion and Impacts the Frequency and Location of Recurrences after Limited Resection for Small Stage I Lung Adenocarcinomas. J Thorac Oncol. 2015 May;10(5):806-814. Walts AE, Marchevsky AM. Current Evidence Does Not Warrant Frozen Section Evaluation for the Presence of Tumor Spread Through Alveolar Spaces. Arch Pathol Lab Med 2018; 142:59-63. Villalba JA, Shih AR, Sayo TMS, et al. Accuracy and Reproducibility of Intraoperative Assessment on Tumor Spread Through Air Spaces in Stage 1 Lung Adenocarcinomas. J Thorac Oncol. 2021 Apr;16(4):619-629. Zhou F, Villalba JA, Sayo TMS, et al. Assessment of the feasibility of frozen sections for the detection of spread through air spaces (STAS) in pulmonary adenocarcinoma. Mod Pathol. 2022;35(2):210-217. Zhang Y, Liao Q, Ding L, et al. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph. 2022;99:102088. Xie H, Chen Z, Deng J, et al. Automatic segmentation of the gross target volume in radiotherapy for lung cancer using transresSEUnet 2.5D Network. J Transl Med. 2022 Nov 12;20(1):524. Kalpathy-Cramer J, Zhao B, Goldgof D, et al. A Comparison of Lung Nodule Segmentation Algorithms: Methods and Results from a Multi-institutional Study. J Digit Imaging. 2016 Aug;29(4):476-87. Sumida I, Magome T, Kitamori H, et al. Deep convolutional neural network for reduction of contrast-enhanced region on CT images. J Radiat Res. 2019;60(5):586-594. Brosch T, Tang LY, Youngjin Yoo, et al. Deep 3D Convolutional Encoder Networks With Shortcuts for Multiscale Feature Integration Applied to Multiple Sclerosis Lesion Segmentation. IEEE Trans Med Imaging. 2016 May;35(5):1229-1239. Detterbeck FC, Boffa DJ, Kim AW, et al. The Eighth Edition Lung Cancer Stage Classification. Chest. 2017;151(1):193-203. Bossuyt PM, Reitsma JB, Bruns DE, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. Published 2015 Oct 28. Kocak B, Baessler B, Bakas S, et al. CheckList for EvaluAtion of Radiomics research (CLEAR): a step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMII. Insights Imaging. 2023;14(1):75. Published 2023 May 4. van Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, R. G. H., Fillon-Robin, J. C., Pieper, S., Aerts, H. J. W. L. (2017). Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research, 77(21), e104–e107. Liu A, Hou F, Qin Y, et al. Predictive value of a prognostic model based on pathologic features in lung invasive adenocarcinoma. Lung Cancer. 2019;131:14-22. doi:10.1016/j.lungcan.2019.03.002 Vaghjiani RG, Takahashi Y, Eguchi T, et al. Tumor Spread Through Air Spaces Is a Predictor of Occult Lymph Node Metastasis in Clinical Stage IA Lung Adenocarcinoma. J Thorac Oncol. 2020;15(5):792-802. doi:10.1016/j.jtho.2020.01.008 Villalba JA, Shih AR, Sayo TMS, et al. Accuracy and Reproducibility of Intraoperative Assessment on Tumor Spread Through Air Spaces in Stage 1 Lung Adenocarcinomas. J Thorac Oncol. 2021;16(4):619-629. doi:10.1016/j.jtho.2020.12.005 Jia C, Jiang HC, Liu C, et al. The correlation between tumor radiological features and spread through air spaces in peripheral stage IA lung adenocarcinoma: a propensity score-matched analysis. J Cardiothorac Surg. 2024;19(1):19. Published 2024 Jan 23. doi:10.1186/s13019-024-02498-0 Wang J, Yao Y, Tang D, Gao W. An individualized nomogram for predicting and validating spread through air space (STAS) in surgically resected lung adenocarcinoma: a single center retrospective analysis. J Cardiothorac Surg. 2023;18(1):337. Published 2023 Nov 21. doi:10.1186/s13019-023-02458-0 Jiang C, Luo Y, Yuan J, et al. CT-based radiomics and machine learning to predict spread through air space in lung adenocarcinoma. Eur Radiol. 2020;30(7):4050-4057. doi:10.1007/s00330-020-06694-z Liao G, Huang L, Wu S, et al. Preoperative CT-based peritumoral and tumoral radiomic features prediction for tumor spread through air spaces in clinical stage I lung adenocarcinoma. Lung Cancer. 2022;163:87-95. doi:10.1016/j.lungcan.2021.11.017 Liu C, Wang YF, Wang P, et al. Predictive value of multiple imaging predictive models for spread through air spaces of lung adenocarcinoma: A systematic review and network meta‑analysis. Oncol Lett. 2024;27(3):122. Published 2024 Jan 25. doi:10.3892/ol.2024.14255 Wang S, Liu X, Jiang C, et al. CT-Based Super-Resolution Deep Learning Models with Attention Mechanisms for Predicting Spread Through Air Spaces of Solid or Part-Solid Lung Adenocarcinoma. Acad Radiol. Published online January 5, 2024. doi:10.1016/j.acra.2023.12.034 Lin MW, Chen LW, Yang SM, et al. CT-Based Deep-Learning Model for Spread-Through-Air-Spaces Prediction in Ground Glass-Predominant Lung Adenocarcinoma. Ann Surg Oncol. 2024;31(3):1536-1545. doi:10.1245/s10434-023-14565-2 Table Table1. Metrics on different signature. Cohort Signature ACC AUC 95% CI SEN SPE PPV NPV train Radiomics 0.89 0.953 0.915-0.992 0.929 0.881 0.65 0.981 DL 0.815 0.899 0.838-0.959 0.786 0.822 0.512 0.942 val Radiomics 0.825 0.833 0.707-0.960 0.588 0.913 0.714 0.857 DL 0.825 0.88 0.780-0.979 0.824 0.826 0.636 0.927 test Radiomics 0.775 0.803 0.693-0.913 0.722 0.79 0.5 0.907 DL 0.85 0.88 0.778-0.982 0.722 0.887 0.65 0.917 DL: DeepLearning ACC: Accuracy SEN: Sensitivity SPE: Specificity PPV: Positive Predictive Value NPV: Negative Predictive Value Additional Declarations No competing interests reported. Supplementary Files Supplementary.pdf Cite Share Download PDF Status: Published Journal Publication published 08 Mar, 2025 Read the published version in Respiratory Research → Version 1 posted Editorial decision: Revision requested 31 Oct, 2024 Reviews received at journal 31 Oct, 2024 Reviewers agreed at journal 10 Oct, 2024 Reviews received at journal 08 Oct, 2024 Reviewers agreed at journal 07 Sep, 2024 Reviewers invited by journal 24 Jul, 2024 Editor assigned by journal 24 Jul, 2024 Submission checks completed at journal 23 Jul, 2024 First submitted to journal 19 Jul, 2024 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-4768392","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":342159109,"identity":"7ff0ff2f-5c05-4fd0-a236-bcc6af168f67","order_by":0,"name":"Cong Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAz0lEQVRIiWNgGAWjYBACNvbm4z8+VPyzk2dvPkCcFj6eYwmSM84cSDbsOZZAnBY5CR8Dac62A4wNN3IMiHSYBIOBMWPbHWbGnjMfb7xhsJPTbSCkRbohIbng3DM+dvbezZZzGJKNzQ4Q0iJz4MDhGWXMQFvObpPmYTiQuI2gFonExmYeNmaQX54RqyWZmZmn7TBICxuRWniOsTHOOJMGCmRjyzkGRPhFvr3/G8OHChtQVD688abCTo6gFhQgwUNk1CBrIVXHKBgFo2AUjAgAAMA+RRGPFkTOAAAAAElFTkSuQmCC","orcid":"","institution":"XuZhou New Health Geriatric Hospital","correspondingAuthor":true,"prefix":"","firstName":"Cong","middleName":"","lastName":"Liu","suffix":""},{"id":342159110,"identity":"e039e4ab-73ae-4194-88db-f8fbd09efedc","order_by":1,"name":"Yu-feng Wang","email":"","orcid":"","institution":"The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yu-feng","middleName":"","lastName":"Wang","suffix":""},{"id":342159113,"identity":"e19db6c2-365b-407c-afaf-3d0bca082a92","order_by":2,"name":"Ping Gong","email":"","orcid":"","institution":"Xuzhou Medical College","correspondingAuthor":false,"prefix":"","firstName":"Ping","middleName":"","lastName":"Gong","suffix":""},{"id":342159116,"identity":"7ac3bb90-d188-470a-bdb8-c3587d6990ef","order_by":3,"name":"Xiu-Qing Xue","email":"","orcid":"","institution":"Yancheng First People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiu-Qing","middleName":"","lastName":"Xue","suffix":""},{"id":342159117,"identity":"bce4c8a2-73ed-41bd-a126-5bec63307644","order_by":4,"name":"Hong-Ying Zhao","email":"","orcid":"","institution":"The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hong-Ying","middleName":"","lastName":"Zhao","suffix":""},{"id":342159119,"identity":"9c906868-7bd6-45da-b3d9-3baa60e1028d","order_by":5,"name":"Hui Qian","email":"","orcid":"","institution":"Jiangsu University","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Qian","suffix":""},{"id":342159121,"identity":"53552c37-85bb-4762-ae15-77f0db688b58","order_by":6,"name":"Chao Jia","email":"","orcid":"","institution":"The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chao","middleName":"","lastName":"Jia","suffix":""},{"id":342159124,"identity":"668ee43b-71c0-4883-8071-8547085348fe","order_by":7,"name":"Xiao-Feng Li","email":"","orcid":"","institution":"The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiao-Feng","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2024-07-19 15:08:42","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4768392/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4768392/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12931-025-03174-0","type":"published","date":"2025-03-08T15:57:55+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":63285684,"identity":"5bbab224-b389-4b99-bdf6-f92b3dad1269","added_by":"auto","created_at":"2024-08-26 13:41:00","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":85189,"visible":true,"origin":"","legend":"\u003cp\u003eScreening flowchart for enrolled patients\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/e9396555adf60cc58ab15a53.png"},{"id":63288402,"identity":"7f9efa91-9bef-4008-940c-190cbd835824","added_by":"auto","created_at":"2024-08-26 13:57:00","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":149406,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of this study.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/9ccb0727f7200f4813bb0b95.png"},{"id":63285685,"identity":"3b3c4569-9067-4b7a-927d-e0e49c70c5eb","added_by":"auto","created_at":"2024-08-26 13:41:00","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":46992,"visible":true,"origin":"","legend":"\u003cp\u003eNumber and ratio of handcrafted features\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/64caeba9808ac41a85271c09.png"},{"id":63287074,"identity":"fe4283a8-269f-4769-bff7-1631cfac2d20","added_by":"auto","created_at":"2024-08-26 13:49:00","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":154887,"visible":true,"origin":"","legend":"\u003cp\u003eCoefficients of 10 fold cross validation(A), MSE of 10 fold cross validation(B), The histogram of the Rad-score based on the selected features(C).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/2abd0026b1aa91f5ab3963c2.png"},{"id":63285686,"identity":"b9562f7e-3c7e-42b8-90c1-a28331c0529c","added_by":"auto","created_at":"2024-08-26 13:41:00","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":217270,"visible":true,"origin":"","legend":"\u003cp\u003eMetric results for Machine Learning Radiomics Signature.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/45104408420e5ad489e5729b.png"},{"id":63285690,"identity":"852cc1b5-37d8-4d15-b96a-6b53ba563ab2","added_by":"auto","created_at":"2024-08-26 13:41:00","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":108770,"visible":true,"origin":"","legend":"\u003cp\u003eROC results for Radiomics Signature of different Machine Learning model.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/ba7223506840b9c0862438ad.png"},{"id":63287078,"identity":"ce78c970-e3f7-42a7-9618-e3fa47bb12ee","added_by":"auto","created_at":"2024-08-26 13:49:00","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":101087,"visible":true,"origin":"","legend":"\u003cp\u003eMetric results for Deep Learning Radiomics Signature.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/18d70238fffe37beafd78648.png"},{"id":63288403,"identity":"64d622f5-d206-431f-942b-0fb11e15e766","added_by":"auto","created_at":"2024-08-26 13:57:00","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":68039,"visible":true,"origin":"","legend":"\u003cp\u003eROC results for Deep Learning Signature of different model.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/3a3bc30e299704385ed0a315.png"},{"id":63285696,"identity":"89f3db85-b188-42a5-b0e3-6c6067b179c0","added_by":"auto","created_at":"2024-08-26 13:41:01","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":61697,"visible":true,"origin":"","legend":"\u003cp\u003eIllustrates the ROC for different signatures across various cohorts, offering a visual comparison of their diagnostic abilities.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/dc157fb4c0297632b331af69.png"},{"id":63287077,"identity":"edece740-10fd-49f5-9218-4ad9d64ffdeb","added_by":"auto","created_at":"2024-08-26 13:49:00","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":58197,"visible":true,"origin":"","legend":"\u003cp\u003eDisplays the calibration curves for different signatures in the test cohort. These curves are instrumental in understanding how well the predicted probabilities of the models match the actual outcomes.\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/b0e7165dc5130945109c1feb.png"},{"id":63285689,"identity":"1aef39f2-59fd-42aa-979a-2c426ea7a2c6","added_by":"auto","created_at":"2024-08-26 13:41:00","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":64784,"visible":true,"origin":"","legend":"\u003cp\u003eDifferent signatures' DCA on test cohort.\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/14dddea7ba237a3b9dc5e2a0.png"},{"id":78190359,"identity":"8f8549b2-7056-4d19-b952-cf472dc66d94","added_by":"auto","created_at":"2025-03-10 19:48:45","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1787673,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/99793f7a-efd6-43d1-9b19-244d07fbe7f4.pdf"},{"id":63285694,"identity":"67e59ade-321b-4ff2-b9c7-894d5a64daa1","added_by":"auto","created_at":"2024-08-26 13:41:00","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1656524,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4768392/v1/da60f75b686d06a9d99be4a8.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Prediction of Tumor Spread Through Air Spaces with an Automatic Segmentation Deep Learning Model in Peripheral Stage I Lung Adenocarcinoma","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn 2015, the World Health Organization introduced the concept of tumor spread through air spaces (STAS) in its lung cancer classification[1]. Subsequent studies have confirmed that STAS is an independent risk factor for recurrence in patients with stage I lung adenocarcinoma (LUAD) who undergo sublobar resection[2]. Eguchi[3] suggested that for patients with T1 stage LUAD who are STAS-positive, lobectomy offers greater survival benefits compared to sublobar resection. Furthermore, STAS is also an independent adverse prognostic factor for patients with stage I LUAD[4-5], significantly associated with recurrence-free survival[6] (HR=1.975, 95% CI: 1.691-2.307). Therefore, accurate preoperative identification of STAS is critical for surgical planning and prognostic evaluation in stage I LUAD.\u003c/p\u003e\n\u003cp\u003eCurrent studies indicate that intraoperative frozen section (FS)\u0026nbsp;analysis has a sensitivity of 50% and a negative predictive value of only 8%, rendering it suboptimal for diagnosing STAS[7]. The limited efficacy of intraoperative FS diagnosis of STAS can affect the extent of resection and the choice of surgical method[8-9]. Additionally, due to the difficulty in obtaining live tissue specimens for pathological diagnosis of tumor cells within alveolar or air spaces, preoperative percutaneous biopsy is also inadequate for definitive STAS diagnosis. Thus, there is an urgent need for a more accurate preoperative method to diagnose STAS.\u003c/p\u003e\n\u003cp\u003eRecently, imaging-based deep learning (DL) tools in the computer vision field have gained significant attention. They have shown great promise in quantifying early-stage lung cancer heterogeneity and providing potential clinical imaging features for patient stratification. Specifically, the clinical malignancy probability assessment based on radiomic features has demonstrated considerable potential [10-11]. Accurate tumor delineation is a priority in radiomics, however, several challenges remain: the accuracy and reproducibility of early-stage lung cancer lesion delineation and the robustness of radiomic feature extraction are still debated topics[12]. Recently, deep convolutional neural networks (CNNs) have achieved significant success in medical image segmentation, and CT images, being volumetric data, require the full exploitation of volumetric information[13-14]. Additionally, two challenges remain with the increased number of parameters: (1) Label scarcity due to the cost of annotations by experienced domain experts, and (2) the higher risk of overfitting. To address these issues, we proposed a automatic segmentation and DL\u0026nbsp; method to utilize volumetric spatial information. We also assess the clinical applicability of DL based on automatic segmentation to predict STAS in peripheral stage I LUAD.\u003c/p\u003e\n\u003cp\u003eThe primary aim of this study was to compare models constructed using conventional radiomic features with a DL model based on automatic segmentation, to evaluate their clinical applicability in preoperatively predicting STAS in peripheral stage I LUAD. Additionally, this study explores the feasibility of using an automatic segmentation algorithm to identify lesion regions of interest (ROI) in peripheral stage I LUAD.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003ePatients and Clinical Data\u003c/p\u003e\n\u003cp\u003eThis retrospective analysis utilized data collected from January 2022 to December 2023. Clinical and radiological data were obtained from patients who underwent surgical treatment for lung tumors at our institution, supplemented with an external validation set from another hospital.\u003c/p\u003e\n\u003ch3\u003eData Sets\u003c/h3\u003e\n\u003cp\u003eInclusion criteria were as follows: i) Clinical stage T1-T2aN0M0, according to the 8th edition of the American Joint Committee on Cancer cancer staging manual [15]; ii) Tumors located in the outer two-thirds of the lung field on chest CT axial images, with the tumor center within this specified area; iii) Radical resection for lung cancer and systematic lymph node dissection with a minimum of 6 lymph nodes excised; iv) Postoperative pathological diagnosis confirmed as adenocarcinoma. Exclusion criteria included: i) Multiple pulmonary neoplastic lesions diagnosed preoperatively or synchronous primary or multiple primary lung cancers (more than 2 lesions) identified postoperatively; ii) Preoperative exposure to radiotherapy, chemotherapy, immunotherapy, or targeted therapy for cancer; iii) A history of other malignant tumors within the past three years. The study received approval from the local Institutional Review Board (2023-02-027-K01) and adhered to the Declaration of Helsinki. Informed consents were waived by the Committee due to the retrospective and anonymous nature of this study. The study was registered in the Research Registry (researchregistry10213). The work has been reported in line with the STARD (Standards for the Reporting of Diagnostic accuracy studies) criteria[16]. Compliance with the CheckList for EvaluAtion of Radiomics research (CLEAR) [17] guidelines was maintained, as detailed in Supplementary Table S1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe dataset was randomly divided in a 7:3 ratio into training and internal validation sets. Additionally, an external validation set was introduced to assess the model's generalizability. The training set utilized manually annotated ROIs, whereas the test set did not involve manual ROI annotations. Automatic segmentation algorithms were employed to predict and obtain the ROI segmentation results.\u003c/p\u003e\n\u003ch3\u003eImage Acquisition\u003c/h3\u003e\n\u003cp\u003eAll CT scans were performed using a GE Discovery 750HD, SIEMENS SOMATOM Definition AS and SOMATOM Definition Flash scanners, spanning from the apex to the base of the lungs. Patients were positioned supine, with scan parameters set at a tube voltage of 120 kV and an automatic tube current ranging from 80 to 350 mA. The rotation time was 0.5-0.6 seconds per rotation. The standard scanning slice thickness and interval were 5 mm, with a reconstructed slice thickness and interval of 0.6-0.625 mm, and a display field of view of 40cm x 45cm. Images were analyzed using both lung (window width 1500 HU, window level -450 HU) and mediastinal (window width 350 HU, window level 35 HU) settings. For contrast-enhanced scans, iodinated contrast agent iohexol (350 mg/ml) was administered intravenously at a rate of 3ml/s, with a dosage of 1.5-2.0ml/kg. Arterial and venous phase scans were conducted 10 seconds and 30 seconds, respectively, after the aortic threshold reached 80 HU.\u003c/p\u003e\n\u003ch3\u003eROI Segmentation\u003c/h3\u003e\n\u003cp\u003eIn our research, we focused on refining the process of automating ROI segmentation by employing the VNet architecture. These models were trained specifically to carry out ROI segmentation autonomously, minimizing the need for manual input. We implemented an early stopping mechanism during training, setting a threshold at 32 checkpoints, to preserve the most efficient configurations of our model. The detailed methodologies utilized for training are described in Supplementary Material 1A.\u003c/p\u003e\n\u003ch3\u003eDeep Learning Signature\u003c/h3\u003e\n\u003ch4\u003eData Preparation\u003c/h4\u003e\n\u003cp\u003e\u003cstrong\u003eROI Cropping\u003c/strong\u003e: For each subject, the most significant ROI slice was selected as the primary image. To focus the analysis and diminish external noise, we extracted the smallest rectangle that encompassed the ROI, adding a margin of 10 pixels to account for the importance of the surrounding tissue, as suggested by recent research.\u003c/p\u003e\n\u003ch4\u003eModel Training\u003c/h4\u003e\n\u003cp\u003e\u003cstrong\u003eData Augmentation\u003c/strong\u003e: We normalized the intensity values across the RGB spectrum via Z-score normalization. These processed images were then fed into our deep learning models. We incorporated real-time augmentation techniques like random cropping and flipping during the training phase, whereas normalization remained the sole preprocessing step for the test images.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTransfer Learning\u003c/strong\u003e: Our study assessed the utility of well-known architectures such as Resnet101, Resnet50, and DenseNet121 to enhance traditional CNN models. We carried out comparative studies to identify the most suitable algorithm for our research needs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOptimizing Hyperparameters\u003c/strong\u003e: We leveraged transfer learning, initializing our models with weights pre-trained on the ImageNet dataset to boost adaptability to varying datasets. One pivotal aspect of our approach was optimizing the learning rate for improved generalization, utilizing a cosine decay learning rate adjustment:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003cbr\u003e\u003c/p\u003e\n\u003ch3\u003eRadiomics Signature\u003c/h3\u003e\n\u003ch4\u003eFeature Extraction\u003c/h4\u003e\n\u003cp\u003eThe extraction of handcrafted features encompasses three distinct categories: (1) geometric features, which encapsulate the three-dimensional form of the tumor; (2) intensity features, which capture the primary statistical distribution of voxel intensities within the tumor; and (3) textural features, which represent the patterns or higher-order spatial distributions of intensities. The extraction of textural features is accomplished through various techniques, including the methods of gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM).\u003c/p\u003e\n\u003ch4\u003eFeature Selection\u003c/h4\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical:\u003c/strong\u003e Initially, a z-score normalization was applied to all extracted features to transform them into a normal distribution. Consequently, the t-test was utilized for statistical analysis and feature selection among all radiomic features, retaining only those with a p-value less than 0.05.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCorrelation:\u003c/strong\u003e For highly repeatable features, Pearson's rank correlation coefficient was utilized to assess the correlation between features. Among features with a correlation coefficient exceeding 0.9, only one feature was preserved. To maximize the descriptive capability of the features, a greedy recursive elimination strategy was adopted for feature filtering, sequentially removing the feature with the highest redundancy in each step.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLASSO:\u003c/strong\u003e The radiomics signature was developed using the least absolute shrinkage and selection operator (LASSO) regression model. This approach uses a regularization weight \u003cem\u003eλ\u003c/em\u003e to reduce all regression coefficients towards zero, effectively eliminating many irrelevant features. Optimal \u003cem\u003eλ\u003c/em\u003e was determined through 10-fold cross-validation, selecting the value that minimized cross-validation error. Features with nonzero coefficients were used to fit the regression model and form the radiomics signature. The radiomics score for each patient was calculated as a linear combination of these retained features, weighted by their respective coefficients.\u003c/p\u003e\n\u003ch3\u003eSignature Construction\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eRadiomic Signature:\u003c/strong\u003e Utilizing Lasso for feature selection, we integrated the selected features into six conventional machine learning models: Logistic Regression (LR), Support Vector Machines (SVM), Random Forest (RF), Light Gradient Boosting Machine (LGBM), eXtreme Gradient Boosting (XGB)and Extra Trees. The model demonstrating superior performance on the internal validation dataset was chosen for further comparative analysis across various signatures.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDL Signature:\u003c/strong\u003e The predicted probabilities derived from our CNN model were designated as the DL Signature. To explore the potential of multi-model integration, we employed three distinct fusion methods-mean, minimum, and maximum value fusion-to amalgamate the outcomes from these models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Evaluation Metrics:\u003c/strong\u003e We assessed the diagnostic accuracy of our models via Receiver Operating Characteristic (ROC) curves. Model calibration was evaluated using calibration curves and Hosmer-Lemeshow (HL) tests, which ascertain the precision of the models' predictions. Decision Curve Analysis (DCA) was also implemented to determine the clinical utility of our predictive models.\u003c/p\u003e\n\u003ch3\u003eStatistical Methodology\u003c/h3\u003e\n\u003cp\u003eOur statistical evaluations and model development were executed using Python version 3.7.12, supplemented by the statsmodels library version 0.13.2. Machine learning frameworks were developed employing the scikit-learn library version 1.0.2. DL training utilized an NVIDIA 4090 GPU, with software frameworks including MONAI version 0.8.1 and PyTorch version 1.8.1. The images in the training set segmentation for 3D regions was conducted using the 3D Slicer software (version 5.3.0-2023-08-03).\u003c/p\u003e"},{"header":"Results","content":"\u003ch3\u003eBaseline Characteristics\u003c/h3\u003e\n\u003cp\u003eA total of 290 cases met the inclusion and exclusion criteria, with 65 cases (22.41%) testing positive for STAS. The cohort comprised 55% males and 45% females, with an average age of 62 years. No statistically significant differences were found between clinical and pathological variables from Center 1 and Center 2, as all p-values exceeded 0.05. Study subject enrollment is depicted in Figure 1 and Figure 2 details the study flow.\u003c/p\u003e\n\u003ch3\u003eRadiomics Signature\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eFeature Statistics\u003c/strong\u003e: A total of 6 categories and 1834 handcrafted features were extracted, including 360 first-order features, 14 shape features, and the remaining texture features. All handcrafted features were extracted using an in-house feature analysis program implemented in Pyradiomics (http://pyradiomics.readthedocs.io)[18]. Figure 3 displays all features and their corresponding p-value results.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLasso Feature Selection\u003c/strong\u003e: After Lasso 12 nonzero coefficients were selected to establish the Rad-score using a least absolute shrinkage and selection operator (LASSO) logistic regression model. The coefficients and the mean standard error (MSE) of the 10-fold validation are shown in Figure 4.\u003c/p\u003e\n\u003cp\u003eIn the validation cohort, the XGBoost model exhibited an Area Under the Curve (AUC) of 0.833 (95% CI 0.707-0.960). This performance metric signifies the model's ability to distinguish between the classes effectively, albeit not the highest among the compared algorithms. The AUC value, while respectable, indicates a moderate discriminative capability in the validation setting, necessitating further evaluation and potentially recalibration before clinical implementation or extensive comparative analysis with other models. As shown in Fig5\u0026amp;6.\u003c/p\u003e\n\u003ch3\u003eDeep Learning Signature\u003c/h3\u003e\n\u003cp\u003eIn the validation cohort, the ResNet101 model achieved an AUC of 0.880 (95% CI 0.780-0.979). This AUC value indicates a high capability of the model to discriminate between positive and negative classes effectively. Despite this strong performance, the decision to utilize ResNet101 for further model comparisons suggests a strategic focus on exploring its utility against other models under varied conditions or specific performance aspects not solely defined by AUC. As shown in Fig7\u0026amp;8.\u003c/p\u003e\n\u003cp\u003eWe also visualized the model's prediction process using Grad-CAM, with detailed information available in Supplementary Material 2A.\u003c/p\u003e\n\u003ch3\u003eSignature Comparison\u003c/h3\u003e\n\u003cp\u003eIn the comparison of traditional radiomics and DL-based models, the focus on AUC across test and validation cohorts reveals a distinct advantage for the DL approach.\u003c/p\u003e\n\u003cp\u003eIn the test cohort, the DL model demonstrates higher efficacy with an AUC of 0.880, compared to the radiomics model's AUC of 0.803. This indicates a superior discriminative capability in the DL model for distinguishing between positive and negative classes. Similarly, in the validation cohort, the DL signature again recorded an AUC of 0.880, which surpasses the traditional radiomics signature's AUC of 0.833. This further confirms the enhanced performance of DL models in generalizing to new, unseen data.\u003c/p\u003e\n\u003cp\u003eThese results underscore the effectiveness of DL models, particularly when integrated with automated delineation techniques, in achieving higher accuracy and reliability in medical diagnostics over traditional radiomics approaches that rely on manual delineation. As shown in Table1\u0026amp;Fig9.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCalibration Curve\u003c/strong\u003e: The HL test is key for assessing a predictive model's calibration, comparing predicted probabilities with actual outcomes. A higher HL statistic indicates better calibration, showing closer alignment between model predictions and observed outcomes. In our study, the DL model demonstrated excellent calibration, evidenced by HL test statistics of 0.828, 0.911 and 0.852 in the train, test and val cohort, suggesting its high effectiveness in reflecting observed data, as shown Fig10.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDCA\u003c/strong\u003e: Figure 11 illustrates the DCA for the train and test sets. These curves reveal that our fusion model provides considerable advantages in terms of its predictive probabilities.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we proposed a DL algorithm based on automatic segmentation for predicting STAS in peripheral stage I LUAD, achieving an AUC of 0.880 (95% CI 0.780\u0026ndash;0.979). This performance surpasses that of the conventional radiomics model, which achieved an AUC of 0.833 (95% CI 0.707\u0026ndash;0.960), in both the test set and the external validation set. The integration of automated segmentation technology significantly enhances the clinical applicability of these results.\u003c/p\u003e \u003cp\u003eFor thoracic surgeons, given the poorer prognosis of STAS-positive patients, determining the presence of STAS in early-stage lung cancer is crucial for selecting the appropriate surgical approach. Research by Suzuki et al.[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] have demonstrated that postoperative local control and long-term survival rates are comparable between sublobar resection and lobectomy in patients with peripheral stage I LUAD, thereby affirming the clinical efficacy of sublobar resection for early-stage lung cancer. Liu et al[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] have indicated that in STAS-positive peripheral stage I lung cancer, STAS is closely associated with reduced recurrence-free survival (HR\u0026thinsp;=\u0026thinsp;4.318, 95% CI: 1.593\u0026ndash;11.701) and overall survival (HR\u0026thinsp;=\u0026thinsp;4.421, 95% CI: 1.273\u0026ndash;15.354). Compared to lobectomy, sublobar resection often results in higher local recurrence rates and lung cancer-specific mortality, suggesting that sublobar resection may not be the optimal choice for patients with STAS-positive peripheral stage I lung cancer. Raj et al.[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] have previously shown that STAS is a predictor of occult lymph node metastasis in clinical stage IA lung adenocarcinoma and may be an important factor leading to poor tumor prognosis. In patients who are eligible for both lobar and sublobar resection, intraoperative identification of STAS can help to determine the most appropriate type of resection to perform. Contradictorily, the diagnosis of STAS depends on the pathological results of the resected lung tissue post-surgery, which poses the disadvantage of delayed diagnosis. Julian et al.[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] reported the value of intraoperative frozen pathology in diagnosing STAS, with FS showing low sensitivity (44%), high specificity (91%), and accuracy (71%), and an AUC of 0.67. Due to the subjective factors of pathologists and technical limitations, FS has many constraints in diagnosing STAS.\u003c/p\u003e \u003cp\u003eOwing to the emergence of radiomics, preoperative imaging might enable earlier diagnosis of STAS. Before starting this study, we reviewed the literature and found that many scholars have made unique contributions to predicting STAS using imaging characteristics. Traditional imaging features such as CTR, pleural indentation, and vessel cancer embolus are closely related to the occurrence of STAS[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. However, due to measurement biases and subjective imaging factors, these indicators have limited value in predicting the occurrence of STAS in lung adenocarcinoma. Radiomics has achieved significant results in diagnosing and predicting the prognosis of various diseases. In predicting LUAD STAS, Jiang et al.[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] constructed a random forest model incorporating 12 radiomics features and age, with a predictive efficacy AUC of 0.754 (sensitivity of 0.880 and specificity of 0.588). Liao[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] conducted a comparative analysis of multiple predictive models combining radiomics features and clinical characteristics, constructing a model with 18 radiomics features and 2 clinical characteristics, achieving an AUC of 0.87 (95% CI: 0.82\u0026ndash;0.92). Our previous network meta-analysis comprehensively evaluated the predictive value of various models for model construction algorithms and whether to include peritumoral features. Current reports indicate that machine learning radiomics models incorporating peritumoral features have better predictive efficacy[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. However, only one study on a deep learning model was reported previously and was not included in the meta analysis. The implementation of machine learning combined with radiomics features in practical clinical applications still faces challenges, primarily due to the variability in model generalizability and external application stability. Even with some scholars open-sourcing their code, the manual delineation and feature extraction step remains an obstacle to practical application.\u003c/p\u003e \u003cp\u003eAccurate identification of ROI is crucial for advancing artificial intelligence in medical diagnostics. The inefficiency associated with manual segmentation significantly limits the clinical applicability of AI-based radiomics technology. Our study explored the role of automatic segmentation technology in improving workflow and yielded satisfactory results. The automatic segmentation model constructed using the VNet algorithm showed minimal differences in recognition accuracy, as visualized in our results. These differences are considered negligible within our DL workflow, effectively validating the feasibility of our proposed automatic delineation process. Relying solely on radiographic scan information avoids potential errors caused by inaccurate or incomplete clinical variables and subjectively determined CT signs. Therefore, radiomics combined with DL models holds promise as a preoperative diagnostic strategy for predicting STAS status, providing reliable support for clinical surgical decisions and other treatment plans. Additionally, we proposed the feasibility of implementing automatic segmentation in peripheral stage I LUAD and constructing a DL model. Our results confirmed our hypothesis, as the predictive model based on the ResNet101 algorithm achieved AUCs of 0.880 (95% CI: 0.778\u0026ndash;0.982) and 0.880 (95% CI: 0.780\u0026ndash;0.979) in the test set and external validation set, respectively. The clinical decision curve and calibration curve results also showed better performance compared to the machine learning model based on the XGBoost algorithm. These findings suggest that a DL model provides superior clinical predictive value compared to traditional machine learning algorithms. Indeed, existing studies have constructed deep learning predictive models. According to our literature review, Wang et al.[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] developed a DL model, SE-Resnet50, which achieved an AUC of 0.933 [95% CI: 0.917\u0026ndash;0.945], indicating higher predictive performance than our study. Unlike our research, their study included LUAD stages I-IV, whereas our focus was on peripheral stage I LUAD. Study by Lin et al.[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], which included 581 LUAD patients [Minimally invasive adenocarcinoma (MIA)-IIIA] from two centers, proposed a STAS DL model that achieved an AUC of 0.82 and an accuracy of 74% (with a sensitivity of 79%). These reports support our findings that DL has high application value in early prediction of STAS. In clinical practice, sublobar resection is often used for patients with tumors less than 2 cm in size located in the outer two-third of the lung parenchyma. Our study focuses on peripheral stage IA LUAD, and to our knowledge, this is the first report specifically addressing this particular stage. Our study results demonstrate the unique value of the DL model in preoperatively predicting the presence of STAS in peripheral stage IA LUAD. Therefore, the DL model based on automatic segmentation technology not only surpasses traditional machine learning in diagnostic efficacy but also facilitates the clinical application of AI technology in this field.\u003c/p\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eThis study has several limitations. Firstly, although we included an external validation dataset, future research should focus on increasing the sample size and incorporating multi-center data to validate the generalizability of the model. Secondly, selection bias is an inherent issue in any retrospective study. Additionally, we only included patients with peripheral stage I LUAD, so our conclusions are only applicable to this patient group. Finally, exploring the integration of multi-modal data, such as combining imaging with genetic or histopathological data, could further enhance the predictive accuracy and clinical applicability of the models. Investigating the potential benefits of incorporating tumor microenvironment features into the predictive models is also a promising direction. Prospective studies are needed to validate the model's performance in a real-world clinical setting and to assess its impact on clinical decision-making and patient outcomes.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn peripheral stage I LUAD, a DL model based on automatic segmentation outperforms the conventional radiomics model in predicting STAS. Additionally, the integrated workflow with automated segmentation technology significantly enhances the clinical applicability of the predictive models.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGuarantor: The scientifc guarantor of this publication is\u0026nbsp;Xiao-Feng\u0026nbsp;Li. Statistics and Biometry: No complex statistical methods were necessary for this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConception and design:Xiao‑Feng Li, Hui Qian; (II) Administrative support:Xiao‑Feng Li, Ping Gong, Yu‑Feng Wang; (III) Provision of study materials or patients:Chao Jia, Xiu-Qing Xue; (IV) Collection and assembly of data:Hong-Ying Zhao, Cong Liu; (V) Data analysis and interpretation:Cong Liu, Xiao‑Feng Li; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eF\u003c/strong\u003e\u003cstrong\u003eunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by Xuzhou Science and Technology Bureau Project [grant number KC23229]; The study was supported by Clinical medicine science and technology development foundation of Jiangsu University, China (Pro. No. JLY2021082).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data and and analysis code used in the current study are available in the open source website github (https://github.com/liucong1994/model-data.git).\u003c/p\u003e\n\u003cp\u003eEthical approval\u003c/p\u003e\n\u003cp\u003eThe study was conducted in accordance with the Declaration of Helsinki (as revised in 2013), The ethics committee of Xuzhou Cancer Hospital approved the study protocol (2023-02-027-K01); This work was supported by the Natural Science Foundation of China (No.82001987).\u003c/p\u003e\n\u003cp\u003eConflict of interest\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003cp\u003eInformed consent\u003c/p\u003e\n\u003cp\u003eDue to the retrospective nature of our study, informed consent from patients was waived.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor details\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e1\u003c/sup\u003eDepartment of Minimally Invasive Oncology, XuZhou New Health Geriatric Hospital, Xuzhou, People\u0026rsquo;s Republic of China. \u003csup\u003e2\u003c/sup\u003eDepartments of Nuclear Medicine, The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital, Xuzhou, People\u0026rsquo;s Republic of China. \u0026nbsp;\u003csup\u003e3\u003c/sup\u003eSchool of Medical Imaging, Xuzhou Medical University, Xuzhou, People\u0026rsquo;s Republic of China. \u003csup\u003e4\u003c/sup\u003eDepartment of Nuclear Medicine, The First People\u0026rsquo;s Hospital of Yancheng, Yancheng, \u0026nbsp;People\u0026rsquo;s Republic of China. \u003csup\u003e5\u003c/sup\u003eDepartment of Radiotherapy, The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospital, Xuzhou, People\u0026rsquo;s Republic of China. \u003csup\u003e6\u003c/sup\u003eMedical College of Jiangsu University, Zhenjiang, People\u0026rsquo;s Republic of China. \u003csup\u003e7\u003c/sup\u003eDepartment of Radiology, The Xuzhou Hospital Affiliated to Jiangsu University, Xuzhou Cancer Hospita, Xuzhou, People\u0026rsquo;s Republic of China.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eTravis WD, Brambilla E, Nicholson AG, et al. WHO Panel. The 2015 World Health Organization Classification of Lung Tumors: Impact of Genetic, Clinical and Radiologic Advances Since the 2004 Classification. J Thorac Oncol. 2015 Sep;10(9):1243-1260.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eRen Y, Xie H, Dai C, et al. Prognostic impact of tumor spread through air spaces in sublobar resection for 1A lung adenocarcinoma patients. Ann Surg Oncol 2019;26:1901-8.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eEguchi T, Kameda K, Lu S, et al. Lobectomy is associated with better outcomes than sublobar resection in spread through air spaces (STAS)-positive T1 lung adenocarcinoma: a propensity score matched analysis. J Thorac Oncol 2019;14:87\u0026ndash;98.\u003c/li\u003e\n \u003cli\u003eSuzuki K, Saji H, Aokage K, et al. Comparison of pulmonary segmentectomy and lobectomy: Safety results of a randomized trial. J Thorac Cardiovasc Surg. 2019 Sep;158(3):895-907.\u003c/li\u003e\n \u003cli\u003eChen D, Mao Y, Wen J, et al. Tumor Spread Through Air Spaces in Non-Small Cell Lung Cancer: a systematic review and meta-analysis. The Annals of Thoracic Surgery 2019 Sep;108(3):945-954.\u003c/li\u003e\n \u003cli\u003eKadota K, Nitadori JI, Sima CS, et al. Tumor Spread through Air Spaces is an Important Pattern of Invasion and Impacts the Frequency and Location of Recurrences after Limited Resection for Small Stage I Lung Adenocarcinomas. J Thorac Oncol. 2015 May;10(5):806-814.\u003c/li\u003e\n \u003cli\u003eWalts AE, Marchevsky AM. Current Evidence Does Not Warrant Frozen Section Evaluation for the Presence of Tumor Spread Through Alveolar Spaces. Arch Pathol Lab Med 2018; 142:59-63.\u003c/li\u003e\n \u003cli\u003eVillalba JA, Shih AR, Sayo TMS,\u0026nbsp;et al. Accuracy and Reproducibility of Intraoperative Assessment on Tumor Spread Through Air Spaces in Stage 1 Lung Adenocarcinomas. J Thorac Oncol. 2021 Apr;16(4):619-629.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eZhou F, Villalba JA, Sayo TMS, et al. Assessment of the feasibility of frozen sections for the detection of spread through air spaces (STAS) in pulmonary adenocarcinoma. Mod Pathol. 2022;35(2):210-217.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eZhang Y, Liao Q, Ding L,\u0026nbsp;et al. Bridging 2D and 3D segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5D solutions. Comput Med Imaging Graph. 2022;99:102088.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eXie H, Chen Z, Deng J,\u0026nbsp;et al. Automatic segmentation of the gross target volume in radiotherapy for lung cancer using transresSEUnet 2.5D Network. J Transl Med. 2022 Nov 12;20(1):524.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eKalpathy-Cramer J, Zhao B, Goldgof D,\u0026nbsp;et al.\u0026nbsp;A Comparison of Lung Nodule Segmentation Algorithms: Methods and Results from a Multi-institutional Study. J Digit Imaging. 2016 Aug;29(4):476-87.\u003c/li\u003e\n \u003cli\u003eSumida I, Magome T, Kitamori H, et al. Deep convolutional neural network for reduction of contrast-enhanced region on CT images. J Radiat Res. 2019;60(5):586-594.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBrosch T, Tang LY, Youngjin Yoo,\u0026nbsp;et al. Deep 3D Convolutional Encoder Networks With Shortcuts for Multiscale Feature Integration Applied to Multiple Sclerosis Lesion Segmentation. IEEE Trans Med Imaging. 2016 May;35(5):1229-1239.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eDetterbeck FC, Boffa DJ, Kim AW,\u0026nbsp;et al. The Eighth Edition Lung Cancer Stage Classification. Chest. 2017;151(1):193-203.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBossuyt PM, Reitsma JB, Bruns DE, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. Published 2015 Oct 28.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eKocak B, Baessler B, Bakas S, et al. CheckList for EvaluAtion of Radiomics research (CLEAR): a step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMII. Insights Imaging. 2023;14(1):75. Published 2023 May 4.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003e\u003cem\u003evan Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, R. G. H., Fillon-Robin, J. C., Pieper, S., Aerts, H. J. W. L. (2017). Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research, 77(21), e104\u0026ndash;e107.\u0026nbsp;\u003c/em\u003e\u003c/li\u003e\n \u003cli\u003eLiu A, Hou F, Qin Y, et al. Predictive value of a prognostic model based on pathologic features in lung invasive adenocarcinoma.\u0026nbsp;Lung Cancer. 2019;131:14-22. doi:10.1016/j.lungcan.2019.03.002\u003c/li\u003e\n \u003cli\u003eVaghjiani RG, Takahashi Y, Eguchi T, et al. Tumor Spread Through Air Spaces Is a Predictor of Occult Lymph Node Metastasis in Clinical Stage IA Lung Adenocarcinoma.\u0026nbsp;J Thorac Oncol. 2020;15(5):792-802. doi:10.1016/j.jtho.2020.01.008\u003c/li\u003e\n \u003cli\u003eVillalba JA, Shih AR, Sayo TMS, et al. Accuracy and Reproducibility of Intraoperative Assessment on Tumor Spread Through Air Spaces in Stage 1 Lung Adenocarcinomas.\u0026nbsp;J Thorac Oncol. 2021;16(4):619-629. doi:10.1016/j.jtho.2020.12.005\u003c/li\u003e\n \u003cli\u003eJia C, Jiang HC, Liu C, et al. The correlation between tumor radiological features and spread through air spaces in peripheral stage IA lung adenocarcinoma: a propensity score-matched analysis.\u0026nbsp;J Cardiothorac Surg. 2024;19(1):19. Published 2024 Jan 23. doi:10.1186/s13019-024-02498-0\u003c/li\u003e\n \u003cli\u003eWang J, Yao Y, Tang D, Gao W. An individualized nomogram for predicting and validating spread through air space (STAS) in surgically resected lung adenocarcinoma: a single center retrospective analysis.\u0026nbsp;J Cardiothorac Surg. 2023;18(1):337. Published 2023 Nov 21. doi:10.1186/s13019-023-02458-0\u003c/li\u003e\n \u003cli\u003eJiang C, Luo Y, Yuan J, et al. CT-based radiomics and machine learning to predict spread through air space in lung adenocarcinoma.\u0026nbsp;Eur Radiol. 2020;30(7):4050-4057. doi:10.1007/s00330-020-06694-z\u003c/li\u003e\n \u003cli\u003eLiao G, Huang L, Wu S, et al. Preoperative CT-based peritumoral and tumoral radiomic features prediction for tumor spread through air spaces in clinical stage I lung adenocarcinoma.\u0026nbsp;Lung Cancer. 2022;163:87-95. doi:10.1016/j.lungcan.2021.11.017\u003c/li\u003e\n \u003cli\u003eLiu C, Wang YF, Wang P, et al. Predictive value of multiple imaging predictive models for spread through air spaces of lung adenocarcinoma: A systematic review and network meta‑analysis.\u0026nbsp;Oncol Lett. 2024;27(3):122. Published 2024 Jan 25. doi:10.3892/ol.2024.14255\u003c/li\u003e\n \u003cli\u003eWang S, Liu X, Jiang C, et al. CT-Based Super-Resolution Deep Learning Models with Attention Mechanisms for Predicting Spread Through Air Spaces of Solid or Part-Solid Lung Adenocarcinoma.\u0026nbsp;Acad Radiol. Published online January 5, 2024. doi:10.1016/j.acra.2023.12.034\u003c/li\u003e\n \u003cli\u003eLin MW, Chen LW, Yang SM, et al. CT-Based Deep-Learning Model for Spread-Through-Air-Spaces Prediction in Ground Glass-Predominant Lung Adenocarcinoma. Ann Surg Oncol. 2024;31(3):1536-1545. doi:10.1245/s10434-023-14565-2\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"648\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"100%\" colspan=\"9\"\u003e\n \u003cp\u003eTable1. Metrics on different signature.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eCohort\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eSignature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eACC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e95% CI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eSEN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eSPE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003ePPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eNPV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"11.11111111111111%\" rowspan=\"2\"\u003e\n \u003cp\u003etrain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eRadiomics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.953\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.915-0.992\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.929\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.881\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.981\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003eDL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.815\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.899\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.838-0.959\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.786\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.822\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.512\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.942\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"11.11111111111111%\" rowspan=\"2\"\u003e\n \u003cp\u003eval\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eRadiomics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.825\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.833\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.707-0.960\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.588\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.913\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.714\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.857\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003eDL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.825\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.780-0.979\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.824\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.826\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.636\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.927\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"11.11111111111111%\" rowspan=\"2\"\u003e\n \u003cp\u003etest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003eRadiomics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.775\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.803\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.693-0.913\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.722\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.11111111111111%\"\u003e\n \u003cp\u003e0.907\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003eDL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.778-0.982\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.722\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.887\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"12.5%\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"100%\" colspan=\"9\"\u003e\n \u003cp\u003eDL: DeepLearning \u0026nbsp; \u0026nbsp; \u0026nbsp;ACC: Accuracy \u0026nbsp;SEN: Sensitivity \u0026nbsp;SPE: Specificity\u003cbr\u003e\u0026nbsp;PPV: \u0026nbsp;Positive Predictive Value \u0026nbsp;NPV: Negative Predictive Value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"respiratory-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"rere","sideBox":"Learn more about [Respiratory Research](http://respiratory-research.biomedcentral.com/)","snPcode":"12931","submissionUrl":"https://submission.nature.com/new-submission/12931/3","title":"Respiratory Research","twitterHandle":"@RespiratoryBMC","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Radiomics, Deep Learning, Lung Adenocarcinoma, Tumor Spread Through Air Spaces","lastPublishedDoi":"10.21203/rs.3.rs-4768392/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4768392/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose:\u003c/strong\u003e To evaluate the clinical applicability of deep learning (DL) models based on automatic segmentation in preoperatively predicting tumor spread through air spaces (STAS) in peripheral stage I lung adenocarcinoma (LUAD).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e This retrospective study analyzed data from patients who underwent surgical treatment for lung tumors from January 2022 to December 2023. An external validation set was introduced to assess the model's generalizability. The study utilized conventional radiomic features and DL models for comparison. ROI segmentation was performed using the VNet architecture, and DL models were developed with transfer learning and optimization techniques. We assessed the diagnostic accuracy of our models via calibration curves, decision curve analysis, and ROC curves.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The DL model based on automatic segmentation achieved an AUC of 0.880 (95% CI 0.780-0.979), outperforming the conventional radiomics model with an AUC of 0.833 (95% CI 0.707-0.960). The DL model demonstrated superior performance in both internal validation and external testing cohorts. Calibration curves, decision curve analysis, and ROC curves confirmed the enhanced diagnostic accuracy and clinical utility of the DL approach.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eThe DL model based on automatic segmentation technology shows significant promise in preoperatively predicting STAS in peripheral stage I LUAD, surpassing traditional radiomics models in diagnostic accuracy and clinical applicability.\u003c/p\u003e","manuscriptTitle":"Prediction of Tumor Spread Through Air Spaces with an Automatic Segmentation Deep Learning Model in Peripheral Stage I Lung Adenocarcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-08-26 13:40:54","doi":"10.21203/rs.3.rs-4768392/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-10-31T16:10:16+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-10-31T15:48:51+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"27248070905989632630334703135294651589","date":"2024-10-10T16:34:26+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-10-08T12:54:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"49461114827054735513673520043984750269","date":"2024-09-08T00:33:06+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-07-25T00:41:49+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-07-24T16:08:28+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-07-24T01:01:19+00:00","index":"","fulltext":""},{"type":"submitted","content":"Respiratory Research","date":"2024-07-19T15:07:19+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"respiratory-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"rere","sideBox":"Learn more about [Respiratory Research](http://respiratory-research.biomedcentral.com/)","snPcode":"12931","submissionUrl":"https://submission.nature.com/new-submission/12931/3","title":"Respiratory Research","twitterHandle":"@RespiratoryBMC","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"399d61c6-9f9b-49a8-b8ae-5687a0191162","owner":[],"postedDate":"August 26th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-03-10T19:45:27+00:00","versionOfRecord":{"articleIdentity":"rs-4768392","link":"https://doi.org/10.1186/s12931-025-03174-0","journal":{"identity":"respiratory-research","isVorOnly":false,"title":"Respiratory Research"},"publishedOn":"2025-03-08 15:57:55","publishedOnDateReadable":"March 8th, 2025"},"versionCreatedAt":"2024-08-26 13:40:54","video":"","vorDoi":"10.1186/s12931-025-03174-0","vorDoiUrl":"https://doi.org/10.1186/s12931-025-03174-0","workflowStages":[]},"version":"v1","identity":"rs-4768392","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4768392","identity":"rs-4768392","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}