A deep learning approach for predicting visceral pleural invasion in cT1 lung adenocarcinoma

Introduction

Surgical resection is currently the most effective treatment for early-stage non-small cell lung cancer (NSCLC) (1,2). Over the past few decades, questions have been raised by some researchers regarding the applicability of lobectomy (3,4). A series of randomized controlled trials (RCTs) have been conducted in this field by research institutions, with the Japan Clinical Oncology Group (JCOG) being represented as one of them (5-9). With the publication of these research results, it has been seen that sublobar resection is feasible in some patients. Studies included in this group are cT1 stage lung nodules less than 2–3 cm in size. Although researchers have adopted some evaluation criteria to screen suitable patients for sublobar resection, there is still a varying proportion (0.4–14.6%) of individuals with postoperative pathological stage reaching stage T2a or higher (5,8-10). This suggests that tumor heterogeneity should not be ignored when making the decision to perform sublobar resection in cT1 stage patients. There are numerous factors that contribute to the upgrading of pT stage, with one notable factor being visceral pleural invasion (VPI). Currently, there is a research challenge in accurately determining the status of VPI prior to surgery for patients eligible for sublobar resection.

Currently, the choice of surgical strategy intraoperatively is predominantly dependent on intraoperative frozen pathology reports, which may be inadequate in cases of VPI. Computed tomography (CT) data can be extracted to identify imaging-based features and used with machine learning methods to predict the malignancy, pathological subtype, and prognosis of pulmonary nodules (11-13). The research on the heterogeneity of pulmonary nodules is still inadequate, particularly in terms of lacking data support for the VPI status of the nodules. The aim of this study is to develop a method for predicting pulmonary nodule VPI status prior to surgery, improving the accuracy of predicting outcomes, identifying cT1 patients who may experience pathological stage upgrading, and guiding surgical decision-making by surgeons. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-601/rc).

Methods

Study design and population

The researchers reviewed information on patients who received treatment at the Department of Thoracic Surgery of the Fifth Affiliated Hospital of Sun Yat-sen University (Figure 1), between 1 January 2018 and 31 October 2022. After screening, 983 eligible patients were eventually included in the study. The 8th edition of the International Association for the Study of Lung Cancer staging system was utilized in this study. Inclusion criteria were as follows: (I) nodules diameter ≤3 cm; (II) located around the periphery of the lungs; (III) ≤5 mm away from the pleura; and (IV) invasive adenocarcinoma. The exclusion criteria were as follows: (I) preoperative distant metastasis; (II) preoperative neoadjuvant therapy; (III) severe CT artifacts; (IV) sublobectomy; (V) multiple primary lung cancer and (IV) cases with incomplete data that cannot be analyzed. After an 8:2 ratio random split, 786 patients were included in the training cohort for model construction, and 197 patients were included in the validation cohort. The purpose of internal validation as part of model development is to check the repeatability of the model development process and to prevent overfitting of the model leading to overestimation of the model performance. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University, Sun Yat-sen University [No. (2024) K83-1] and individual consent for this retrospective analysis was waived.

Figure 1 Flow diagram of the study population. CT, computed tomography.

Feature extraction and selection

The details of CT protocol and image segmentation are presented in Appendices 1,2. Pyradiomics in Python (version 3.7) was used to extract radiomics features from CT images (Appendix 3). The ResNet-50 model was used to extract deep learning features from 2D and 2.5D data, while the 3D ResNet-50 model is specifically designed for extracting deep learning features from 3D data (Figure 2, Appendix 3). These extracted features were normalized to a standard dataset with a mean of 0 and a variance of 1 (Figure 2).

Figure 2 The workflow of this study. In step 1, the ROI is defined based on the segmented visible contour of the tumor. The MCS of the ROI region on the transverse plane is defined as the 2D data. The 2.5D data include the MCS of the sagittal plane, coronal plane, and transverse plane, as well as the planes of two voxels above and below the MCS plane of the transverse section. The definition that only includes the ROI region is referred to as ROI-only, while the definition that includes the rectangular region of that plane is referred to as ROI-rect. The 3D data are also divided into two categories, where the minimum bounding hexahedron of the ROI region is defined as ROI-rect, and the data that only include the ROI region is referred to as ROI-only. In step 2, the data generated from step 1 were respectively inputted into the ResNet-50, Pyradiomics, and 3D-ResNet-50 models to extract features. Step 3, the extracted features are filtered in the training cohort, followed by model establishment and subsequent validation in the validation cohort. ROI-only: data that exclusively focused on the ROI; ROI-rect: the minimum bounding rectangle or cuboid of the ROI. ROI, region of interest; MCS, maximum cross-section; ROC, receiver operating characteristic; LASSO, least absolute shrinkage and selection operator.

All radiomic and deep learning features were statistically tested and screened using Mann-Whitney U test, and only those with a P value <0.05 were retained. Features with high repeatability were evaluated using Spearman’s rank correlation coefficient to calculate the correlation between features, and it was determined that one feature with a correlation coefficient greater than 0.9 between any two features was retained. Subsequently, the least absolute shrinkage and selection operator (LASSO) regression model was used on the discovery data set for signature construction. Employing 10-fold cross validation with minimum criteria, an optimal λ was found that yielded minimum cross validation error. The retained features with nonzero coefficients were used for regression model fitting and combined into a radiomics signature. Subsequently, radiomics scores were obtained for each patient by linearly combining retained features weighted by their model coefficients using the Python scikit-learn package.

Model construction and validation

Models were developed for each region on the training set, and then tested on the validation cohort. Receiver operating characteristic (ROCs) curves were generated for each model using multi-layer perceptron (MLP). The performance of these models was evaluated and validated using the training and test cohorts, including areas under the curve (AUCs), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The AUC and cut-off values of the corresponding model were derived using the Youden Index, with a comprehensive account of the cut-off acquisition procedure detailed in Appendix 4.

Statistical analysis

Statistical analysis was conducted using SPSS (version 26.0, IBM, Armonk, NY, USA), and Python software (version 3.7). Continuous variables were analyzed using Student’s t-test and were expressed as mean ± standard deviation. Categorical variables were analyzed using the Chi-square test or Fisher’s exact test and were presented as ratios. Multivariable logistic regression analysis was performed using the forward stepwise selection method. According to the results of ROC curve analysis, optimal cut-off values for the final best model were used to classify the patients into high- and low-risk groups. Subsequently, survival curves were plotted using the Kaplan-Meier method and then compared between groups using the log-rank test. A two-tailed P value of <0.05 was considered statistically significant. The statistical significance of the ROCs of models was demonstrated through Delong’s test (Appendix 5).

Results

Patients’ characteristics and models

A total of 983 patients (380 male, 603 female) were recruited for our study, including 786 patients in the training cohort and 197 patients in the validation cohort. Clinical characteristics, pathological and radiological characteristics in CT images are shown in Table 1. These characteristics showed no statistical difference (P>0.05) between the cohorts. Univariate and multivariable logistic regression analyses showed that smoke, carcinoembryonic antigen (CEA), lobulated sign, and pleural indentation were independent predictors of VPI (Table 2). The clinical model, radiological model, and combined model were constructed using MLP to evaluate the diagnostic efficacy of different models uniformly and objectively (Figure 3A-3C). The ROC curves of the three models in the training and validation cohorts are presented in Table 3. The clinical characteristics model demonstrated poor predictive performance in the validation set, with an AUC of 0.532 [95% confidence interval (CI): 0.441–0.623] (Figure 3A). The radiological model showed better predictive performance, however, the fusion model, which incorporated clinical characteristics, did not improve the overall performance (AUC: 0.563, 95% CI: 0.474–0.651). Statistical significance was observed in the difference through Delong’s test (Appendix 5).

Figure 3 The efficacy of predicting VPI across various models. (A-D) The clinical, radiological, clinical + radiological, and radiomic models. (E-G) The 2D, 2.5D and 3D models. (H) The DCA and calibration curves of the models in the validation cohort. ROI-only: data that exclusively focused on the ROI; ROI-rect: the minimum bounding rectangle or cuboid of the ROI. VPI, visceral pleural invasion; ROI, region of interest; AUC, area under the curve; CT, computed tomography; DCA, decision curve analysis.

Radiomic and deep learning models

A total of 1,288 radiomic features were extracted by Pyradiomics, and 2,048 deep learning features were extracted by ResNet-50/3D-ResNet-50 from a single region of interest (ROI). The features, which had nonzero coefficients, were retained after undergoing feature scaling and selection (Appendix 3). Subsequently, a modeling process was conducted using MLP. The ROC curves of the models in the training and validation cohorts are shown in Figure 3. Similar predictive capabilities were demonstrated by the radiomic model and the radiological model (AUC: 0.823 vs. 0.810), with no statistically significant differences observed between the two models (Appendix 5). The decision curve analysis (DCA) curve and calibration curve also indicated similar predictive performance between the two models (Figure 3H).

Deep learning models were employed to conduct an analysis on 2D, 2.5D, and 3D data (Figure 3E-3G). From an overall perspective, the 3D-ROI-only model exhibited the highest AUC value (0.952, 95% CI: 0.921–0.983). However, it was observed that both the DCA curve and calibration curve indicated a suboptimal predictive performance of this model (Figure 3H). In comparison to the radiomic model, the deep learning models also demonstrated favorable AUC values in the ROI-rect (the minimum bounding rectangle or cuboid of the ROI) groups of 2D and 2.5D (AUC_2D-ROI-rect: 0.877, 95% CI: 0.827–0.927; AUC_{2.5D-ROI-rect}: 0.857, 95% CI: 0.795–0.919). However, it was indicated by both the DCA curve and the calibration curve that the 2D-rect model exhibited superior predictive capabilities (Figure 3H). The results of the Delong test for each model can be found in Appendix 5.

Overall survival (OS)

The OS curves for all patients are displayed in Appendix 6. A total of 34 patients succumbed to mortality. The median follow-up time is 38 months (95% CI: 36.346–38.654). The OS rates at 1, 3, and 5 years were 99.0%, 94.7%, and 88.7%, respectively. Survival analysis was conducted in the deep learning models by selecting the models with higher AUC values in the 2D, 2.5D, and 3D models, respectively (Figure 4). Within each model, the validation cohort patients were categorized into low-risk and high-risk groups based on their corresponding cut-off values. In the 2D-rect and 2.5D-rect models, a statistically significant suppression between the two groups was observed (P=0.02 and 0.04). The radiological and 3D-ROI-only models displayed a discernible trend of differentiation between the groups. However, the potential of these two models to predict survival duration under prolonged follow-up conditions remains to be observed.

Figure 4 Comparison of Kaplan-Meier curves for prognosis in the validation cohort. Each model demonstrated the capability to stratify patients based on overall survival. However, there were no statistically significant differences in stratification ability among the radiological (A), the radiomic (B), and the 3D ROI-only (E) model. Notably, the 2D (C) and the 2.5D (D) models exhibited more pronounced stratification abilities, with P values of 0.02 and 0.04, respectively. ROI-only: data that exclusively focused on the ROI; ROI-rect: the minimum bounding rectangle or cuboid of the ROI. ROI, region of interest.

Discussion

In recent years, an increasing number of research results have shown that in cT1 stage patients, the OS and/or disease-free survival (DFS) after sublobar resection may be comparable to those after lobectomy (5-10). This suggests that a more limited resection, preserving more lung tissue, may be feasible for some patients. However, there are still several issues in need of urgent resolution in these studies. Despite the preoperative selection of cT1 patients for inclusion in the studies, there are still individual cases where postoperative pathological results indicate a higher stage, and patients who underwent sublobar resection show a higher rate of local recurrence (11% vs. 5%, P=0.02) (2,5,8-10). The long-term follow-up from the JCOG0201 study suggests that VPI is a risk factor for the recurrence of stage I lung adenocarcinoma (hazard ratio: 2.17; 95% CI: 1.23–3.81; P=0.07). Furthermore, subgroup analysis data shows that patients with VPI have a lower 10-year recurrence-free survival (RFS) (64.0% vs. 87.2%) (14). Therefore, it is particularly important to identify patients who may have VPI among T1 patients undergoing planned sublobar resection. Historically, the presence of VPI has been preliminarily predicted to a certain extent through certain CT signs, such as visceral pleural traction and visceral pleural concavity. Despite strict definitions for these signs, subjective judgment sometimes influences the prediction, and the sensitivity and specificity of the predictions are unsatisfactory (15).

ResNet-50, a classic convolutional neural network (CNN), has been widely applied in medical image recognition and semantic segmentation (16,17). In the current investigation, CT images were utilized for the extraction of deep learning features, which were employed in the development of a ResNet-50 model. In this study, peripheral pulmonary nodules with a maximum diameter of <3 cm, nodular edge distance from the pleura of less than 5 mm, and postoperative pathological confirmation of invasive adenocarcinoma were selected.

A better performance was achieved in the validation cohort compared to radiomic and other models in this study. In both the 2D and 2.5D models, the AUC of ROI-rect was higher than that of ROI-only. Surprisingly, the models in the ROI-only group did not even outperform the radiological model. This may be due to the presence of valuable features outside the ROI region when significant extranodal invasion features, such as VPI, are present. An interesting finding in this study was that the 3D-ROI-only model achieved an AUC of 0.952, but did not show improved predictive performance in terms of DCA curve, calibration curve, and survival analysis evaluation. Additionally, the 3D-ROI-rect model performed worse than the 3D-ROI-only model. It is speculated that this may be caused by the inclusion of too much irrelevant information when taking the minimum enclosing cuboid in the 3D image, especially in the planes close to the edges where the ROI region occupies a small proportion. It is worth noting that, according to this study, a predictive model established based on radiological features also achieved an accuracy of 0.801 in the validation set (Table 3). Although the radiomic group exhibited higher accuracy and sensitivity, no statistical difference in the ROC curves between the radiomic group and the radiological group was found, and the AUCs were similar (AUC: 0.823 vs. 0.810, P=0.69) (Appendix 5). However, other studies have demonstrated that radiomic models exhibit a superior predictive effect compared to radiological models, potentially attributed to a greater focus on factors such as pathological type, depth of invasion, and intrapulmonary dissemination (12,16,18). When VPI occurs, specific CT morphological features, including pleural retraction, pleural attachment, and lobulation, are observed, thereby enhancing the predictive capability (19-21). The likelihood of pleural invasion has been observed to positively correlate with the extent of contact between the lesion and the pleura by Imai et al. (22). Additionally, pleural indentation has been confirmed as an independent predictor of VPI, with reports indicating that it results from scar formation within the tumor, which is associated with VPI induced by tumor-induced lung atelectasis (23). In fact, these CT manifestations are widely considered as CT features of both pulmonary malignancies and potential VPI. Kim et al. found that CT features exhibited an accuracy level ranging from 62.7% to 72.3% in predicting VPI status (15). Although there is also evidence suggesting that other CT features such as lobulation can serve as predictive factors for VPI (15,20), these findings align with the observations in this study. However, the accuracy and stability of the aforementioned research on these CT features are not satisfactory, as the recognition of these radiological features relies on the expertise and subjective perception of radiologists, leading to heterogeneity and suboptimal accuracy (20,24,25).

In the JCOG0802 study, the 5-year OS for the segmentectomy group and lobectomy group were 94.1% and 91.1%, with statistically significant results (8). The JCOG0802 study became the first prospective RCTs to demonstrate the superiority of segmentectomy over lobectomy in terms of OS in early-stage lung cancer. However, the local recurrence rate after segmentectomy (11%) is nearly twice that of lobectomy (5%). Although this higher local recurrence rate is not reflected in OS, whether this advantage of segmentectomy can be maintained in longer-term follow-up remains to be observed. This suggests that sublobar resection may not be suitable for certain patients. In performing sublobar resection, there is insufficient evaluation for patients who may have positive lymph nodes between segments or high-risk nodules that have spread beyond the surgical margin (VPI, spread through air spaces, etc.). Further investigation is warranted to ascertain the potential benefits of sublobar resection in these specific patient cohorts. However, this issue is also commonly observed in other similar studies (5-10,14). At present, there is a lack of reliable preoperative or intraoperative methods to accurately identify patients who may not derive benefits from the sublobar resection. In this study, a predictive method based on machine learning was developed using preoperative CT image features to classify patients into high-risk and low-risk groups. Figure 4 suggests that in T1 lung cancer, there may be a subset of patients with poorer prognosis based on VPI status. These patients can be identified to some extent using non-invasive machine learning methods preoperatively. Although this study lacks data on this subset, a hypothesis can still be proposed: the benefit of sublobar resection may be limited in this subgroup. However, this indicates that surgeons should still consider lobectomy as the standard surgical procedure for curative treatment of lung cancer and exercise caution when deciding to perform sublobar resection.

The findings of this study indicate that the VPI status of patients with cT1-stage lung adenocarcinoma could be predicted using a deep learning model based on ResNet-50 and established with 2D maximum cross-section (MCS) imaging data, as compared to 2.5D and 3D data. Furthermore, a preliminary investigation was conducted in this study to explore the subset of patients who may receive lesser benefits from sublobar resection. Despite the favorable performance of the predictive model proposed in this study, it is essential to acknowledge certain limitations. Firstly, although the sample size in this study is relatively large, it remains a single-center retrospective investigation. Secondly, the follow-up period is relatively short, with only 34 patients reaching the study endpoint, potentially impacting the model’s performance. Nonetheless, the implementation of this model aids in the development of treatment strategies for individuals with early-stage T1 lung adenocarcinoma and may offer preoperative guidance for surgeons contemplating sublobar resection.

Conclusions

An advanced computational model, developed through the implementation of a deep learning signature, has provided a robust tool for accurately predicting vascular invasion in stage cT1 lung adenocarcinoma, thereby improving stratification for prognostic assessment. Additionally, this sophisticated model assists in the optimization of therapeutic strategies for individuals diagnosed with cT1 lung adenocarcinoma.

Acknowledgments

We are grateful for the valuable assistance provided by Jingzhi Wu in the areas of image acquisition, preprocessing, and segmentation.

Funding: None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-601/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-601/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-601/prf

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-601/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the Ethics Committee of the Fifth Affiliated Hospital of Sun Yat-sen University, Sun Yat-sen University [No. (2024) K83-1] and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ginsberg RJ, Rubinstein LV. Randomized trial of lobectomy versus limited resection for T1 N0 non-small cell lung cancer. Lung Cancer Study Group. Ann Thorac Surg 1995;60:615-22; discussion 622-3. [Crossref] [PubMed]
2. Ginsberg RJ, Rubinstein L. The comparison of limited resection to lobectomy for T1N0 non-small cell lung cancer. LCSG 821. Chest 1994;106:318S-9S.
3. Kodama K, Higashiyama M, Yokouchi H, et al. Prognostic value of ground-glass opacity found in small lung adenocarcinoma on high-resolution CT scanning. Lung Cancer 2001;33:17-25. [Crossref] [PubMed]
4. Aoki T, Tomoda Y, Watanabe H, et al. Peripheral lung adenocarcinoma: correlation of thin-section CT findings with histologic prognostic factors and survival. Radiology 2001;220:803-9. [Crossref] [PubMed]
5. Aokage K, Suzuki K, Saji H, et al. Segmentectomy for ground-glass-dominant lung cancer with a tumour diameter of 3 cm or less including ground-glass opacity (JCOG1211): a multicentre, single-arm, confirmatory, phase 3 trial. Lancet Respir Med 2023;11:540-9. [Crossref] [PubMed]
6. Altorki N, Wang X, Kozono D, et al. Lobar or Sublobar Resection for Peripheral Stage IA Non-Small-Cell Lung Cancer. N Engl J Med 2023;388:489-98. [Crossref] [PubMed]
7. Suzuki K, Watanabe SI, Wakabayashi M, et al. A single-arm study of sublobar resection for ground-glass opacity dominant peripheral lung cancer. J Thorac Cardiovasc Surg 2022;163:289-301.e2. [Crossref] [PubMed]
8. Saji H, Okada M, Tsuboi M, et al. Segmentectomy versus lobectomy in small-sized peripheral non-small-cell lung cancer (JCOG0802/WJOG4607L): a multicentre, open-label, phase 3, randomised, controlled, non-inferiority trial. Lancet 2022;399:1607-17. [Crossref] [PubMed]
9. Ito H, Suzuki K, Mizutani T, et al. Long-term survival outcome after lobectomy in patients with clinical T1 N0 lung cancer. J Thorac Cardiovasc Surg 2020;S0022-5223(20)30054-4.
10. Speicher PJ, Gu L, Gulack BC, et al. Sublobar Resection for Clinical Stage IA Non-small-cell Lung Cancer in the United States. Clin Lung Cancer 2016;17:47-55. [Crossref] [PubMed]
11. Shi F, Chen B, Cao Q, et al. Semi-Supervised Deep Transfer Learning for Benign-Malignant Diagnosis of Pulmonary Nodules in Chest CT Images. IEEE Trans Med Imaging 2022;41:771-81. [Crossref] [PubMed]
12. Liu K, Li K, Wu T, et al. Improving the accuracy of prognosis for clinical stage I solid lung adenocarcinoma by radiomics models covering tumor per se and peritumoral changes on CT. Eur Radiol 2022;32:1065-77. [Crossref] [PubMed]
13. Park S, Park G, Lee SM, et al. Deep learning-based differentiation of invasive adenocarcinomas from preinvasive or minimally invasive lesions among pulmonary subsolid nodules. Eur Radiol 2021;31:6239-47. [Crossref] [PubMed]
14. Tsutani Y, Suzuki K, Koike T, et al. High-Risk Factors for Recurrence of Stage I Lung Adenocarcinoma: Follow-up Data From JCOG0201. Ann Thorac Surg 2019;108:1484-90. [Crossref] [PubMed]
15. Kim H, Goo JM, Kim YT, et al. CT-defined Visceral Pleural Invasion in T1 Lung Adenocarcinoma: Lack of Relationship to Disease-Free Survival. Radiology 2019;292:741-9. [Crossref] [PubMed]
16. Zhang H, Liao M, Guo Q, et al. Predicting N2 lymph node metastasis in presurgical stage I-II non-small cell lung cancer using multiview radiomics and deep learning method. Med Phys 2023;50:2049-60. [Crossref] [PubMed]
17. Teng Y, Ran X, Chen B, et al. Pathological Diagnosis of Adult Craniopharyngioma on MR Images: An Automated End-to-End Approach Based on Deep Neural Networks Requiring No Manual Segmentation. J Clin Med 2022;11:7481. [Crossref] [PubMed]
18. Wang Y, Ding Y, Liu X, et al. Preoperative CT-based radiomics combined with tumour spread through air spaces can accurately predict early recurrence of stage I lung adenocarcinoma: a multicentre retrospective cohort study. Cancer Imaging 2023;23:83. [Crossref] [PubMed]
19. Onoda H, Higashi M, Murakami T, et al. Correlation between pleural tags on CT and visceral pleural invasion of peripheral lung cancer that does not appear touching the pleural surface. Eur Radiol 2021;31:9022-9. [Crossref] [PubMed]
20. Hsu JS, Han IT, Tsai TH, et al. Pleural Tags on CT Scans to Predict Visceral Pleural Invasion of Non-Small Cell Lung Cancer That Does Not Abut the Pleura. Radiology 2016;279:590-6. [Crossref] [PubMed]
21. Zhao LL, Xie HK, Zhang LP, et al. Visceral pleural invasion in lung adenocarcinoma ≤3 cm with ground-glass opacity: a clinical, pathological and radiological study. J Thorac Dis 2016;8:1788-97. [Crossref] [PubMed]
22. Imai K, Minamiya Y, Ishiyama K, et al. Use of CT to evaluate pleural invasion in non-small cell lung cancer: measurement of the ratio of the interface between tumor and neighboring structures to maximum tumor diameter. Radiology 2013;267:619-26. [Crossref] [PubMed]
23. Wang F, Lin X, Lin C, et al. Ability of three-dimensional 3-Tesla ultrashort echo time magnetic resonance imaging to display the morphological characteristics of pulmonary nodules: a sensitivity analysis. Quant Imaging Med Surg 2023;13:1792-801. [Crossref] [PubMed]
24. Kim HJ, Cho JY, Lee YJ, et al. Clinical Significance of Pleural Attachment and Indentation of Subsolid Nodule Lung Cancer. Cancer Res Treat 2019;51:1540-8. [Crossref] [PubMed]
25. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]