Contrast-enhanced computed tomography-based intratumoral and peritumoral radiomics for identifying malignancy in pulmonary ground-glass nodules

Introduction

Lung cancer remains one of the most prevalent malignancies and a leading cause of cancer-related mortality worldwide. According to data from the American Cancer Society, lung cancer ranks first in cancer-related deaths (1), highlighting an urgent need for early detection and intervention. A critical radiological indicator of early-stage lung cancer is the pulmonary ground-glass nodules (GGNs), which appears on computed tomography (CT) images as a appear as hazy, slightly increased attenuation areas with preserved bronchial and vascular structures (2). These nodules represent a spectrum of pathological entities, ranging from benign inflammatory or fibrotic changes to pre-invasive and early-stage adenocarcinomas, and are categorized as pure GGNs and mixed GGNs.

The clinical significance of GGNs has been amplified in recent years due to the growing use of low-dose CT screening and heightened public awareness of lung health, leading to a marked increase in their detection rate. This rise presents a double-edged sword: while the timely resection of malignant GGNs can lead to excellent long-term survival rates of 95–100% at 5 and 10 years (3-5), overtreatment of benign lesions can cause considerable harm. Unnecessary surgical intervention may result in significant physical trauma, increased healthcare costs, and a range of postoperative complications, including chronic chest pain, paresthesia, and respiratory dysfunction (5). Consequently, the challenge for clinicians has become a delicate balance between ensuring curative treatment for malignancies and avoiding the iatrogenic consequences of overtreating benign nodules. Therefore, more accurate image evaluation is of paramount importance for the effective differentiation of benign and malignant GGNs.

Traditional evaluation methods based on morphological characteristics—such as nodule size, shape, and margins—often yield inconsistent results, particularly for slowly progressing or atypically shaped GGNs (6). In this context, radiomics has emerged as a promising tool for extracting high-dimensional quantitative features from medical images, offering new opportunities for tumor characterization and predictive modeling (7). Several studies have applied radiomics to GGNs with promising results (8-10), yet most have focused exclusively on intratumoral features, while peritumoral information remains relatively underexplored.

The peritumoral region is not merely a passive surrounding area but is often influenced by tumor infiltration and associated microenvironmental changes. Radiomics features derived from peritumoral tissues may capture important biological processes such as tumor growth, inflammatory response, cellular migration, and angiogenesis (11). Prior studies incorporating both intratumoral and peritumoral features to evaluate the invasiveness and histologic subtypes of GGNs have primarily relied on non-contrast or low-dose CT data (11-13). However, Tumor vasculature plays a critical role in the progression and metastasis of tumors, antitumor immunity, drug delivery, and resistance to therapies (14). contrast-enhanced CT provides additional insights into tumor vascularity, as enhancement patterns reflect microvascular density and neovascularization (15), potentially influencing both intratumoral and peritumoral feature expression and, in turn, the diagnostic performance of radiomics models.

This study aimed to systematically investigate the radiomics characteristics of both intratumoral and peritumoral regions based on contrast-enhanced CT imaging. By developing a comprehensive and efficient radiomics model, we seek to enhance the diagnostic accuracy for distinguishing benign from malignant GGNs. Furthermore, we critically evaluate the model’s performance in comparison with existing approaches and explore the clinical applicability and limitations of contrast-enhanced CT-based radiomics in early lung cancer diagnosis, thereby contributing to the advancement of precision imaging and individualized treatment planning. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1608/rc).

Methods

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the institutional ethics committee of Ganzhou Cancer Hospital (No. 2025247), and individual consent for this retrospective analysis was waived.

Patient and data collection

We initially identified 240 patients who underwent pulmonary nodule surgery at Ganzhou Cancer Hospital between 2019 and 2023. Among these, a total of 147 patients with GGNs were retrospectively enrolled, comprising 87 patients with malignant nodules and 60 with benign nodules. Patients were randomly divided into a training set (n=103) and a testing set (n=44) in a 7:3 ratio.

The inclusion criteria were as follows: (I) histopathologically confirmed diagnosis of benign or malignant GGNs following surgical resection; (II) preoperative contrast-enhanced chest CT or non-contrast plus contrast-enhanced CT performed; and (III) an interval of no more than one month between CT examination and surgery. The exclusion criteria were as follows: (I) CT images with severe artifacts that impeded lesion evaluation; (II) incomplete clinical or imaging data; (III) patients who received chemotherapy or radiotherapy prior to CT examination; and (IV) patients with other concomitant malignancies. Figure 1 displays the patient enrollment flowchart.

Figure 1 The flowchart shows the patient enrollment process. CT, computed tomography.

Clinical and imaging data were retrieved from the hospital’s electronic medical record system and picture archiving and communication system (PACS). Clinical information included age, sex, and postoperative pathological diagnosis. Two radiologists, each with 8 years of experience in thoracic imaging, independently reviewed and evaluated the imaging features while blinded to pathological results. The analyzed imaging features included lesion location, morphology, lobulation, spiculation, vascular convergence, and average CT attenuation values. In cases of inter-reader disagreement, final consensus was achieved through discussion.

Image acquisition and processing

CT examinations were performed using 64-slice (uCT760) and 320-slice (uCT960+) scanners (United Imaging Healthcare, Shanghai, China). Scanning parameters were standardized with a tube voltage of 120 kV, automatic tube current modulation, and a matrix size of 512×512. Iodine contrast agent (e.g., Omnipaque, 300 mgI/mL, GE Healthcare, Boston, USA) was administered by power injector at a flow rate of 3.0 to 3.5 mL/s and the contrasted images were acquired after the injection of iodine contrast agent at 30–40 seconds. No adverse reactions occurred after the injection of contrast agent in all patients. Tumor segmentation was performed in a two-step process. Initially, a radiologist with 8 years of experience in thoracic imaging manually delineated the intratumoral region using ITK-SNAP software (version 3.8.0; www.itksnap.org). The segmentation was subsequently reviewed and, if necessary, revised by another senior radiologist with equivalent experience. Discrepancies were resolved by consensus discussion. The peritumoral region was generated automatically as a three-dimensional (3D) 1-mm isotropic expansion surrounding the segmented tumor using 3D Slicer software (version 5.6.2; www.slicer.org), with the segmentation process of both the intratumoral and peritumoral regions shown in Figure 2.

Figure 2 The segmentation process of both the intratumoral and peritumoral regions. (A) Original GGN image; (B,C) ROIs of intratumoral and peritumoral regions; (D) 3D segmentation image. 3D, three-dimensional; GGN, ground-glass nodule; ROIs, regions of interest.

Feature extraction

To minimize the influence of scanner heterogeneity and enhance feature reproducibility, all images underwent pre-processing prior to feature extraction. This included gray-level normalization (Z-score standardization) and isotropic resampling to a voxel size of 1×1×1 mm³ using B-spline interpolation. Radiomics feature extraction was conducted using PyRadiomics (version 3.0), an open-source Python-based package (16). Features were extracted separately from the intratumoral and peritumoral regions of interest (ROIs) based on the contrast-enhanced CT images. A total of 1,409 features were derived from each region per patient, encompassing first-order statistics, shape-based descriptors, and multiple texture features including gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), gray-level dependence matrix (GLDM), and neighborhood gray-tone difference matrix (NGTDM). In addition, 14 types of image filters were applied to generate high-dimensional features, including Laplacian of Gaussian, wavelet decomposition, and local binary patterns.

Feature dimensionality reduction and model construction

For radiomics features, a two-step dimensionality reduction strategy was employed. First, the minimum redundancy maximum relevance (mRMR) algorithm was applied to select the top 20 features with the highest relevance and lowest redundancy. These features were further refined using least absolute shrinkage and selection operator (LASSO) regression to identify the final feature subset for model construction. Separate logistic regression models were developed based on intratumoral and peritumoral features. For clinical variables, univariate logistic regression was first performed to identify candidate predictors (P<0.05), which were subsequently included in a multivariate logistic regression model to determine the final set of independent predictors. A clinical model was then constructed based on these variables. Finally, the radiomics model (intratumoral or peritumoral) with superior performance was combined with the clinical model using logistic regression to develop the integrated model. The calibration curve was employed to assess the agreement between the predicted probabilities of benignancy and malignancy for GGNs generated by the combined model and their actual pathological status, both in the training and testing sets. The net clinical benefit of each model was compared using decision curve analysis (DCA).

Statistical analysis

All statistical analyses were performed using IBM SPSS Statistics (version 26.0; IBM Corp., Armonk, NY, USA) and R software (R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were presented as mean ± standard deviation or median with interquartile range (IQR), depending on the distribution, and compared using independent samples t-tests or Wilcoxon rank-sum tests, respectively. Categorical variables were summarized as counts and percentages, and group comparisons were conducted using the Chi-squared test or Fisher’s exact test as appropriate. A two-tailed P<0.05 was considered statistically significant. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and its 95% confidence interval (CI), along with sensitivity, specificity, and accuracy. The DeLong test was used to compare the AUC values among the four models.

Results

Patient and imaging characteristics

A total of 147 patients with GGNs were included in this study, comprising 46 males and 101 females. Among them, 87 patients had malignant GGNs and 60 had benign GGNs. Patients were randomly assigned to a training set (n=103) and a testing set (n=44) at a 7:3 ratio. No significant differences in baseline clinical or imaging characteristics were observed between the training and testing sets (Table 1).

Univariate logistic regression analysis identified GGN type, spiculation, lobulation, carcinoembryonic antigen (CEA), and pro-gastrin-releasing peptide (ProGRP) as significant factors in distinguishing malignant from benign GGNs (Table 2). However, multivariate logistic regression analysis revealed that only lobulation [odds ratio (OR), 5.17; 95% CI: 1.99–13.42), CEA (OR, 1.72; 95% CI: 1.10–2.70], and ProGRP (OR, 1.04; 95% CI: 1.00–1.07) remained independent predictors of malignancy (Table 3). These three variables were subsequently incorporated into the clinical model and the clinical-radiomics combined model.

Using the described feature extraction approach, a total of 1,409 radiomics features were extracted from both intratumoral and peritumoral regions of the GGNs. mRMR was initially applied for dimensionality reduction, followed by LASSO regression for feature selection within the intratumoral and peritumoral models, respectively (Figure 3). Eight features were ultimately selected for the intratumoral model, and two features were selected for the peritumoral model.

Figure 3 Feature dimensionality reduction and selection using LASSO regression for intratumoral and peritumoral models. (A) Tuning parameter λ selection and coefficient path plot for the LASSO regression in the intratumoral model; (B) tuning parameter λ selection and coefficient path plot for the LASSO regression in the peritumoral model. LASSO, least absolute shrinkage and selection operator.

Receiver operating characteristic (ROC) curves for the four models—intratumoral, peritumoral, clinical, and clinical-intratumoral combined—are presented in Figure 4. A statistically significant difference was observed between the intratumoral and peritumoral models (D =3.36, df =171.28, P=0.001), whereas no significant difference was found between the intratumoral and clinical-intratumoral combined models (Z=–0.506, P=0.61). In the training set, the area under the curve (AUC) values for the intratumoral, peritumoral, clinical, and combined models were 0.86, 0.63, 0.79, and 0.87, respectively. In the testing set, the corresponding AUC values were 0.84, 0.52, 0.71, and 0.87.

Figure 4 ROC curves and AUC values of the four models in the training and testing set. (A) ROC curves of different models in the training set; (B) ROC curves of different models in the testing set. Intra_tumor: intratumoral model; Peri_tumor: peritumoral model; Clinics: clinical model; Combine: clinical-intratumoral combined model. AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

Both the intratumoral and the clinical-intratumoral combined models demonstrated superior predictive performance. The intratumoral model achieved sensitivities of 0.918 (training) and 0.846 (testing), specificities of 0.667 and 0.722, positive predictive values (PPVs) of 0.800 and 0.815, and negative predictive values (NPVs) of 0.848 and 0.765. The clinical-intratumoral combined model yielded sensitivities of 0.787 (training) and 0.654 (testing), specificities of 0.833 and 0.889, PPVs of 0.873 and 0.895, and NPVs of 0.729 and 0.640 (Table 4).

Calibration curves for the clinical-intratumoral combined model demonstrated good agreement between predicted and observed outcomes in both sets, indicating satisfactory calibration (Figure 5). DCA further revealed a favorable net clinical benefit of the combined model across a wide range of threshold probabilities (Figure 6).

Figure 5 Calibration curves of the clinical-intratumoral combined model in the training and testing set. The X-axis represents the predicted probability of GGNs benign and malignant based on the combined model, and the Y-axis represents the actual probability of GGNs’ benign and malignant. GGNs, ground-glass nodules.

Figure 6 Decision curve analysis curves for the four models. X-axis: high-risk threshold; Y-axis: net benefit. Intra_tumor: intratumoral model; Peri_tumor: peritumoral model; Clinics: clinical model; Combine: clinical-intratumoral combined model.

Discussion

The intratumoral radiomics model and the combined clinical-intratumoral radiomics model developed in this study demonstrated excellent performance in differentiating benign from malignant pulmonary GGNs, with the AUC values of 0.86 and 0.87, respectively. DCA further confirmed that these two models provided greater clinical net benefit compared to the peritumoral radiomics model and the clinical model alone.

Surgical resection is the standard treatment for suspected malignant GGNs, with reported 5- and 10-year overall survival rates reaching 95–100% (3-5). In contrast, incidentally detected GGNs are typically managed by surveillance, and lesions that regress or resolve during follow-up are presumed benign (17). Therefore, accurately distinguishing between benign and malignant GGNs prior to intervention is critical for determining the necessity and timing of surgery and for prognostic assessment.

In previous studies (18-20), preoperative diagnosis of GGNs largely relied on subjective evaluation of conventional CT signs—such as spiculation, lobulation, pleural indentation, and vessel convergence—leading to variable accuracy and inter-reader consistency due to dependence on radiologist experience. Radiomics, by contrast, enables the extraction of high-throughput quantitative features from standard imaging, capturing subtle image characteristics beyond human perception (21).

Many earlier radiomics studies focused exclusively on intratumoral features (22,23), overlooking the peritumoral region, which may provide complementary information. Malignant GGNs are known to induce changes in the surrounding microenvironment, including alterations in stromal architecture, immune infiltration, and angiogenesis (24). Therefore, incorporating peritumoral features may enhance model performance.

Recent studies combining intratumoral and peritumoral radiomics features for GGN assessment have reported promising results in predicting invasiveness and pathological subtypes (11,12,24), consistent with the findings of our study. However, unlike previous work that primarily used low-dose CT, this study utilized contrast-enhanced CT data. Following contrast administration, GGNs may exhibit increased enhancement of intratumoral and peritumoral microvessels (14), potentially altering radiomics feature expression. Our results indicate that models based solely on intratumoral features, or those incorporating clinical variables, achieve superior performance in this setting. This suggests that in contrast-enhanced CT, intratumoral features may carry greater diagnostic value compared to peritumoral features.

Notably, the peritumoral model in this study demonstrated relatively lower AUC values, implying that peritumoral features extracted from enhanced CT may be less informative for predicting malignancy. This finding contrasts with the results of Wu et al. (25), who used non-contrast CT and reported higher diagnostic value for peritumoral features. Also, we found that the combined model achieved the highest AUC, but its sensitivity was unexpectedly lower than that of the other models (intratumoral model and clinical model). We attribute this counterintuitive finding to a strategic adjustment of the model’s decision boundary following the integration of clinical data. This integration appeared to instill greater overall confidence in the model’s predictions, thereby enhancing its discriminative capacity. As a result, the model adopted a more stringent threshold for classifying positive cases, which resulted in higher specificity and PPV at the expense of sensitivity. This outcome underscores a fundamental trade-off in model performance, where the combined model prioritizes a higher rate of correct overall classification and a lower false-positive rate over the detection of all potential positive cases.

There are several limitations in this study. First, it is a single-center, retrospective analysis with a relatively small sample size. Second, the absence of repeated segmentation and assessment of inter-/intra-observer reproducibility using metrics like the intraclass correlation coefficient (ICC) may impact the reliability and generalizability of radiomics features. These factors may introduce selection and information biases and limit the generalizability of our results. Additionally, the lack of external validation further restricts the applicability of our findings across different institutions. Notably, our reliance on manual segmentation—while consistent with the study’s retrospective design—also poses a practical constraint: manual operations are time-consuming and inherently prone to observer variability, which may hinder the model’s scalability in routine clinical practice. Therefore, future multicenter, prospective studies with larger cohorts and standardized segmentation protocols are warranted to validate and refine our models; exploring and validating automated segmentation methods (e.g., deep learning-based approaches) will also be a key focus to address the limitations of manual segmentation and further enhance clinical applicability.

Conclusions

The preoperative prediction models constructed using enhanced CT images, namely the intratumoral radiomics model and the intratumoral combined with clinical features model, demonstrated favorable performance in differentiating benign from malignant GGNs. Among these, the model that integrates both intratumoral radiomics features and clinical factors exhibited the best predictive performance. This suggests its substantial potential in assisting clinicians with the preoperative assessment of the nature of GGNs, thereby supporting more informed diagnostic and treatment decision-making in clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1608/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1608/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1608/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1608/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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the institutional ethics committee of Ganzhou Cancer Hospital (No. 2025247), 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. Wender R, Fontham ET, Barrera E Jr, et al. American Cancer Society lung cancer screening guidelines. CA Cancer J Clin 2013;63:107-17. [Crossref] [PubMed]
2. Sun JD, Sugarbaker E, Byrne SC, et al. Clinical Outcomes of Resected Pure Ground-Glass, Heterogeneous Ground-Glass, and Part-Solid Pulmonary Nodules. AJR Am J Roentgenol 2024;222:e2330504. [Crossref] [PubMed]
3. Li D, Deng C, Wang S, et al. Ten-Year Follow-up Results of Pure Ground-Glass Opacity-Featured Lung Adenocarcinomas After Surgery. Ann Thorac Surg 2023;116:230-7. [Crossref] [PubMed]
4. Ye X, Fan W, Wang Z, et al. Expert consensus on thermal ablation therapy of pulmonary subsolid nodules (2021 Edition). J Cancer Res Ther 2021;17:1141-56. [Crossref] [PubMed]
5. Li Q, Dai J, Zhang P, et al. Management of Pulmonary Ground Glass Nodules: Less Is More. Ann Thorac Surg 2021;112:1-2. [Crossref] [PubMed]
6. Sun Y, Li C, Jin L, et al. Radiomics for lung adenocarcinoma manifesting as pure ground-glass nodules: invasive prediction. Eur Radiol 2020;30:3650-9. [Crossref] [PubMed]
7. Zhang YP, Zhang XY, Cheng YT, et al. Artificial intelligence-driven radiomics study in cancer: the role of feature engineering and modeling. Mil Med Res 2023;10:22. [Crossref] [PubMed]
8. Feng H, Shi G, Xu Q, et al. Radiomics-based analysis of CT imaging for the preoperative prediction of invasiveness in pure ground-glass nodule lung adenocarcinomas. Insights Imaging 2023;14:24. [Crossref] [PubMed]
9. Sun Y, Ma Z, Zhao W, et al. Computed tomography radiomics in growth prediction of pulmonary ground-glass nodules. Eur J Radiol 2023;159:110684. [Crossref] [PubMed]
10. Liang L, Zhang H, Lei H, et al. Diagnosis of Benign and Malignant Pulmonary Ground-Glass Nodules Using Computed Tomography Radiomics Parameters. Technol Cancer Res Treat 2022;21:15330338221119748. [Crossref] [PubMed]
11. Zou Y, Mao Q, Zhao Z, et al. Intratumoural and peritumoural CT-based radiomics for diagnosing lepidic-predominant adenocarcinoma in patients with pure ground-glass nodules: a machine learning approach. Clin Radiol 2024;79:e211-8. [Crossref] [PubMed]
12. Jiang W, Qu T, Liu W, et al. Intra- and Peritumoral-Based Radiomics for Preoperatively Assessing the Pathological Subtype of T1-Stage Lung Adenocarcinoma Presenting as Pure Ground-Glass Nodules. Technol Cancer Res Treat 2024;23:15330338241305432. [Crossref] [PubMed]
13. Zeng Y, Zhou X, Zhou T, et al. Peritumoral radiomics increases the efficiency of classification of pure ground-glass lung nodules: a multicenter study. J Cardiothorac Surg 2024;19:505. [Crossref] [PubMed]
14. Lin Q, Choyke PL, Sato N. Visualizing vasculature and its response to therapy in the tumor microenvironment. Theranostics 2023;13:5223-46. [Crossref] [PubMed]
15. Chen N, Li R, Jiang M, et al. Progression-Free Survival Prediction in Small Cell Lung Cancer Based on Radiomics Analysis of Contrast-Enhanced CT. Front Med (Lausanne) 2022;9:833283. [Crossref] [PubMed]
16. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
17. Kim YT. Management of Ground-Glass Nodules: When and How to Operate? Cancers (Basel) 2022;14:715. [Crossref] [PubMed]
18. Qi H, Zuo Z, Lin S, et al. Assessment of intratumor heterogeneity for preoperatively predicting the invasiveness of pulmonary adenocarcinomas manifesting as pure ground-glass nodules. Quant Imaging Med Surg 2025;15:272-86. [Crossref] [PubMed]
19. Yang Y, Xu J, Wang W, et al. Meta-analysis of the correlation between CT-based features and invasive properties of pure ground-glass nodules. Asian J Surg 2023;46:3405-16. [Crossref] [PubMed]
20. Zhang P, Li T, Tao X, et al. HRCT features between lepidic-predominant type and other pathological subtypes in early-stage invasive pulmonary adenocarcinoma appearing as a ground-glass nodule. BMC Cancer 2021;21:1124. [Crossref] [PubMed]
21. Singh S, Mohajer B, Wells SA, et al. Imaging Genomics and Multiomics: A Guide for Beginners Starting Radiomics-Based Research. Acad Radiol 2024;31:2281-91. [Crossref] [PubMed]
22. Gao C, Yan J, Luo Y, et al. The Growth Trend Predictions in Pulmonary Ground Glass Nodules Based on Radiomic CT Features. Front Oncol 2020;10:580809. [Crossref] [PubMed]
23. Gao Y, Hua M, Lv J, et al. Reproducibility of radiomic features of pulmonary nodules between low-dose CT and conventional-dose CT. Quant Imaging Med Surg 2022;12:2368-77. [Crossref] [PubMed]
24. Wang X, Chen K, Wang W, et al. Can peritumoral regions increase the efficiency of machine-learning prediction of pathological invasiveness in lung adenocarcinoma manifesting as ground-glass nodules? J Thorac Dis 2021;13:1327-37. [Crossref] [PubMed]
25. Wu L, Gao C, Ye J, et al. The value of various peritumoral radiomic features in differentiating the invasiveness of adenocarcinoma manifesting as ground-glass nodules. Eur Radiol 2021;31:9030-7. [Crossref] [PubMed]