Development and validation of a computed tomography-based radiomics-clinical model to preoperatively predict high-grade patterns within lung invasive adenocarcinoma

Introduction

Lung cancer is the most prevalent malignant tumor worldwide and the leading cause of cancer-related death (1). Non-small cell lung cancer (nSCLC) accounts for approximately 85% of all lung cancer cases (2), with adenocarcinoma being the most common histopathological type (3).

According to the 2015 World Health Organization (WHO) classification of lung tumors, lung invasive adenocarcinoma (IAC) is categorized into five distinct histopathological subtypes: lepidic, acinar, papillary, micropapillary, and solid subtypes (4). Among these subtypes, high-grade patterns (HGPs) of lung adenocarcinoma (primarily micropapillary, solid, and complex glandular patterns) are associated with poor prognosis. Notably, in 2020, the International Association for the Study of Lung Cancer (IASLC) formally included the complex glandular pattern in the postoperative pathological grading system for IAC (5).

Previous studies have demonstrated that patients with HGPs frequently have a poor prognosis (6,7), and even a small amount of HGP (<5%) can adversely affect patient prognosis (8). Furthermore, the surgical approach can impact outcomes; specifically, when HGPs are present, more extensive anatomical resection is recommended, regardless of the proportion of the HGP (9). For patients with a high percentage of HGP, lobectomy is the preferred surgical option to improve survival rates (10). Therefore, the accurate identification of patients with potential HGPs before surgery is critical, as it can guide surgical decision-making and ultimately improve patient prognosis.

Currently, histopathological examination remains the gold standard for diagnosing lung cancer. However, preoperative or intraoperative pathology is limited by the volume and scope of the tissue samples, which may not fully represent the entire lesion. As a result, lesions with small amounts of HGPs can be overlooked. To obtain a complete diagnosis, additional histopathological examinations are required after full surgical resection.

Chest computed tomography (CT), which is the most commonly used preoperative imaging modality, offers a rapid, convenient, and noninvasive evaluation. However, its effectiveness is limited by the subjective interpretation of the imaging features, and it can be challenging to detect minimal HGPs, particularly in cases in which the imaging features are subtle. Consequently, some patients who have later undergone postoperative pathological confirmation of HGPs may have received suboptimal treatment, such as the use of sublobar instead of lobectomy.

Radiomics, which is a technique that extracts quantitative features from medical images, has emerged as a promising approach for improving diagnostic accuracy. By analyzing features such as texture, intensity, and shape (information that is often invisible to the human eye), radiomics can enhance the detection and characterization of tumors (11). Although several studies have used radiomics to predict lung cancer subtypes, research on HGPs in lung adenocarcinoma is more limited. Most studies have focused on micropapillary and solid patterns, with few addressing complex glandular patterns. Moreover, prior research has generally employed basic radiomics techniques without incorporating clinical indicators.

In this study, we aimed to construct and conduct preliminary validation of a preoperative prediction model for HGPs in IAC patients, specifically the patients with complex glandular patterns. This model combines CT radiomics, artificial intelligence, and clinical indicators to explore its predictive value and provide novel insights for guiding surgical decision-making. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-224/rc).

Methods

Patients

This study included 278 patients (122 males and 156 females) with a mean age of 63.3±10.6 years. All of the patients were admitted to The First Affiliated Hospital of Ningbo University between January 2020 and August 2024 and underwent surgical resection. The cohort comprised an internal cohort (n=240; retrospectively collected from January 2020 to May 2024) and a preoperative prediction cohort (n=38; consecutively collected from June to August 2024). The internal cohort, balanced with 120 HGP-positive and 120 HGP-negative patients (mean age 63.7±10.2 years), was randomly split into a training set (n=168, 70%) and a test set (n=72, 30%). The preoperative prediction cohort (mean age 60.9±12.5 years) included 11 HGP-positive and 27 HGP-negative cases for model generalizability assessment.

The inclusion criteria were as follows: (I) the patient underwent complete surgical resection at our hospital with postoperative pathological confirmation of IAC; (II) the patient did not receive neoadjuvant radiotherapy, chemotherapy, or immunotherapy before surgery; and (III) complete clinical, pathological, and preoperative CT imaging data were available. The following exclusion criteria were used: (I) postoperative pathology confirmed the presence of preinvasive adenocarcinoma or non-adenocarcinoma lesions, such as squamous cell carcinoma, small cell lung carcinoma, or benign lesions; (II) no chest CT examination was performed within one month prior to surgery; (III) a previous or current combination of other systemic malignant tumors was present; and (IV) poor quality CT images or incomplete clinical data was demonstrated. The flow charts are shown in Figures 1,2.

Figure 1 The flow chart of patient selection in the internal cohort. CT, computed tomography; HGP, high-grade pattern; IAC, invasive adenocarcinoma.

Figure 2 The flow chart of patient selection in the preoperative prediction cohort. HGP, high-grade pattern; IAC, invasive adenocarcinoma.

The sample size was determined by collecting all available HGP cases with complete pathological data from our institution, matched with an equal number of non-HGP cases, while also considering sample sizes from previous studies. All cases in the internal cohort were consecutively collected without selective exclusion.

In this study, we employed complete-case analysis, including only cases with no missing data across all variables, to ensure the integrity and reliability of the analysis.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The First Affiliated Hospital of Ningbo University ethics committee approved the study (approval No. 2024-048-RS) and waived informed consent from the patients.

Image acquisition

Chest CT images for all of the patients who were included in the study were acquired within one month prior to surgery via a Philips CT Big Bore scanner. During the scanning procedure, patients were positioned in the supine position, instructed to take a deep breath, and asked to hold their breath while the scan was performed. The entire lung parenchyma (ranging from apex to base) was included in the scan. The slice thickness was set to a maximum of 2 mm to ensure the acquirement of high-resolution images. After the scanning procedure was completed, the images were reviewed and uploaded to the hospital’s imaging system by a radiologist with extensive clinical experience.

Image segmentation and feature extraction

Chest CT images in DICOM format were uploaded to the United Imaging Intelligence platform for lesion segmentation. Target lesions on lung window CT images were segmented using automated software with semi-automatic adjustments and manual corrections by two blinded thoracic surgeons. Regions of interest (ROIs) were reviewed for accuracy, with discrepancies (e.g., vessels, bronchioles, margins) resolved through consensus and manual refinement. Radiomics features were derived from all contiguous CT slices covering the entire tumor volume, with the number of slices per lesion varying according to tumor size. Following this, the radiomics features were extracted from the segmented and preprocessed ROIs via the United Imaging Intelligence platform. Feature fusion was allowed, and empty rows were removed during this process. A total of 15 radiomics filters were selected, which automatically identified, calculated, and extracted the various features. The extracted features were grouped into seven categories: first-order, shape, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), gray level dependence matrix (GLDM), and neighboring gray tone difference matrix (nGTDM). After Z-score normalization, 2,264 features were retained.

Selection of radiomics features

Radiomics feature selection was performed through a multi-step statistical process. Initially, all 2,264 extracted features underwent the Mann-Whitney U test (P<0.05), retaining 1,727 features. To reduce redundancy, pairwise Pearson’s correlation analysis (threshold: r>0.9) was applied, resulting in 413 features. The maximal relevance minimal redundancy (mRMR) method further prioritized features with high discriminative power and low redundancy, yielding 30 candidates. Finally, the least absolute shrinkage and selection operator (LASSO) regression model with 10-fold cross-validation (λ=0.0126) was implemented to penalize irrelevant features by shrinking coefficients to zero, ultimately selecting 14 non-zero coefficient features. All analyses were conducted using Python scikit-learn.

Establishment of the radiomics model

After LASSO feature screening, the selected features were input into various machine learning algorithms to construct the radiomics risk model. For this process, a fivefold cross-validation method was employed to ensure the robustness of the final model.

Establishment of the clinical model

Features for the clinical model were selected based on baseline statistics, with those statistical results with a P value of less than 0.05 being retained. Univariate and multivariate logistic regression (LR) analyses were then applied to further refine the feature set. The clinical model was constructed using the same fivefold cross-validation framework and machine learning algorithms as the radiomics model.

Establishment of the combined radiomics-clinical model

To assess the incremental prognostic value of the radiomics model over the clinical factors, a combined radiomics-clinical model was developed. This combined model integrated the radiomics features and clinical factors via LR.

Performance assessment and model comparison

The receiver operating characteristic (ROC) curves were plotted separately in the training and test sets to assess the diagnostic performance of the predictive models, and the corresponding area under the curve (AUC), diagnostic accuracy, sensitivity, and specificity values were analyzed. The Delong test was employed for the purpose of comparing the AUCs between the different models.

Further evaluation of the combined radiomics-clinical model

Subgroup analysis

The combined model was applied to conduct a subgroup analysis in the test set of the internal cohort, thus aiming to assess the prediction model’s accuracy for different types of lesions. This analysis provides valuable insights for the model’s further development. The test set (n=72) was subdivided based on two criteria: the maximum diameter (MD) of the lesions observed on the CT images and the nature of the lesions. Lesions were initially categorized by their MD, thus resulting in two groups: MD ≤2 cm (n=34) and MD >2 cm (n=38). The subgroup with MD ≤2 cm was further stratified into two subgroups: 1 cm < MD ≤1.5 cm (n=15) and 1.5 cm < MD ≤2 cm (n=16). A separate subgroup for MD ≤1 cm was not established due to the limited sample size (n=3) within the MD ≤1 cm category in the test set and the absence of HGP-positive cases. The lesions were then classified according to their nature, after which two subgroups were obtained: mixed ground-glass lesion (mGGL) (n=34) and solid lesion (SL) (n=33). Due to the limited number of lesions with a pure ground-glass appearance (n=5), a separate group of pure ground-glass lesions was not established. In addition to the aforementioned groups, this study further stratified mGGL cases into two subgroups based on MD: MD ≤2 cm (n=24) and MD >2 cm (n=10). This stratification allowed for a comparative evaluation of the predictive efficacy of the model within the same lesion type but across different MD categories. Given that all SL cases with a MD of 2 cm within the cohort exhibited HGP (n=6), further stratification of SL cases based on MD was deemed unnecessary. The stratification of ≤2 cm lesions into 1.0–1.5 and 1.5–2 cm subgroups was based on previous research findings (12). This subdivision also optimized subgroup sample size distribution, thereby strengthening the statistical validity of comparative analyses.

Preoperative prediction cohort test

After the completion of internal testing, the combined model was preoperatively applied to evaluate its performance in a clinical setting. From June to August 2024, preoperative CT image data and blood carcinoembryonic antigen (CEA) levels were continuously collected from patients undergoing pulmonary resection in our department. The model was then used to predict the pathological types of the lesions without prior knowledge of the pathology, and the predictions were compared to the final postoperative pathological outcomes to verify the model’s accuracy.

Statistical analysis

The Python statistical model (version 0.13.2) package was employed for the purpose of performing the statistical analysis, and a P value <0.05 was considered to indicate statistical significance. We analyzed the differences between groups via Student’s t-test or the Mann-Whitney U test for continuous variables; moreover, the Chi-squared test or Fisher’s exact test was applied for categorical variables.

Results

Patient characteristics in the internal cohort

A total of 240 patients were included in the internal cohort, which was randomly divided into a training set (n=168) and a test set (n=72) at a 7:3 ratio. A pathologist reviewed the pathological data for all of the patients, all of whom underwent surgical treatment.

Among the training set, 81 patients (48.2%) had HGPs, and 87 patients (51.8%) lacked HGPs; in the test set, 39 patients (54.2%) had HGPs, and 33 patients (45.8%) lacked HGPs. The pathological patterns of patients within the internal cohort are shown in Table 1. Table 2 summarizes the baseline characteristics and main clinical factors for the training and test sets. The CT imaging characteristics of lesions in the internal cohort are summarized in Table S1.

The comparison of demographic factors such as age, sex, smoking history, and family history of cancer demonstrated no statistically significant differences between the training and test sets (P>0.05), thus indicating that the classification of patients into these groups was reasonable. However, significant differences were observed between the sets for the clinical factor CEA (P<0.05).

Feature selection and radiomics model establishment

Feature selection

From the initial 2,264 radiomics features, sequential statistical screening and dimensionality reduction identified 14 non-zero coefficient features through LASSO regression with tenfold cross-validation (λ=0.0126, Figures 3,4). These features comprised 3 GLSZM features, 3 GLCM features, 3 first-order features, 3 NGTDM features, 1 GLRLM feature, and 1 GLDM feature (Figure 5).

Figure 3 Coefficients of 10 folds cross validation. The dashed line indicates the optimal regularization parameter λ, where the model achieves the best feature selection.

Figure 4 MSE of 10 folds cross validation. The dashed line indicates the regularization parameter λ that minimizes the model’s prediction error. MSE, mean standard error.

Figure 5 Radiomics features with non-zero coefficients after LASSO regression screening. LASSO, least absolute shrinkage and selection operator; LLH, low-low-high.

Establishment of the radiomics model

Fifteen distinct machine learning algorithms were tested to develop the radiomics model, with the K-nearest neighbors (KNN) model ultimately identified as being the optimal choice. The model was constructed via fivefold cross-validation. The model’s performance (in terms of AUC) is presented in Figure 6 and Table 3 for both the training and test sets.

Figure 6 ROC analysis of three models on the training set (A) and the test set (B). AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

Establishment of the clinical model

Analysis of the clinical data revealed that CEA was an independent clinical predictor for HGPs (Table 4). A clinical model was subsequently constructed based on the CEA factor. The performance of the clinical model is illustrated in Figure 6 and Table 3.

Establishment of the combined radiomics-clinical model

A combined prediction model was developed by integrating the radiomics and clinical prediction models. To enhance the clinical applicability of the prediction model for individualized risk assessment, the complete model specification is provided, including all regression coefficients and the model intercept. The model is constructed as a LR model, which estimates the log-odds of the outcome based on the predictor variables. The specific parameter estimates for the model are: Intercept =−2.2020; Coefficient for Radiomics =4.2178; Coefficient for CEA =0.0234. Thus, the final model equation is:

logit(p)=−2.2020+4.2178×Radiomics+0.0234×CEA[1]

To estimate the predicted probability of HGP for an individual patient, the following steps should be implemented: First, the values of the predictor variables, comprising radiomics score and CEA level, are acquired for the individual. The radiomics score is derived by inputting the extracted radiomics features into the KNN model for computation. These values are then substituted into the LR model equation to compute the linear predictor, denoted as logit(p). Finally, the log-odds logit(p) are transformed into the predicted probability (p) by applying the inverse logit transformation, expressed as:

p=1/(1+e−logit(p))[2]

where: p is the predicted probability of HGP; logit(p) is the log-odds; e is the base of the natural logarithm.

This combined model demonstrated excellent performance, with AUC values of 0.934 for the training set and 0.854 for the test set (Table 3, Figure 6). The diagnostic accuracy, sensitivity, and specificity of all three models are presented in Table 3. Additionally, the calibration curve indicated that the HGP predictions that were made by the combined model were strongly correlated with the actual results in both datasets (Figure S1). Decision curve analysis (DCA) further revealed the superior predictive performance of the combined model in both datasets (Figure S2). Finally, a nomogram was developed to visualize the combined model’s predictions (Figure S3).

Comparison of the combined model with two other models

Table 3 provides a comprehensive overview of the performance metrics for the three prediction models in the internal cohort. In the training set, the combined model demonstrated superior performance, with an AUC of 0.934, thus outperforming both the radiomics and clinical models. It also exhibited a higher prediction accuracy (0.869) and sensitivity (0.765), both of which were superior to those of the other models. The specificity of the combined model was 0.966, which was comparable to that of the radiomics model and higher than that of the clinical model.

In the test set, the combined model again had the highest AUC (0.854) among the three models. It also demonstrated superior predictive accuracy (0.806) and sensitivity (0.846) compared with the radiomics and clinical models. However, the specificity of the combined model was slightly lower than that of the clinical (0.970) and radiomics models.

Figure S1 shows the calibration curves of all three models within the internal cohort. The calibration curves for the combined model showed a high degree of concordance between the predicted and observed HGP outcomes in both the training and test cohorts, thus indicating strong model performance.

Additionally, we evaluated each model via DCA, and the results are presented in Figure S2. The DCA revealed that, compared with a no-model approach (i.e., a treat-all or treat-none strategy), the combined model provided a significant benefit for patient intervention based on prediction probabilities, thus demonstrating superior clinical utility over the radiomics and clinical models.

The Delong test was conducted to compare the AUCs of the models (Figure S4). Statistically significant differences were observed between the combined radiomics-clinical model and the clinical model (P=0.003), as well as between the radiomics and clinical models (P=0.008) in the test set. However, no statistically significant difference was found between the combined radiomics-clinical model and the radiomics model (P=0.54) in the test set.

Subgroup analysis

As shown by the ROC curves (Figures 7-9), the results of the subgroup analysis demonstrated that the combined model exhibited superior predictive efficacy for lesions with a MD of ≤2 cm, with an AUC of 0.907 [95% confidence interval (CI): 0.809–1.000]. Notably, in the subgroup of lesions with MD >1.5 cm and ≤2 cm, the model’s performance was particularly superior, with an AUC of 0.933 (95% CI: 0.812–1.000). Similarly, in the lesion nature classification subgroup, the model showed enhanced predictive accuracy for mGGLs, with an AUC of 0.909 (95% CI: 0.811–1.000). Further subgroup analysis based on MD indicated that the model maintained high predictive accuracy for mGGL cases, irrespective of MD being ≤2 or >2 cm, with AUC values of 0.874 (95% CI: 0.711–1.000) and 0.905 (95% CI: 0.683–1.000), respectively.

Figure 7 The ROC curves of the subgroup analysis on the lesion maximum diameter classification. AUC, area under the curve; CI, confidence interval; MD, maximum diameter; ROC, receiver operating characteristic.

Figure 8 The ROC curves of the subgroup analysis on the lesion nature classification. AUC, area under the curve; CI, confidence interval; mGGL, mixed ground-glass lesion; ROC, receiver operating characteristic; SL, solid lesion.

Figure 9 The ROC curves based on maximum diameter in mixed ground-glass lesions. AUC, area under the curve; CI, confidence interval; MD, maximum diameter; mGGL, mixed ground-glass lesion; ROC, receiver operating characteristic.

Preoperative prediction cohort test

A total of 38 patients were included in the preoperative prediction cohort, all of whom met the inclusion and exclusion criteria and were confirmed via postoperative pathology to have lung IAC. Among these patients, 11 patients exhibited HGPs, whereas 27 patients did not exhibit HGPs. The model’s predictions were accurate in 35 cases and inaccurate in 3 cases, thus yielding an overall accuracy rate of 92.1% (Table 5).

Discussion

HGPs, including micropapillary, solid, and complex glandular subtypes, are well-established predictors of poor prognosis in IAC, associated with higher risks of recurrence, metastasis, and lymphatic invasion (13-16). Crucially, the presence of HGPs directly impacts surgical decision-making. Sublobar resection for tumors with ≥5% micropapillary components correlates with a 34.2% 5-year recurrence rate, compared to 12.4% for those with <5% components (17). Inadequate surgical margins (<1 cm) further elevate recurrence risk (32.0% vs. 13.0% for margins ≥1 cm) (12). These findings highlight the critical importance of the accurate assessment of HGP presence in lung adenocarcinoma patients prior to treatment, which could significantly inform clinical decision-making.

High-resolution CT enables preliminary identification of HGPs through quantifiable imaging features such as tumor size, density, and lobulation (18). However, its reliance on subjective interpretation limits diagnostic consistency. This challenge has been addressed through the emergence of radiomics, which extracts high-throughput features, thus offering a more objective and reproducible alternative to traditional semiquantitative imaging methods (17). Several predictive studies of HGP types have utilized radiomics technology. For example, Xu et al. (19) developed a model focusing solely on the micropapillary subtype (AUC =0.739), integrating clinical factors like smoking history, yet excluding other HGP subtypes. Another study by Chen et al. (20) achieved higher performance (AUC =0.84) by analyzing “near-pure” micropapillary/solid patterns but omitted complex glandular pattern. Another study (21) reported exceptional discrimination between lepidic and solid/micropapillary subtypes (AUC =0.984) but similarly excluded complex glandular pattern and relied on single-center data. While these studies validate radiomics’ potential, their limited HGP subtype coverage and heterogeneous methodologies hinder clinical generalizability.

In 2015, the WHO classified the complex glandular pattern as a distinct HGP subtype; moreover, in 2020, the IASLC incorporated this pattern into its postoperative pathological grading system for invasive lung adenocarcinoma. Previous studies have often excluded patients with complex glandular patterns from the HGP group; however, our study included these patients (in addition to those patients with micropapillary and solid patterns) to expand the scope of the prediction model. Compared with previous methods, this approach broadens the applicability of the model. Previous studies have predominantly used software to automatically extract radiomics features and build prediction models from limited feature sets. In contrast, our study constructed a radiomics model that includes both radiomics features and the CEA level as optimal clinical variables, which were identified via correlation analysis of clinical indicators. The integration of these two models into a combined clinical-radiomics model resulted in superior prediction performance. The combined model yielded impressive results in the training set and test set, with AUC values of 0.934 (95% CI: 0.898–0.970) and 0.854 (95% CI: 0.760–0.947).

Subgroup analysis demonstrated superior predictive performance of the combined model for lesions with MD ≤2 cm (AUC =0.907) and mGGL (AUC =0.909), with consistently high accuracy across all MD ≤2 cm subgroups. Notably, the model maintained robust predictive performance even for mGGL lesions exceeding 2 cm in MD (AUC =0.905). These findings suggest that the combined model may be particularly effective for predicting lesions with a mixed ground-glass nature or an MD of ≤2 cm on CT images. This discrepancy may originate from the pathological-radiological correlation: IAC >2 cm predominantly present as solid high-density lesions with homogeneous internal density, lacking the characteristic heterogeneous attenuation and contrast variations observed in mGGL. Pathologically, progressive fibrotic tissue proliferation and alveolar collapse accompanying tumor growth lead to gradual disappearance of ground-glass opacity features, thereby diminishing extractable radiomic features reflective of biological aggressiveness (22). Consequently, SL exhibit impoverished discriminative radiomic features for identifying HGP, leading to model limitations in reliably differentiating solid IAC lesions with versus without HGP components. Therefore, future studies should focus on expanding the sample size, particularly for SLs, to improve the model’s predictive accuracy for these subtypes. In accordance with the 2024 edition of the National Comprehensive Cancer Network (NCCN) guidelines, standard lobectomy continues to be endorsed as the principal surgical intervention for lung lesions exceeding 2 cm in diameter in cases of NSCLC (23). Consequently, the precise identification of HGP components within IAC lesions measuring 2 cm or less in diameter is of paramount clinical significance. The findings of this study not only contribute to the enhancement of early diagnostic precision for HGP but also establish a critical foundation for the development of individualized therapeutic strategies for IAC patients with lesions ≤2 cm in MD, thereby significantly improving patient prognosis.

Finally, the model was prospectively applied to a cohort of 38 patients, and the preoperative predictions were compared with the postoperative pathological results. The predicted outcomes were consistent with the actual pathological results in 35 of 38 patients, thus yielding an accuracy of 92.1%. These findings suggest that the combined clinical-radiomics model has good predictive performance and optimal generalizability.

Although the results of this study are promising, it is important to acknowledge its limitations. First, the overall sample size was relatively small, which could introduce potential selection bias. Future studies should aim to include a larger number of cases to enrich the dataset for model training. In addition, an increase in the number of training iterations and enhancement of the predictive model’s learning capacity would improve its performance, thereby reducing the impact of the small sample size on the results. Second, advancements in medical imaging technology have led to the development of imaging equipment with higher resolution and thinner scans, thus allowing for the acquisition of more precise and complex radiomics data. However, the retrospective nature of this study and the use of conventional imaging equipment limited the inclusion of more advanced imaging techniques. Additionally, segmentation of ROI within the lesion presents a challenge, as it is difficult to entirely exclude structures such as small blood vessels and bronchi. In this study, we employed a combination of automated recognition via software and manual corrections to accurately identify the lesion to the greatest possible extent. It is anticipated that future improvements in segmentation methodologies will increase accuracy, thus allowing for better delineation of the ROI. Notably, the case data that were used in the training and test sets of this study were sourced from a single medical institution, thus making this a single-center study. Data from other centers have not yet been incorporated for external validation. Therefore, further verification of the model’s generalizability by involving a larger sample size and multicenter case data is necessary.

Conclusions

In conclusion, the use of a CT-based combined radiomics-clinical model for predicting HGP components in invasive lung adenocarcinoma patients prior to surgical intervention offers several advantages over traditional diagnostic methods, including objectivity, speed, safety, and repeatability. The model has demonstrated high predictive accuracy, and it is expected that it will become a valuable clinical tool for diagnosing HGPs in the future, particularly for pulmonary lesions with a diameter smaller than 2 cm or presenting as mixed ground-glass nodules, thus guiding personalized diagnosis and treatment strategies for patients.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-224/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-224/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-224/prf

Funding: This work was supported by the China Youth Entrepreneurship and Employment Foundation, ‘China Youth Medical Innovation Research Project’ (Phase V) to X.Z., L.W., M.S., L.Z. and Z.L.; Key Research and Development Project of Ningbo (No. 2024Z220) to Y.P.; HwaMei Research Foundation of Ningbo No. 2 Hospital (No. 2023HMKY50) to Y.P. The funding source had no role in the study design, data collection, analysis, interpretation, or writing of the manuscript.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-224/coif). X.Z., L.W., M.S., L.Z. and Z.L. report that this work was supported by the China Youth Entrepreneurship and Employment Foundation, ‘China Youth Medical Innovation Research Project’ (Phase V). Y.P. reports that this work was supported by Key Research and Development Project of Ningbo (No. 2024Z220) and HwaMei Research Foundation of Ningbo No. 2 Hospital (No. 2023HMKY50). The other 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 First Affiliated Hospital of Ningbo University ethics committee approved the study (approval No. 2024-048-RS) and waived informed consent from the patients.

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. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Kratzer TB, Bandi P, Freedman ND, et al. Lung cancer statistics, 2023. Cancer 2024;130:1330-48. [Crossref] [PubMed]
3. Alduais Y, Zhang H, Fan F, et al. Non-small cell lung cancer (nSCLC): A review of risk factors, diagnosis, and treatment. Medicine (Baltimore) 2023;102:e32899. [Crossref] [PubMed]
4. Marx A, Chan JK, Coindre JM, et al. The 2015 World Health Organization Classification of Tumors of the Thymus: Continuity and Changes. J Thorac Oncol 2015;10:1383-95. [Crossref] [PubMed]
5. Moreira AL, Ocampo PSS, Xia Y, 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;15:1599-610. [Crossref] [PubMed]
6. Peng B, Li G, Guo Y. Prognostic significance of micropapillary and solid patterns in stage IA lung adenocarcinoma. Am J Transl Res 2021;13:10562-9.
7. Bai J, Deng C, Zheng Q, et al. Comprehensive analysis of mutational profile and prognostic significance of complex glandular pattern in lung adenocarcinoma. Transl Lung Cancer Res 2022;11:1337-47. [Crossref] [PubMed]
8. Choi SH, Jeong JY, Lee SY, et al. Clinical implication of minimal presence of solid or micropapillary subtype in early-stage lung adenocarcinoma. Thorac Cancer 2021;12:235-44. [Crossref] [PubMed]
9. Su H, Xie H, Dai C, et al. Procedure-specific prognostic impact of micropapillary subtype may guide resection strategy in small-sized lung adenocarcinomas: a multicenter study. Ther Adv Med Oncol 2020;12:1758835920937893. [Crossref] [PubMed]
10. Song W, Hou Y, Zhang J, et al. Comparison of outcomes following lobectomy, segmentectomy, and wedge resection based on pathological subtyping in patients with pN0 invasive lung adenocarcinoma ≤1 cm. Cancer Med 2022;11:4784-95. [Crossref] [PubMed]
11. Mayerhoefer ME, Materka A, Langs G, et al. Introduction to Radiomics. J Nucl Med 2020;61:488-95. [Crossref] [PubMed]
12. Nitadori J, Bograd AJ, Kadota K, et al. Impact of micropapillary histologic subtype in selecting limited resection vs lobectomy for lung adenocarcinoma of 2cm or smaller. J Natl Cancer Inst 2013;105:1212-20. [Crossref] [PubMed]
13. Wang Y, Hu J, Sun Y, et al. Micropapillary or solid component predicts worse prognosis in pathological IA stage lung adenocarcinoma: A meta-analysis. Medicine (Baltimore) 2023;102:e36503. [Crossref] [PubMed]
14. Choi Y, Kim J, Park H, et al. Rethinking a Non-Predominant Pattern in Invasive Lung Adenocarcinoma: Prognostic Dissection Focusing on a High-Grade Pattern. Cancers (Basel) 2021;13:2785. [Crossref] [PubMed]
15. Chang C, Sun X, Zhao W, et al. Minor components of micropapillary and solid subtypes in lung invasive adenocarcinoma (≤ 3 cm): PET/CT findings and correlations with lymph node metastasis. Radiol Med 2020;125:257-64. [Crossref] [PubMed]
16. Li Y, Byun AJ, Choe JK, et al. Micropapillary and Solid Histologic Patterns in N1 and N2 Lymph Node Metastases Are Independent Factors of Poor Prognosis in Patients With Stages II to III Lung Adenocarcinoma. J Thorac Oncol 2023;18:608-19. [Crossref] [PubMed]
17. Scapicchio C, Gabelloni M, Barucci A, et al. A deep look into radiomics. Radiol Med 2021;126:1296-311. [Crossref] [PubMed]
18. Dong H, Qiu Y, Wang X, et al. Predictive value of logistic regression model based on high-resolution CT signs for high-grade pattern in stage IA lung adenocarcinoma. China Oncology 2023;33:768-75.
19. Xu Y, Ji W, Hou L, et al. Enhanced CT-Based Radiomics to Predict Micropapillary Pattern Within Lung Invasive Adenocarcinoma. Front Oncol 2021;11:704994. [Crossref] [PubMed]
20. Chen LW, Yang SM, Wang HJ, et al. Prediction of micropapillary and solid pattern in lung adenocarcinoma using radiomic values extracted from near-pure histopathological subtypes. Eur Radiol 2021;31:5127-38. [Crossref] [PubMed]
21. Park S, Lee SM, Noh HN, et al. Differentiation of predominant subtypes of lung adenocarcinoma using a quantitative radiomics approach on CT. Eur Radiol 2020;30:4883-92. [Crossref] [PubMed]
22. Woodworth CF, Frota Lima LM, Bartholmai BJ, et al. Imaging of Solid Pulmonary Nodules. Clin Chest Med 2024;45:249-61. [Crossref] [PubMed]
23. Riely GJ, Wood DE, Ettinger DS, et al. Non-Small Cell Lung Cancer, Version 4.2024, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2024;22:249-74. [Crossref] [PubMed]