An interpretable machine learning model for predicting visceral pleural invasion in cT1 lung adenocarcinoma based on habitat analysis

Introduction

Sub-lobar resection is increasingly significant in treating early-stage non-small cell lung cancer (NSCLC). For cT1 patients, sub-lobar resection achieves comparable overall survival (OS) rates to traditional lobectomy, with less trauma and greater pulmonary function preservation (1-6). However, 0.4% to 14.6% of these patients may experience postoperative T staging upgrades, associated with poorer OS outcomes (1,4-6). According to the International Association for the Study of Lung Cancer’s 8th edition lung cancer staging, visceral pleural invasion (VPI) significantly influences T staging for NSCLC tumors under 3 cm and is a negative prognostic indicator (7). In solid nodules with a consolidation-to-tumor ratio (CTR) >0.5, VPI incidence is higher, leading to increased local recurrence rates (10.5% vs. 5.4%) and lower 10-year recurrence-free survival rates (73.2–76.7% vs. 90.5%) (1,3-5,8). Currently, no large-scale randomized controlled trials address whether VPI-positive solid lung nodule patients benefit from sub-lobar resection. VPI status is generally identified after surgery through pathological elastic fiber staining, as there are currently no reliable methods for preoperative or intraoperative evaluation (9). This limitation hinders surgeons from considering VPI during candidate selection for sub-lobar resection.

For lung nodules ≤3 cm, preoperative computed tomography (CT) imaging can partially predict VPI status (10,11). CT images contain rich information, challenging for manual discernment, including shape, intensity, gradient, and texture variations contributing to tumor heterogeneity (12,13). Radiomics-based studies have explored tumor heterogeneity from a macroscopic perspective (14-16). Habitat analysis has been employed to examine regions of interest (ROI) at the voxel level, offering a microscopic perspective on tumor heterogeneity. In fields such as immune therapy resistance, Ki-67 status in ovarian cancer, and chemotherapy responses in breast and esophageal cancers, habitat feature clustering-based predictive models often present superiorly (17-20). Despite advances in machine learning-based prediction models, their interpretability remains under-researched, limiting clinical application (21). Therefore, integrating model interpretability into clinical decision support systems and medical research is essential. Sufficient interpretability enables clinicians to make informed decisions, apply models correctly, and reproduce image-based reasoning and disease diagnosis (22). This study utilizes preoperative CT imaging data for habitat analysis of solid lung nodules ≤3 cm and the peritumoral region’s characteristics. A predictive VPI model is established using machine learning algorithms to explore its relationship with OS. The SHapley Additive exPlanations (SHAP) method is employed to investigate the radiomics model’s interpretability, establishing the relationship between radiomics features and tumor biological characteristics. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2890/rc).

Methods

Study design and population

This retrospective study involved patients diagnosed with invasive pulmonary adenocarcinoma through pathological examination, all of whom underwent lobectomy. These patients were selected based on preoperative CT scans that detected solid lung nodules with a maximum diameter of 3 cm or less. Patients were consecutively enrolled at the Fifth Affiliated Hospital of Sun Yat-sen University (Center 1) between June 2017 and December 2022, and at Sun Yat-sen Memorial Hospital (Center 2) from June 2017 until December 2020 (Figure 1). Following the screening, the study incorporated 509 patients from Center 1 and 216 patients from Center 2 (see Figure 1), all of whom were of East Asian ethnicity. Additionally, a publicly accessible radiogenomics dataset (23) provided data on 211 patients of diverse ethnic backgrounds from North America. After review by the researchers, all lesions in the dataset were confirmed to be purely solid nodules, leading to the inclusion of 77 patients in this study (see Figure 1). The dataset from Center 1 was randomly divided into a training cohort (TC) and a validation cohort (VC) in a 7:3 ratio. The datasets from Center 2 and the public cohort were combined to create the external test cohort (ETC). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Ethical approval was obtained from the Institutional Review Boards of the Fifth Affiliated Hospital of Sun Yat-sen University [No. (2024) K83-1] and Sun Yat-sen Memorial Hospital (No. SYSKY-2025-297-01), and individual consent for this retrospective analysis was waived. The pleural invasion status of all pulmonary nodules was confirmed through elastic fiber staining, as documented in the pathology reports. Detailed follow-up strategies are provided in Appendix 1.

Figure 1 Flow diagram of the study population. Center 1, the Fifth Affiliated Hospital of Sun Yat-sen University; Center 2, Sun Yat-sen Memorial Hospital. CT, computed tomography.

Image preprocessing and segmentation

All CT scans were resampled to isotropic 1×1×1 mm³ voxels using lung window settings (window width: 1,500 HU; window level: −600 HU) for optimal pulmonary parenchyma visualization, following the Fleischner Society guidelines on thoracic imaging. Each tumor ROI was meticulously delineated, followed by an extension of 5 mm around each ROI to incorporate the peritumoral area (Figure 2). In this study, the P5 region was defined to include the entirety of the tumor as well as the adjacent tissue. Should any ROI intersect with non-pulmonary areas, only the segments that included lung tissue were retained. Additional information about the CT protocol and evaluation methods is provided in Appendix 2. For accurate segmentation, ITK-SNAP, a freely accessible software tool (http://www.itksnap.org/pmwiki/pmwiki.php), was utilized. The assessment of each feature was conducted using the intraclass correlation coefficient, with only those features that achieved a value greater than 0.8 allowed to move forward into the next stages of feature selection and model development.

Figure 2 The workflow of this study. Step 1: in the lung window of CT images, ROIs are delineated. An intratumoral mask is drawn along the tumor margin, which is subsequently expanded by 5 mm to achieve an automated delineation of the peritumoral mask. Non-lung tissue portions involved in the mask are then removed. The designation P5 represents the combined intratumoral and peritumoral ROI, while “intra/peri” denotes the combination of the intratumoral ROI and the peritumoral ROI. From the intra-tumor ROI and P5 ROI, 19 habitat features are extracted. Heatmaps of four of these features are presented in the figure. Step 2: habitat features are analyzed using k-means unsupervised clustering. For the intra-tumor analysis, the CH index reached its highest value of 2.19×10⁶ when the number of clusters was set to 3, while the DB index achieved its lowest value of 0.89. However, the Silhouette score reached its maximum of 0.43 when the number of clusters was set to 2. In the P5 analysis, the CH index also peaked at 2.70×10⁵ with 3 clusters, and the BD index reached its lowest value of 1.00. The Silhouette score again attained its highest value of 0.47 when the number of clusters was set to 2. The voxel percentages for subregions H1, H2, and H3 were 50.23%, 36.30%, and 13.47% in the intra-tumor analysis, and 29.81%, 58.74%, and 11.45% in the P5 analysis, respectively. Step 3: a total of 1,834 radiomics features were extracted from each subregion and the peritumoral region. Missing features in the subregions were imputed using the k-nearest neighbors method. Following the fusion of features from the intra-tumor subregions, the features were subjected to selection and modeling, resulting in the establishment of the intra-tumor model. A similar approach was utilized to create the P5 model. By merging the intra-tumor features with the peritumoral features, the same methodology was applied to establish the intra/peri model. CH, Calinski-Harabasz; CT, computed tomography; DB, Davies-Bouldin; GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run-length matrix; LightGBM, light gradient boosting machine; ROI, region of interest; SHAP, SHapley Additive exPlanations.

Habitat sub-region clustering and image cropping

To investigate tumor and peritumoral heterogeneity, the intra-tumoral region and P5 zone were delineated. Radiomic features were extracted from a 5×5×5 voxel matrix centered around each voxel (Figure 2), incorporating 19 distinct features (Appendix 3). An unsupervised approach employing the K-means clustering algorithm was applied to segment voxels from normalized images into distinct sub-regions. To identify the most appropriate number of clusters, the clustering results were evaluated for cluster counts ranging from 2 to 9 using metrics such as the Calinski-Harabasz (CH) index, Davies-Bouldin (DB) index, and Silhouette score to measure the quality of the clustering.

Feature extraction and model ensemble

A total of 1,834 radiomic features were systematically extracted from each sub-region through the use of Pyradiomics in a Python (v 3.7), as described in Appendix 4. Receiver operating characteristic (ROC) curves for the models were developed utilizing the light gradient boosting machine (LightGBM) algorithm. Model performance was assessed and validated using three distinct cohorts: TC, VC, and ETC. The evaluation relied on key indicators such as the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value to comprehensively measure effectiveness. The determination of AUCs and their corresponding threshold values for each model was achieved through the application of the Youden index, with a detailed account of the threshold extraction procedure provided in Appendix 5.

Model interpretability with SHAP

To understand radiomic features in radiomics models, we used SHAP, a game theory-based method for interpreting tree model outputs. SHAP provides both global and local interpretability, helping clinicians understand machine learning results. It quantifies each feature’s contribution to changes in output probability, prioritizing features by importance. SHAP value plots visually represent the positive and negative contributions of features. We used a summary SHAP plot to showcase the most critical features in the final model, enhancing the transparency and interpretability of the results.

Statistical analysis

Statistical analyses were conducted using SPSS version 26.0 and Python version 3.7. For variables following a normal distribution, the z-test was applied, whereas variables that did not meet this criterion were analyzed using the Wilcoxon test. Categorical variables were examined using either the Chi-squared test or Fisher’s exact test, depending on the data characteristics. To identify significant predictors, multivariable logistic regression was conducted employing a forward stepwise selection approach. The comparison of Kaplan-Meier survival curves was carried out with the log-rank test. It is stipulated that P<0.05 indicates a significant statistical difference. Comparative analysis of models was achieved using the DeLong test, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Model effectiveness was evaluated using decision curve analysis (DCA) to examine clinical utility and calibration curves to assess the agreement between predicted and observed outcomes.

Results

Patient characteristics

Based on the defined inclusion and exclusion criteria (refer to Figure 1), the study ultimately included 843 patients. Of these, 509 patients from Center 1 were randomly split into training and validation groups at a ratio of 7:3. Center 2 contributed 216 patients, and an extra 77 cases from an external database were added, together constituting the ETC. Comprehensive patient characteristics are detailed in Table 1. Additionally, Kaplan-Meier survival analysis was performed, illustrating OS curves for the three cohorts as well as for each center, which are presented in Figure S1. The median follow-up time was 42 months [95% confidence interval (CI): 40.130–43.870] for Center 1, 53 months (95% CI: 51.607–58.393) for Center 2, and 45 months (95% CI: 40.701–49.299) for the external database. The mean survival time was 81.28 months (95% CI: 78.381–84.70) for Center 1, 84.76 months (95% CI: 81.527–87.984) for Center 2, and 83.17 months (95% CI: 73.1387–93.206) for the external database. The ability of clinical features to predict VPI status and OS was limited, demonstrating low predictive accuracy (see Appendix 6, Figure S2). This may be attributed to the fact that only the Spiculated sign and Vacuole sign among the included clinical features were found to be statistically significant in the multivariate analysis (Table S1). Furthermore, these two imaging features are subject to subjective interpretation, which may introduce a certain degree of noise. Additionally, the underlying mechanism of VPI is highly complex, potentially involving tumor invasiveness, the tumor microenvironment, and genomic alterations, which cannot be comprehensively reflected by these two features. As a result, clinical models were omitted from the later stages of model development.

Sub-region cluster and feature extraction

Habitat features were derived from both the intra-tumor and P5 regions. Subsequently, unsupervised K-means clustering was performed independently on these areas, with clustering quality assessed using CH index, DB index, and Silhouette score. The analysis identified three as the optimal cluster number within the TC (Figure 2), resulting in the segmentation of the intra-tumoral and P5 regions into three distinct subregions each. From these subregions, along with the peri-tumor region, a total of 1,834 radiomics features were extracted. For samples missing subregion features, imputation was carried out utilizing the K-nearest neighbors algorithm. After feature scaling and selection (Appendix 3), features with non-zero coefficients were retained. The modeling process was then carried out using LightGBM. The intra/peri model was composed of various subregions from both the intra-tumor and peri-tumor areas.

VPI status prediction model and OS stratification

The model incorporating peri-tumor features demonstrated superior performance in predicting VPI status. Within the ETC (Figure 3 and Table 2), the P5 model demonstrated superior predictive ability, attaining the highest AUC of 0.787 (95% CI: 0.7275–0.8465) as detailed in Table 2. In contrast, the intra-tumor model showed the weakest performance, achieving an AUC of just 0.698 (95% CI: 0.6334–0.7627). The AUC of the intra/peri-model was situated between the two. The DeLong test indicated a statistically significant improvement in the predictive performance of the P5-model compared to the intra-model, with P=0.008. The NRI test also revealed that the P5 model outperformed the intra-model by 0.113, and compared to the intra/peri-model, the NRI value reached 0.044. Although the P5-model and intra/peri-model exhibited increases in IDI values of 0.083 and 0.068, respectively, indicating a notable improvement in model accuracy, the P values did not show statistical significance. Calibration curves demonstrated that the predictions of the P5-model were more consistent with the actual outcomes. DCA curve analysis indicated that the P5-model provided a higher net benefit across a wide range of threshold values. The results for TC and VC are presented in Figure S3. The prediction results for VPI status in the ETC (Center 2 and external database) are presented in Table S2.

Figure 3 The predictive capability for VPI status and OS stratification in the ETC for each model. (A) The ROC curves reveal that the P5-model exhibits the highest predictive performance, with an AUC of 0.787 (95% CI: 0.728–0.846). Calibration curves and DCA for each model are presented in the accompanying figure. The P values for both the Delong test and IDI are indicated in the figure. Additionally, NRI analysis indicates that the P5-model demonstrates a positive improvement over the other models. Compared to the intra-model, the NRI for the P5-model and the intra/peri-model were found to be 0.113 and 0.044, respectively. The NRI for the P5-model relative to the intra/peri-model was also determined to be 0.068. The IDI test indicated that the IDI values for the P5-model and the intra/peri-model were 0.083 and 0.068, respectively, compared to the intra-model; however, the differences were not statistically significant. (B) Kaplan-Meier analysis demonstrated that in the test cohort, the P5-model and intra/peri-model exhibited better patient stratification capabilities, with P values of 0.042 and 0.022, respectively. Although the intra-model also showed a trend for stratification, the difference was not statistically significant. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ETC, external test cohort; IDI, integrated discrimination improvement; NRI, net reclassification improvement; OS, overall survival; ROC, receiver operating characteristic; VPI, visceral pleural invasion.

The Youden index for each model was used as the cutoff value to stratify patients into high-risk and low-risk groups based on OS. The results of the Kaplan-Meier survival analysis indicated that, in the ETC, the P5-model and the intra/peri-model, which incorporated peri-tumor regions, effectively stratified patients into high-risk and low-risk groups, with P values of 0.042 and 0.022, respectively. A trend towards stratification was also observed in the intra-model; however, this difference was not statistically significant, with a P value of 0.120. The results for the TC and VC are presented in Figure S4.

Model interpretability

In the intra-, P5-, and intra/peri-model, the final number of features included in model construction was 10, 12, and 11, respectively (Appendix 3). The summary plot illustrates global interpretability (Figure 4). As shown, in the intra-model, the most important feature identified was Wavelet_LLL_NGTDM_Coarseness_h1; in the P5-model, it was Gradient_Firstorder_RobustMeanAbsoluteDeviation_h2; and in the Intra/Peri-model, wavelet_HLH_GLCM_InverseVariance emerged as the most significant feature. The force plot and waterfall plot provided insights into local interpretability (interpretability for individual samples) (Figure 5). As depicted, for two samples with a similar degree of contact with the visceral pleura but differing VPI statuses, the importance of individual features varied. For each predicted sample, positive SHAP values indicated an increased VPI risk, while negative values indicated a decreased risk. Within each sample, positive SHAP values for individual features signified a contribution to VPI risk, whereas negative values indicated a mitigating effect. The larger the SHAP value, the more substantial its contribution.

Figure 4 Summary plot delineating the influence of features on the model’s decision-making process and the interactions among the radiomics features within the models. Positive SHAP values denote a heightened likelihood of VPI prediction, with larger values reflecting a greater risk. Each point on the graph signifies a prediction for a specific patient. SHAP, SHapley Additive exPlanations; VPI, visceral pleural invasion.

Figure 5 Force plot and waterfall plot showing local interpretability. (A) The figure illustrates that in a VPI-positive patient, the importance of these features varies across different models. (B) The figure shows the local interpretability for a VPI-negative patient. Interestingly, the contribution of each feature to the model varies depending on the VPI status. However, in the analysis of the two randomly selected samples in this study, regardless of VPI status, the feature Gradient_Firstorder_RobustMeanAbsoluteDeviation_h2 consistently had the highest contribution in the P5 model, with a similar directional influence on the model. Here, h1, h2, and h3 represent the sub-regions in the respective models, while p5 denotes the tumor periphery region. VPI, visceral pleural invasion.

Discussion

The extended follow-up of the Japan Clinical Oncology Group (JCOG) 0201 study has identified VPI as a significant risk factor for recurrence in stage I lung adenocarcinoma (hazard ratio: 2.17; 95% CI: 1.23–3.81). Particularly in nodules with a CTR greater than 0.5 and solid nodules, the prognosis is less favorable. Further subgroup analyses revealed that patients with VPI have a significantly reduced 10-year recurrence-free survival rate compared to those without VPI, with rates of 64.0% versus 87.2%, respectively (24). Currently, VPI status is predominantly determined postoperatively through pathological elastic fiber staining, with no available preoperative or intraoperative diagnostic methods (9). Effective preoperative or intraoperative techniques to inform surgeons of the VPI status of pulmonary nodules remain unavailable. This study leveraged preoperative CT imaging data to devise a novel non-invasive approach for predicting the VPI status of solid pulmonary nodules. Habitat characteristics of both the tumor and the peritumoral region were analyzed, segmenting the ROIs into various sub-regions to extract unique features from each. Predictive models were constructed based on different combinations of ROI regions. The findings indicated that, within the ETC, the P5 model achieved an AUC value of 0.787 (95% CI: 0.7275–0.8465), surpassing the intra-model, while the intra/peri-model also demonstrated an improvement. Additionally, the P5- and intra/peri-models, which incorporated features from the peritumoral region, exhibited robust patient stratification capabilities. SHAP analysis underscored the significant role of peritumoral region features in the final model. Conversely, the intra-model, which excluded peritumoral features, displayed lower predictive and prognostic stratification capabilities. This study underscores the utility and generalization potential of the model across multinational and multiethnic independent datasets, highlighting its applicability in diverse clinical settings.

Conventionally, the presence of VPI has been preliminarily predicted to some extent by certain CT indicators, such as visceral pleural traction and bulging. In the presence of VPI, distinct CT morphological characteristics such as pleural retraction, adhesion, and lobulation tend to emerge, thereby enhancing the predictive accuracy of these imaging features (11,25,26). According to Imai et al., the likelihood of pleural invasion is positively correlated with the degree of contact between the lesion and the pleura (27). Despite strict definitions for these signs, subjective judgment can sometimes affect predictions, leading to unsatisfactory sensitivity and specificity in prediction (26,28,29). Kim et al. found that the accuracy levels of CT features in predicting VPI status ranged from 62.7% to 72.3% (28). Currently, some studies have employed radiomics or deep learning methods to investigate the VPI status of pulmonary nodules; however, few multicenter studies have included external validation data (30-34). However, these studies included ground-glass nodules with a CTR of less than 0.5, a subset of patients that rarely experiences upgrades in clinical T staging. Additionally, these patients may benefit from both lobectomy and sublobar resection, with no significant differences in long-term survival reported (1-6). To date, there have been no studies concentrating exclusively on nodules with a CTR exceeding 0.5. Additionally, prior research employing deep learning or radiomics techniques to predict VPI status generally shows AUC scores below 0.70 when tested on external validation datasets (33,35). Lin and colleagues constructed a convolutional neural network using a 3D-ResNet-9 architecture to predict VPI status. In the external VC, the clinical model achieved an AUC of 0.66, the deep learning model reached an AUC of 0.62, and the integrated model combining both approaches attained a modest AUC of 0.69 (35). Choi and colleagues created a deep learning model aimed at predicting VPI, obtaining an AUC of 0.75 that was similar to the diagnostic accuracy of radiologists. Nevertheless, their study population comprised patients with more advanced disease stages, specifically those classified as cT1 or higher (32). To derive more valuable insights from the current CT imaging data, our research focuses on analyzing sub-regions within the tumor to capture detailed intratumoral features. Features from the peritumoral region are also extracted. By combining these areas, ROI are created, and habitat analysis is performed to further define sub-regions for detailed study. No similar research has been reported to date.

Current research results indicate that, in predicting VPI, first-order features and gray level size zone matrix (GLSZM) features are comparatively important among the final features included in the modeling, while 2D/3D shape features and other gray level features contribute minimally. In terms of image transformations, wavelet, exponential, and square transformations are considered significant, with local binary pattern 2D/3D transformations rarely observed, and features based on the original image are also infrequently reported (30,32-34). These findings from single-center studies align with those of the current research. However, there are no existing multi-center studies that report on how these features function within the model or their contribution levels. Additionally, the contribution of intratumoral heterogeneity and peri-tumoral region characteristics to the overall model performance remains unexamined. In this study, the SHAP analysis results are visually indicated to reveal that the peritumoral region exhibits distinct characteristics across pulmonary nodules with varying VPI statuses, which are considered crucial for predicting VPI within the model. Furthermore, the SHAP analysis conducted on samples with different VPI conditions reveals that the same features can exert contrasting effects across different samples, such as Wavelet_LHL_Firstorder_Maximum_h2, although their significance remains relatively stable. This comprehensive approach provides a nuanced understanding of the role of peritumoral and intratumoral regions in the context of VPI prediction.

These radiomic features demonstrate their effectiveness in predicting VPI status from both physiological and clinical perspectives. Wavelet_LLL_NGTDM_Coarseness_h1, which describes the coarseness of texture, is likely associated with the increased complexity and heterogeneity of the tumor tissue, which are closely linked to enhanced tumor invasiveness. During the invasion process, the structural irregularities within the tumor increase, and the variations in gray-level intensities become more pronounced, enabling this feature to effectively capture the potential aggressiveness of intratumoral characteristics. Gradient_Firstorder_RobustMeanAbsoluteDeviation_h2, which measures the dispersion of gray-level gradient changes, is thought to reflect the density differences between intratumoral and peritumoral regions, particularly at the interface where the tumor approaches or invades the visceral pleura. In VPI-positive cases, tumor invasion may result in more pronounced gradient variations at this tumor-pleura boundary, allowing this feature to capture critical radiological differences in these regions. Wavelet_HLH_GLCM_InverseVariance, a texture feature assessing the uniformity of gray-level patterns, is believed to reflect the disruption of microstructural organization in both intratumoral and peritumoral regions during the invasion process. When VPI occurs, peritumoral tissue is modified and invaded, leading to significant alterations in texture patterns, which this feature is particularly sensitive to. Collectively, these radiomic features, by capturing intratumoral heterogeneity, density variations at the tumor-pleura interface, and microstructural disruption in peritumoral regions, provide valuable insights into the physiological and pathological characteristics associated with VPI status, forming a robust basis for the accurate prediction of VPI.

This research is subject to several limitations that should be acknowledged. Firstly, the sample size employed is relatively small, and the retrospective design inherently introduces a degree of selection bias, impacting the sample’s representativeness. Secondly, for a comprehensive evaluation of the study’s generalizability and robustness, it is crucial to emphasize the effects of varying CT scanner models, institutional protocols, and scanning parameters on the outcomes. Thirdly, manual segmentation of ROI was utilized, presenting persistent challenges such as being time-consuming, subjective, poorly reproducible, and difficult to scale. Moreover, the methodologies implemented demand significant computational resources and are time-intensive, thereby restricting an exhaustive comparison across different peritumoral region ranges (e.g., 1, 3, 5 mm). This limitation constrains the methods’ broader applicability and potential for clinical adoption. Addressing these factors in future research will be essential for enhancing the robustness and clinical relevance of the findings.

Conclusions

This study utilized both intratumoral and peritumoral regions for habitat analysis, developing a machine learning model to predict VPI status. By combining information from the peritumoral areas with internal tumor features, a comprehensive assessment of tumor heterogeneity was achieved. SHAP analysis effectively highlighted the importance of each feature in the model. The model demonstrates strong predictive accuracy for VPI status in cT1 stage lung adenocarcinoma patients and effective prognosis stratification. This non-invasive method offers valuable preoperative decision support for surgeons, significantly enhancing clinical decision-making and patient outcomes.

Acknowledgments

The authors would like to extend the gratitude to the OneKey platform (https://github.com/OnekeyAI-Platform) for its invaluable contribution to this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2890/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2890/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2890/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 and its subsequent amendments. The study was approved by the Institutional Review Boards of the Fifth Affiliated Hospital of Sun Yat-sen University [No. (2024) K83-1] and Sun Yat-sen Memorial Hospital (No. SYSKY-2025-297-01), 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. Aokage K, Suzuki K, Saji H, Wakabayashi M, Kataoka T, Sekino Y, 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]
2. Altorki N, Wang X, Kozono D, Watt C, Landrenau R, Wigle 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]
3. Suzuki K, Watanabe SI, Wakabayashi M, Saji H, Aokage K, Moriya Y, Yoshino I, Tsuboi M, Nakamura S, Nakamura K, Mitsudomi T, Asamura HWest Japan Oncology Group and Japan Clinical Oncology Group. 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]
4. Saji H, Okada M, Tsuboi M, Nakajima R, Suzuki K, Aokage K, 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]
5. Ito H, Suzuki K, Mizutani T, Aokage K, Wakabayashi M, Fukuda H, Watanabe SI; Japan Clinical Oncology Group Lung Cancer Surgical Study Group. Long-term survival outcome after lobectomy in patients with clinical T1 N0 lung cancer. J Thorac Cardiovasc Surg 2020;S0022-5223(20)30054-4.
6. Speicher PJ, Gu L, Gulack BC, Wang X, D’Amico TA, Hartwig MG, Berry MF. Sublobar Resection for Clinical Stage IA Non-small-cell Lung Cancer in the United States. Clin Lung Cancer 2016;17:47-55. [Crossref] [PubMed]
7. Goldstraw P, Chansky K, Crowley J, Rami-Porta R, Asamura H, Eberhardt WE, Nicholson AG, Groome P, Mitchell A, Bolejack VInternational Association for the Study of Lung Cancer Staging and Prognostic Factors Committee, Advisory Boards, and Participating Institutions. International Association for the Study of Lung Cancer Staging and Prognostic Factors Committee Advisory Boards and Participating Institutions. The IASLC Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groupings in the Forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol 2016;11:39-51. [Crossref] [PubMed]
8. Nakagawa K, Watanabe SI, Wakabayashi M, Yotsukura M, Mimae T, Hattori A, et al. Risk Factors for Locoregional Relapse After Segmentectomy: Supplementary Analysis of the JCOG0802/WJOG4607L Trial. J Thorac Oncol 2025;20:157-66. [Crossref] [PubMed]
9. Travis WD, Brambilla E, Noguchi M, Nicholson AG, Geisinger KR, Yatabe Y, et al. International association for the study of lung cancer/american thoracic society/european respiratory society international multidisciplinary classification of lung adenocarcinoma. J Thorac Oncol 2011;6:244-85. [Crossref] [PubMed]
10. Sun Q, Li P, Zhang J, Yip R, Zhu Y, Yankelevitz DF, Henschke CI. CT Predictors of Visceral Pleural Invasion in Patients with Non-Small Cell Lung Cancers 30 mm or Smaller. Radiology 2024;310:e231611. [Crossref] [PubMed]
11. Onoda H, Higashi M, Murakami T, Tao H, Yokoyama S, Kunihiro Y, Kawano R, Tanabe M, Tanaka N, Matsumoto T. 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]
12. Bera K, Velcheti V, Madabhushi A. Novel Quantitative Imaging for Predicting Response to Therapy: Techniques and Clinical Applications. Am Soc Clin Oncol Educ Book 2018;38:1008-18. [Crossref] [PubMed]
13. Han F, Wang H, Zhang G, Han H, Song B, Li L, Moore W, Lu H, Zhao H, Liang Z. Texture feature analysis for computer-aided diagnosis on pulmonary nodules. J Digit Imaging 2015;28:99-115. [Crossref] [PubMed]
14. Hou M, Zhou L, Sun J. Deep-learning-based 3D super-resolution MRI radiomics model: superior predictive performance in preoperative T-staging of rectal cancer. Eur Radiol 2023;33:1-10. [Crossref] [PubMed]
15. Fan M, Liu Z, Xu M, Wang S, Zeng T, Gao X, Li L. Generative adversarial network-based super-resolution of diffusion-weighted imaging: Application to tumour radiomics in breast cancer. NMR Biomed 2020;33:e4345. [Crossref] [PubMed]
16. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
17. Syed AK, Whisenant JG, Barnes SL, Sorace AG, Yankeelov TE. Multiparametric Analysis of Longitudinal Quantitative MRI data to Identify Distinct Tumor Habitats in Preclinical Models of Breast Cancer. Cancers (Basel) 2020;12:1682. [Crossref] [PubMed]
18. Wang X, Xu C, Grzegorzek M, Sun H. Habitat radiomics analysis of pet/ct imaging in high-grade serous ovarian cancer: Application to Ki-67 status and progression-free survival. Front Physiol 2022;13:948767. [Crossref] [PubMed]
19. Vaidya P, Bera K, Patil PD, Gupta A, Jain P, Alilou M, Khorrami M, Velcheti V, Madabhushi A. Novel, non-invasive imaging approach to identify patients with advanced non-small cell lung cancer at risk of hyperprogressive disease with immune checkpoint blockade. J Immunother Cancer 2020;8:e001343. [Crossref] [PubMed]
20. Xie C, Yang P, Zhang X, Xu L, Wang X, Li X, Zhang L, Xie R, Yang L, Jing Z, Zhang H, Ding L, Kuang Y, Niu T, Wu S. Sub-region based radiomics analysis for survival prediction in oesophageal tumours treated by definitive concurrent chemoradiotherapy. EBioMedicine 2019;44:289-97. [Crossref] [PubMed]
21. Son J, Lee SE, Kim EK, Kim S. Prediction of breast cancer molecular subtypes using radiomics signatures of synthetic mammography from digital breast tomosynthesis. Sci Rep 2020;10:21566. [Crossref] [PubMed]
22. Elshawi R, Al-Mallah MH, Sakr S. On the interpretability of machine learning-based model for predicting hypertension. BMC Med Inform Decis Mak 2019;19:146. [Crossref] [PubMed]
23. Bakr S, Gevaert O, Echegaray S, Ayers K, Zhou M, Shafiq M, Zheng H, Benson JA, Zhang W, Leung ANC, Kadoch M, Hoang CD, Shrager J, Quon A, Rubin DL, Plevritis SK, Napel S. A radiogenomic dataset of non-small cell lung cancer. Sci Data 2018;5:180202. [Crossref] [PubMed]
24. Tsutani Y, Suzuki K, Koike T, Wakabayashi M, Mizutani T, Aokage K, Saji H, Nakagawa K, Zenke Y, Takamochi K, Ito H, Aoki T, Okami J, Yoshioka H, Okada M, Watanabe SIJapan Clinical Oncology Group Lung Cancer Surgical Study Group (JCOG-LCSSG). High-Risk Factors for Recurrence of Stage I Lung Adenocarcinoma: Follow-up Data From JCOG0201. Ann Thorac Surg 2019;108:1484-90. [Crossref] [PubMed]
25. Zhao LL, Xie HK, Zhang LP, Zha JY, Zhou FY, Jiang GN, Chen C. 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]
26. Hsu JS, Han IT, Tsai TH, Lin SF, Jaw TS, Liu GC, Chou SH, Chong IW, Chen CY. 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]
27. Imai K, Minamiya Y, Ishiyama K, Hashimoto M, Saito H, Motoyama S, Sato Y, Ogawa J. 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]
28. Kim H, Goo JM, Kim YT, Park CM. CT-defined Visceral Pleural Invasion in T1 Lung Adenocarcinoma: Lack of Relationship to Disease-Free Survival. Radiology 2019;292:741-9. [Crossref] [PubMed]
29. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
30. Zha X, Liu Y, Ping X, Bao J, Wu Q, Hu S, Hu C. A Nomogram Combined Radiomics and Clinical Features as Imaging Biomarkers for Prediction of Visceral Pleural Invasion in Lung Adenocarcinoma. Front Oncol 2022;12:876264. [Crossref] [PubMed]
31. Yuan M, Liu JY, Zhang T, Zhang YD, Li H, Yu TF. Prognostic Impact of the Findings on Thin-Section Computed Tomography in stage I lung adenocarcinoma with visceral pleural invasion. Sci Rep 2018;8:4743. [Crossref] [PubMed]
32. Choi H, Kim H, Hong W, Park J, Hwang EJ, Park CM, Kim YT, Goo JM. Prediction of visceral pleural invasion in lung cancer on CT: deep learning model achieves a radiologist-level performance with adaptive sensitivity and specificity to clinical needs. Eur Radiol 2021;31:2866-76. [Crossref] [PubMed]
33. Kong L, Xue W, Zhao H, Zhang X, Chen S, Ren D, Duan G. Predicting pleural invasion of invasive lung adenocarcinoma in the adjacent pleura by imaging histology. Oncol Lett 2023;26:438. [Crossref] [PubMed]
34. Liu S, Li H, Xiao X, Huo W, Li X, Zhong B, Gan X, Cao Q. A deep learning approach for predicting visceral pleural invasion in cT1 lung adenocarcinoma. J Thorac Dis 2024;16:5675-87. [Crossref] [PubMed]
35. Lin X, Liu K, Li K, Chen X, Chen B, Li S, Chen H, Li L. A CT-based deep learning model: visceral pleural invasion and survival prediction in clinical stage IA lung adenocarcinoma. iScience 2024;27:108712. [Crossref] [PubMed]