Sage Journals: Discover world-class research

Abstract

Purpose: To evaluate the diagnostic performance of CT radiomic features extracted from tumor and peritumoral regions in identifying pathological subtypes of T1-stage lung adenocarcinoma presenting as pure ground−glass nodules (pGGNs). Methods: A retrospective analysis was conducted on the data of T1-stage lung adenocarcinoma patients who underwent surgical resection and whose preoperative CT scans revealed pGGNs from June 2020 to June 2023 in our hospital. 3D Slicer was used to extract radiomic features of the intratumoral (VOI entire) and peritumoral regions (VOI +2 mm), and Rad-scores were calculated from the coefficients of features obtained after dimensionality reduction with LASSO regression. Results: A total of 131 patients with T1-stage lung adenocarcinoma presenting as pGGNs were included in this study; of these, 84 were pathologically diagnosed with the lepidic-predominant (LPA) subtype, and 47 were diagnosed with non-LPA. The diagnostic performance of the VOI entire and VOI +2 mm features for the pathological subtype of pGGN was superior to that of conventional features, with the VOI +2 mm features showing the best performance: the area under the curve, sensitivity, specificity, and accuracy in the training set were 0.883, 0.964, 0.667, and 0.761, respectively. Conclusion: Intra- and especially peritumoral-based radiomic features have high diagnostic performance for the pathological subtype of T1-stage lung adenocarcinoma presenting as pGGNs.

Keywords

CT radiomics tumor peritumoral pure ground-glass lung nodule pathological subtype

Introduction

Advances in and the widespread use of low-dose CT for screening lung cancer have helped significantly increase the detection of ground-glass nodules (GGNs).¹ Studies have shown that pure ground-glass nodules (pGGNs) mostly exhibit indolent biological characteristics, with a long window of progression; the results of surgical pathology of pGGNs often indicate a diagnosis of T1-stage lung adenocarcinoma or precancerous lesions.^2,3 Among T1-stage lung adenocarcinomas presenting as pGGNs, the main pathological type is lepidic-predominant growth adenocarcinoma, which is characterized by low malignancy, a good prognosis, and a low rate of lymph node metastasis. Surgical methods mostly include segmentectomy or wedge resection, while systematic mediastinal lymph node dissection is less frequently employed.⁴ According to the 2015 World Health Organization classification of lung adenocarcinoma prognostic risk levels, the lepidic-predominant type is classified as low risk, having the best prognosis.⁵ Additionally, in the 2020 revised lung adenocarcinoma prognostic risk classification, lung adenocarcinoma is divided into three grades based on its pathological components, with lepidic-predominant and high-risk components (complex glandular type, solid or micropapillary predominant adenocarcinoma, MPA) accounting less than 20% classified as Grade 1, which also has the best prognosis.⁶ Therefore, accurate preoperative assessment of the pathological subtype of pGGNs is highly important for adjusting follow-up strategies and formulating surgical plans for treating the patient.

Recent research has focused mainly on evaluating the conventional imaging features of GGNs on CT to further predict their postoperative pathological subtype.^7–9 However, conventional imaging is limited to the morphological evaluation of pGGNs, making it difficult to conduct a multidimensional, in-depth analysis of their texture characteristics. Moreover, the results heavily depend on the radiologist's diagnostic experience, introducing significant subjectivity. Radiomics is a technique that uses computer algorithms to analyze medical imaging data and extract numerous image features, allowing for a more objective assessment of properties of the lesion, quantification of lesion characteristics, and prediction of disease progression and prognosis. It has extensive applications in the diagnosis, treatment strategy selection, and prognosis evaluation of lung nodules.¹⁰ Furthermore, considering the potential for lung cancer to infiltrate surrounding lung parenchyma, which may alter the peritumoral microenvironment, the peritumoral characteristics has potential to serve as important tools for assessing tumor growth, cell migration, inflammation, and microvessel formation.¹¹

This study builds on previous research by extracting radiomic features of the tumor and peritumoral regions of pGGNs and collecting their clinical imaging data to establish a predictive model for the pathological subtype of pGGNs. The model aims to enhance diagnostic accuracy at an earlier preoperative stage, enabling clinicians to identify pathologic subtypes more clearly and to determine appropriate patient management strategies and treatment options in a shorter period of time, which will help reduce the chance of disease progression.

Materials and Methods

The retrospective studies involving human participants were reviewed and approved by the Ethics Committee of the Universal Cloud Medical Imaging Diagnostic Center, Shanghai University (No. 2024KY056). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements. In this study, informed consent was obtained from all study participants. The informed consent was verbal. The patients consented to participate in the research study and for the use of their data for scientific analysis and publication while ensuring the confidentiality and anonymity of their personal information. The reporting of this study conforms to TRIPOD guidelines.¹²

Patient Selection

A retrospective analysis was conducted on T1-stage lung cancer patients who were diagnosed with pGGNs by CT and underwent surgical resection from June 2020 to June 2023 and identified according to strict inclusion criteria as follows: i. postoperative pathology confirming a diagnosis of T1-stage lung adenocarcinoma (T1N0M0); ii. preoperative chest CT images showing pGGNs; and iii. no history of radiotherapy, chemotherapy, immunotherapy, or other surgeries before the CT examination. The exclusion criteria were as follows: i. no preoperative CT volumetric scan (slice thickness: 0.625 mm–1.25 mm); ii. poor image quality due to motion or respiratory artifacts, affecting subsequent tumor delineation and image processing; and iii. maximum pGGN diameter on CT volumetric images larger than 30 mm or smaller than 5 mm. According to the Fleischner Society's recommendations, pGGNs are defined as hazy patches on lung windows with visible lung texture and faintly discernible lung parenchyma.¹ In this study, the enrolled patients were divided at a 2:1 ratio into a training cohort and a validation cohort. The data of the training cohort were used for model training, and those of the validation cohort were used for model validation.

CT Image Acquisition

The CT scans in this study were performed with a GE Revolution CT scanning system. During the lung CT scans, patients were instructed to take a shallow breath and hold it for 5 s. The scanning range extended from the lung apices to the bilateral adrenal glands. The scanning parameters of the system were as follows: tube voltage of 120 kV, tube current of 100–150 mA, pitch of 0.828–1.2, matrix size of 512 × 512, standard kernel algorithm, and reconstruction with a slice thickness of 0.625–1.250 mm and an interslice distance of 0.625–1.250 mm. The completed scans were automatically transmitted to the picture archiving and communication system.

Segmentation, Feature Extraction, and Selection

Two senior radiologists with over 10 years of experience in chest imaging analysis interpreted the conventional CT radiological features of the lung nodules while being blinded to the pathological results. When the conclusions between the doctors differed, they reached a consensus through discussion. The assessed conventional features included boundaries, lobulation, spiculation, vascular convergence, vacuole signs, and pleural retraction signs. The pGGNs of the lung were segmented with the open-source platform 3D Slicer (v4.9.0, www.slicer.org) and its “segment editor” module. Semiautomatic segmentation techniques, namely, “level tracing” and “smoothing”, were employed to delineate the tumor contour according to the appearance of the tumor on CT images in the axial, sagittal, and coronal planes, including the entire tumor (VOI entire) and the tumor boundary extended to the 2 mm peritumoral region (VOI +2 mm) (Figure 1). The segmentation was subsequently reviewed by radiologists and thoracic surgeons, with additional adjustments made in case of differing opinions.

Figure 1.

Process of VOI entire and VOI +2 mm lesion segmentation.

To extract radiomic features of the tumor and peritumoral regions from CT images, the PyRadiomics plugin 3.0, an open-source Python package for radiomic feature extraction, was employed.¹³

The radiomic features obtained in this study covered multiple categories, including first-order statistical features, gray-level histogram statistical features, 2D and 3D morphological features, and texture features (such as gray-level dependence matrix and gray-level size zone matrix features).

Intra- and Peritumoral-Based Radiomic Model Development

Before dimensionality reduction, the patient feature values were standardized by calculating their z scores as follows: z score = (x - μ)/σ, where x represents the feature value, μ is the mean of the feature values across all patients, and σ is the corresponding standard deviation; this equation transforms the data to have a mean of 0 and a variance of 1. Subsequently, Pearson or Spearman correlation redundancy analysis was performed on the standardized radiomic features, and features with a correlation coefficient exceeding 0.9 were identified. Finally, least absolute shrinkage and selection operator (LASSO) and 10-fold cross-validation were used to reduce the dimensionality of the remaining radiomic features (Figure 2), retaining those with nonzero coefficients, which were then summed to obtain the Rad score.

Figure 2.

Least absolute shrinkage and selection operator regression model with 10-fold cross-validation.

Statistical Analysis

Statistical analysis was performed with R.4.3.1, and the diagnostic performance for the pathological subtype of pGGNs in the lungs was assessed with receiver operating characteristic (ROC) curve analysis separately for models constructed with VOI entire, VOI +2 mm, and conventional features, including calculation of the area under the curve (AUC), sensitivity, specificity, and accuracy.

Results

Comparison of Baseline Data

A total of 131 patients were included in the study after application of the predefined inclusion and exclusion criteria. Among them, 84 patients were diagnosed with the lepidic-predominant subtype, whereas the remaining 47 patients were diagnosed with the non-lepidic-predominant subtype. Univariate analysis (Table 1) between the training set, which consisted of the data of 2/3 (n = 88) of the patients, and the validation set, which consisted of the data of 1/3 (n = 43) of the patients, revealed no statistically significant differences in the baseline characteristics (all P > 0.05).

Table 1.

Comparative Evaluation of Clinico-Radiological Features Between the Training and Validation Cohort.

Variables	Total (n = 131)	Training cohort (n = 88)	Validation cohort (n = 43)	p value
Pathological subtypes, n (%)				0.233
LPA	84 (64.1)	60 (68.2)	24 (55.8)
Non-LPA	47 (35.9)	28 (31.8)	19 (44.2)
Boundary, n (%)				0.122
Well-Defined	59 (45)	35 (39.8)	24 (55.8)
Ill-Defined	72 (55)	53 (60.2)	19 (44.2)
lobulation, n (%)				0.316
Absence	108 (82.4)	70 (79.5)	38 (88.4)
Presence	23 (17.6)	18 (20.5)	5 (11.6)
Spiculation, n (%)				1.000
Absence	102 (77.9)	69 (78.4)	33 (76.7)
Presence	29 (22.1)	19 (21.6)	10 (23.3)
Vascular Convergence Sign, n (%)				0.765
Absence	107 (81.7)	73 (83)	34 (79.1)
Presence	24 (18.3)	15 (17)	9 (20.9)
Vacuole Sign, n (%)				1.000
Absence	120 (91.6)	80 (90.9)	40 (93)
Presence	11 (8.4)	8 (9.1)	3 (7)
Pleural Indentation, n (%)				0.776
Absence	89 (67.9)	61 (69.3)	28 (65.1)
Presence	42 (32.1)	27 (30.7)	15 (34.9)

Evaluation of Different Models for Predicting the Pathological Subtype of Pure Ground-Glass Lung Nodules

Univariate and multivariate logistic regression analyses were performed on the conventional characteristics of pGGNs of the lung in the predominantly attached and nonattached groups. The univariate analysis results revealed that lobulation and spiculation were associated with the pathological subtype, whereas the multivariate results indicated that only spiculation was an independent subtype predictor (Table 2).

Table 2.

Univariate and Multivariate Logistic Analysis for Distinguishing Histological Subtypes in Clinical Stage T1 pGGNs Lung Adenocarcinoma.

	univariate logistic	P value	multivariate logistic	P value
	OR (95%CI)	P value	OR (95%CI)	P value
Variables	0.78 (0.31-2)	0.600
Lobulation	0.28 (0.095-0.81)	0.019	1.21(0.937∼1.56)	0.149
Spiculation	0.24 (0.082-0.69)	0.008	1.29(1.005∼1.657)	0.049
Vascular Convergence Sign	0.46 (0.15-1.4)	0.180
Vacuole Sign	3.60 (0.42-30)	0.250
Pleural Indentation	0.56 (0.22-1.5)	0.230

Evaluation of Lesion and Perilesional Radiomic Features in Predicting the Pathological Subtype of Pure Ground-Glass Lung Nodules

In the selection of lesional and perilesional radiomic features for predicting the pathological subtype of pGGNs of the lung, LASSO dimensionality reduction resulted in the final identification of 5 and 9 radiomic predictors for VOI entire and VOI +2 mm, respectively (Figure 3). In a comparison of the Rad-scores constructed from the VOI entire and VOI +2 mm features, the Rad-scores of the lepidic-predominant patients were lower than those of the non lepidic-predominant patients (both P < 0.05) (Figure 4).

Figure 3.

Radiomics signature was built for VOI entire and VOI +2 mm lesion.

Figure 4.

Violin diagrams comparing Radscore between LPA and non-LPA in the training and validation cohort.

Using conventional features as well as the VOI entire and VOI +2 mm features to diagnose the pathological subtype of lung pGGNs, ROC curve analysis (Figure 5) revealed that the diagnostic performance of the VOI entire- and VOI +2 mm-based models was superior to that of conventional characteristics. Among them, the VOI +2 mm model had the best diagnostic performance: the AUC, sensitivity, specificity, and accuracy in the training set were 0.883, 0.964, 0.667, and 0.761, respectively, and those in the validation set were 0.886, 0.895, 0.750, and 0.815, respectively (Table 3).

Figure 5.

ROC curve comparing performance across image finding, VOI entire, and VOI +2 mm in the training and validation cohort.

Table 3.

Evaluating Diagnostic Performance Across Models in Distinguishing LPA.

Models	AUC	ACC	SEN	SPN
Training cohort
Image finding	0.658	0.693	0.5	0.783
VOI _entire	0.872	0.75	0.964	0.650
VOI _+2mm	0.883	0.761	0.964	0.667
Validation
Image finding	0.768	0.791	0.632	0.917
VOI _entire	0.833	0.814	0.947	0.708
VOI _+2mm	0.886	0.815	0.895	0.75

Discussion

This study conducted radiomic analysis on CT images of pGGNs in the lungs to explore the diagnostic efficacy of conventional features, VOI entire features, and VOI +2 mm for the pathological subtype of pGGNs. The results showed that all three methods effectively identified the pathological subtype of lepidic-predominant GGNs, with radiomics models (VOI entire, VOI +2 mm) outperforming the conventional feature models, and the VOI +2 mm model having the highest diagnostic efficacy.

Among these subtypes, the lepidic-predominant subtype has the better prognosis.^5,6 Previous studies^5–9 have indicated that for lung pGGNs, the most common pathological subtype is the lepidic-predominant subtype, characterized by relatively low malignancy and a low lymph node metastasis rate. Sublobar resection is frequently recommended for treating these pGGNs, while systematic mediastinal lymph node dissection is less frequently suggested. Therefore, preoperative evaluation of the pathological subtype of pGGN in the lungs is highly valuable for follow-up management, surgical timing and planning, postoperative monitoring, and prognosis assessment. However, the preoperative CT diagnosis of the pathological subtype also presents new challenges for radiologists. In previous studies, preoperative evaluations of the pathological subtypes of GGNs in the lungs mainly consisted of the analysis of conventional CT features, which include spiculation, lobulation, pleural retraction, bubble-like lucency, and vascular convergence signs.^7–9 However, owing to the significant subjectivity of these signs, given their large dependence on the radiologist's diagnostic experience, their consistency is poor, limiting their generalizability.

Radiomics can delve deeper into conventional images, mapping image pixels to the tumor's biological microenvironment and thereby extracting multidimensional features that are not observable by the naked eye.¹⁰ The value of CT radiomics in pGGNs has been reported with promising results.^14–17 Xiong et al¹⁸ demonstrated that radiomic features based on CT improved the recognition performance of pathological subtypes of pGNNs lung adenocarcinoma, and achieved a good prediction performance of 0.792 AUC. Currently, most studies focus only on intra-tumor features, ignoring the fact that peri-tumor features contain equally important information. And CT radiomic features of the peritumoral region has long been shown to be a very valuable factor to study.^19–21 Considering the potential for tumor infiltration, which may alter the peritumoral lung tissue microenvironment, peritumoral features are also important for evaluating tumor growth, cell migration, inflammation, and microvascular formation.¹¹ Therefore, it is necessary to consider peritumoral features when assessing pGGNs of the lung. Wu et al¹¹ reported that when evaluating the invasiveness of pGGNs, the diagnostic efficacy of features of the peritumoral region 2 mm beyond the tumor boundary was not greater than the diagnostic efficacy of the features of the tumor region itself.

Based on this finding, two models were employed in this study, one based on features from the entire VOI and another based on the features from the entire VOI with a 2-mm peritumoral region, for the radiomic analysis of pGGNs in the lungs following dimensionality reduction to select optimal radiomic features and construct the Rad-score. The Rad-score can quantify tumor heterogeneity, the biological characteristics of malignancies, and the tumor microenvironment. The study results revealed that for lesions with a postoperative pathological subtype of lepidic-predominant, the Rad score was lower than that of the other pathological subtypes, corresponding to the lower degree of malignancy of lepidic-predominant lung adenocarcinoma. In addition, the results of the present study fully exposed the great potential of radiomics features of the peritumoral region in diagnosing the pathological subtypes of pGGNs, reaching an AUC of 0.886 on the validation group, which was superior to the conventional features of 0.768 and the whole VOI of 0.833. In contrast to other related studies that have utilized only tumor-area radiomic features, this study provides a valid comparison regarding the predictive effects of peritumoral and tumor-area radiomic features as well as conventional features, and achieves good predictive results.

This study has certain limitations: first, it is a single-center study and lacks external validation, and the generalizability of the model needs to be improved and it is prone to poor generalization of the model to data from external centers; second, it is a retrospective study with a small sample size, which made it vulnerable to selection bias and information bias, and the small sample size may have led to an increase in chance outcomes. Finally, our failure to conduct long-term follow-ups in our study may render us susceptible to bias. Future studies will consider multicenter, prospective data collection and larger sample sizes to improve the validity and stability of results. It will also aim to enhance the clinical application value of the model.

Conclusion

In summary, prediction models were constructed for the pathological subtype of pGGNs in the lungs with radiomic features extracted from the tumor and peritumoral regions. The results revealed that both the VOI entire model and the VOI +2 mm model outperformed the conventional feature model, with the VOI +2 mm model showing the best performance, highlighting the importance of the peritumoral microenvironment in evaluating the biological characteristics of tumors.

Supplemental Material

sj-docx-1-tct-10.1177_15330338241305432 - Supplemental material for Intra- and Peritumoral-Based Radiomics for Preoperatively Assessing the Pathological Subtype of T1-Stage Lung Adenocarcinoma Presenting as Pure Ground-Glass Nodules

Supplemental material, sj-docx-1-tct-10.1177_15330338241305432 for Intra- and Peritumoral-Based Radiomics for Preoperatively Assessing the Pathological Subtype of T1-Stage Lung Adenocarcinoma Presenting as Pure Ground-Glass Nodules by Wenting Jiang, Tingting Qu, Weiran Liu, Huazheng Shi and Yali Zhang in Technology in Cancer Research & Treatment

Footnotes

Acknowledgments

None.

Author Contributions

The literature search, analysis, data explanation, and manuscript draft were finished by WJ, TQ, and WL. WL and HS are responsible for the analysis and explanation of the radiomics imaging features data. WJ, TQ, and WL and YZ acquired the clinical information and imaging. YZ designed the study, explained the data, and made multiple revisions to the manuscript. All authors contributed to the article and approved the submitted version.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors without undue reservation.

Ethics Statement

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Shanghai Municipal Health Commission, (grant number No.202340225).

ORCID iD

Yali Zhang

Supplemental Material

Supplemental material for this article is available online.

References

Zuo

Wang

Zeng

Zhang

. Measuring pure ground-glass nodules on computed tomography: Assessing agreement between a commercially available deep learning algorithm and radiologists’ readings. Europe PMC. 2023;64(4):1422-1430.

Zhang

Wang

, et al. Computed tomography density is not associated with pathological tumor invasion for pure ground-glass nodules. J Thorac Cardiovasc Surg. 2021;162(2):451-459.e3.

Lee

Park

, et al. Lung adenocarcinoma invasiveness risk in pure ground-glass opacity lung nodules smaller than 2cm. Thorac Cardiovasc Surg. 2019;67(4):321-328.

Cheng

Zheng

, et al. Tumor histology predicts mediastinal nodal status and may be used to guide limited lymphadenectomy in patients with clinical stage I non-small cell lung cancer. J Thorac Cardiovasc Surg. 2018;155(6):2648-2656.e2.

Murakami

Ito

Tsubokawa

, et al. Prognostic value of the new IASLC/ATS/ERS classification of clinical stage IA lung adenocarcinoma. Lung Cancer. 2015;90(2):199-204.

Moreira

Ocampo

PSS

Xia

, et al. A grading system for invasive pulmonary adenocarcinoma: A proposal from the international association for the study of lung cancer pathology committee. J Thorac Oncol. 2020;15(10):1599-1610.

Zhang

Tao

, 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(1):1124.

Shang

Gao

, et al. Correlation between computed tomography imaging and pathological stages and subtypes in early lung adenocarcinoma. J Cancer Res Ther. 2020;16(7):1569-1574.

Han

, et al. Comparative study of imaging and pathological evaluation of pneumonic mucinous adenocarcinoma. Oncol Lett. 2021;21(2):125.

10.

Thawani

McLane

Beig

, et al. Radiomics and radiogenomics in lung cancer: A review for the clinician. Lung Cancer. 2018;115:34-41.

11.

Gao

, et al. The value of various peritumoral radiomic features in differentiating the invasiveness of adenocarcinoma manifesting as ground-glass nodules. Eur Radiol. 2021;31(12):9030-9037.

12.

Collins

Reitsma

Altman

Moons

. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD statement. Br Med J. 2014;350:g7594. doi:https://doi.org/10.1136/bmj.g7594

13.

van Griethuysen

JJM

Fedorov

Parmar

, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104-e107.

14.

Xiong

Jiang

Che

, et al. Use of CT radiomics to differentiate minimally invasive adenocarcinomas and invasive adenocarcinomas presenting as pure ground-glass nodules larger than 10 mm. Eur J Radiol. 2021;141:109772.

15.

Kao

Hsieh

Chen

, et al. CT-Based radiomic analysis for preoperative prediction of tumor invasiveness in lung adenocarcinoma presenting as pure ground-glass nodule. Cancers (Basel). 2022;14(23):5888.

16.

Jiang

Xiong

Zhao

, et al. Computed tomography radiomics-based distinction of invasive adenocarcinoma from minimally invasive adenocarcinoma manifesting as pure ground-glass nodules with bubble-like signs. Gen Thorac Cardiovasc Surg. 2022;70(10):880-890.

17.

Zou

Mao

Zhao

, 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(2):e211-e218.

18.

Xiong

Jiang

Tian

, et al. Radiomics for identifying lung adenocarcinomas with predominant lepidic growth manifesting as large pure ground-glass nodules on CT images. PLoS One. 2022;17(6):e0269356.

19.

Tang

Huang

, et al. Intratumoral and peritumoral CT-based radiomics strategy reveals distinct subtypes of non-small-cell lung cancer. J Cancer Res Clin Oncol. 2022;148(9):2247-2260.

20.

Shang

Chen

, et al. Computed tomography-derived intratumoral and peritumoral radiomics in predicting EGFR mutation in lung adenocarcinoma. Radiol Med. 2023;128(12):1483-1496.

21.

Lou

Kong

, et al. Can quantitative peritumoral CT radiomics features predict the prognosis of patients with non-small cell lung cancer? A systematic review. Eur Radiol. 2023;33(3):2105-2117.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.97 MB