Enhancing the intraoperative identification of high-grade patterns in invasive lung adenocarcinoma via radiomics

Introduction

Lung cancer ranks first in cancer-related mortality worldwide (1,2), with lung adenocarcinoma (LUAD) accounting for the majority of lung cancer cases. Fortunately, with development and prevalence of screening technology, lung tumors are being detected earlier, facilitating timely surgical intervention and ultimately reducing mortality (3). In recent years, there has been extensive research conducted on the surgical procedures for lung cancer. For early-stage lung cancer [tumor size ≤2 cm, American Joint Committee on Cancer (AJCC) stage I], several clinical trials have confirmed that sublobar resection, including wedge resection and segmentectomy, is not only noninferior to lobectomy (4) with respect to disease-free survival but also provides greater lung function preservation and fewer perioperative complications (5-8). However, the probability of recurrence after sublobar resection and incomplete systematic lymph node dissection is increased in some early-stage lung cancers with high-risk features (9,10).

LUAD is a histologically heterogeneous disease with complex subtypes and growth patterns. The International Association for the Study of Lung Cancer (IASLC), the American Thoracic Society (ATS) and the European Respiratory Society (ERS) have devised a new multidisciplinary classification for LUAD as described in 5% increments, which includes the lepidic, acinar, papillary, micropapillary, and solid patterns (11). Classic high-grade patterns (HGPs) include micropapillary and solid patterns. A study has confirmed that patients in whom this histological subtype predominates are more likely to develop lymphatic, vascular, and pleural invasion (12). Even if not the predominant subtype, micropapillary and solid patterns nonetheless indicate a worse prognosis for patients (13). Moreover, some research suggests that these subtypes significantly shorten disease-free survival regardless of the proportion, with micropapillary being an independent predictor of overall survival (14), even when constituting less than 5% of the entire tumor (15) or in IA stage disease (16). In addition to these five classic subtypes, complex glandular patterns (CGPs), including cribriform and fused gland patterns, have been identified and classified as HGPs (17). Even the presence of a small portion of the cribriform subtype can predict the relapse of tumor (18). CGPs and micropapillary patterns have also been associated with a higher incidence of lymph node micrometastasis, and thus adequate lymph node dissection is considered highly beneficial to improving survival (19). Consequently, the selection of more thorough the surgical procedures, which range from the systematic lymph node dissection to anatomic lobectomy, can be considered according to risk stratification results provided by pre- and intraoperative information. Moreover, HGPs also influence postoperative adjuvant therapy. Studies have shown that stage I/IB/I–III LUAD with predominant micropapillary or solid patterns are significantly associated with a reduced likelihood of recurrence and improved survival after postoperative adjuvant chemotherapy (13,20,21). Additionally, patients with stage I LUAD with HGPs who undergo postoperative adjuvant therapy exhibit a reduced risk of recurrence and prolonged disease-free survival (22), even when the micropapillary component constitutes as little as 1% of the whole tumor (23).

For certain lesions, surgeons often send the resected sublobar lesions to be examined via intraoperative frozen section (FS) analysis. Based on the rapid pathological assessment of the tumor malignancy, the final surgical approach can be determined, which includes deciding whether to proceed with additional lobectomy and more comprehensive lymph node dissection. However, there are presently no fully reliable and quantitative methods that can select suitable candidates for sublobar resection. As FS is critical to guiding surgical procedures (24), there is an emerging consensus that FS is required to identify HGPs in LUAD, which can select a group for whom sublobar resection will be relatively safe (25). However, in determining whether HGPs are present in invasive lung adenocarcinoma (IAC), FS exhibits unsatisfactory sensitivity (26,27). Errors and variations in sampling and interpretation, as well as poor FS quality, have been reported as major causes of FS-based misdiagnosis.

Preoperative imaging provides additional information for characterizing the tumor (28), and radiomics (29,30), which extracts high-throughput quantitative features from medical images and uses data-characterization algorithms, is a pipeline that could be productively integrated with FS to enhance quantitative analysis. Compared to FS, the radiomics analyzes the whole tumor volume and can be performed preoperatively. Moreover, as expertise of different pathologists differs and the judgment is inherently subjective, there may be high cross-reader variation and error in interpretation in FS analysis. In this context, artificial intelligence (AI) and computer-aided diagnosis (CAD) may represent more trustworthy alternatives, as they are less prone to variation and error if appropriately trained and adopted in a rationale manner (31).

Previous studies on this subject have focused exclusively on either computed tomography (CT) imaging machine learning (including deep learning) or FS to predict HGPs (32-35); however, none have compared and comprehensively integrated both methods. Therefore, the objective of this study is to evaluate the value of combining radiomics with FS for HGPs prediction in patients with clinical T1 (cT1) LUAD. Patients with lesions smaller than 3 cm (cT1 in the AJCC T stage) (36,37) receiving surgical treatments and FS examinations were retrospectively recruited, and a radiomics (R) model was built to complement FS analysis in predicting HGPs intraoperatively. The validity and interpretability of the radiomics model and the additional value that it may provide to FS were fully assessed to determine its potential to inform surgical procedures, including the preparation for positive adjuvant therapy after surgery in patients with cT1 LUAD. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-504/rc).

Methods

Study population

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Zhongshan Hospital, Fudan University (approval No. B2021-128). Informed consent was waived by the ethics committee due to the retrospective nature of this study.

A total of 666 patients who underwent surgery for LUAD in Zhongshan Hospital, Fudan University from January 2019 to April 2019 were reviewed retrospectively. The inclusion criteria were as follows: (I) confirmation of primary IAC via final pathology (FP); (II) intraoperative FS performed; and (III) completion of a non-contrast chest CT scan (slice thickness ≤1 mm) within a month before surgery. Meanwhile, the exclusion criteria were as follows: (I) administration of neoadjuvant therapy; (II) a maximum diameter of the lesion greater than 3 cm; (III) severe respiratory motion artifacts or with missing relevant data; (IV) invasive mucinous adenocarcinoma or other rare subtypes such as fetal and enteric type adenocarcinoma; (V) small biopsies and cytology specimens obtained from lesions before CT examination; and (VI) a history of malignant tumor. When multiple nodules were observed in one resected tissue, the largest one was selected for subsequent analysis. Based on these criteria, 490 patients were included. All lesions were randomly allocated to either a training set (n=392) or test set (n=98). A flowchart of the population selection is shown in Figure 1.

Figure 1 Enrollment of patients. CT, computed tomography; FS, frozen section; IAC, invasive adenocarcinoma; HGP, high-grade pattern; non-HGP, non-high-grade pattern.

CT scans

All CT images were exported from the picture archiving and communication system (PACS) in Digital Imaging and Communications in Medicine (DICOM) format. All patients underwent supine position CT scans for thoracic imaging and were required to hold their breath while scanning. The bone algorithm was employed for reconstruction. Chest CT scans were acquired using spiral CT scanners from Siemens (16/64/128-slice), United Imaging (80/128/160-slice), GE Healthcare (64-slice), and Canon (128/320-slice). The tube voltage was set to 120 kVp with automatic tube current modulation. Slice thickness was set as 0.625 or 1.0 mm. Detailed acquisition protocols are listed in Table S1.

Radiological and pathological evaluation

Two radiologists with more than 5 years of working experience evaluated these lesions in both soft tissues and lung windows. They measured the maximum diameter in the whole three-dimensional (3D) tumors and classified them into pure ground-glass nodules (pGGNs), part-solid nodules (PSNs), and pure-solid tumors (PSTs) without knowledge of any other pathological information. PSNs were defined as nodules containing both ground-glass and solid opacity. In cases of disagreement, a senior radiologist made the final decision.

All FP slides were interpreted by two pathologists with more than 5 years of working experience. We also conducted a review of FS results according to previously stored images and reports. The key focus of review was to check for the presence of HGPs (micropapillary, solid, and CGPs). If present, they were marked as 1; if absent, they were marked as 0. When disagreements arose between pathologists, a third expert was responsible for the final decision.

Extraction of radiomics features

CT images were imported into 3D Slicer version 5.6.2 (https://www.slicer.org) in anonymized DICOM format. All tumors were segmented by a radiologist with 5 years working experience. The delineation of the region of interest (ROI) was based on the tumor–lung interface, with structures such as blood vessels and bronchi being excluded to the fullest extent. Tumors were outlined layer by layer, allowing for the automatic reconstruction of the 3D tumor volume. Thirty lesions were randomly selected for segmentation by another radiologist with 10 years of working experience to assess interobserver reliability. Subsequently, ROI radiomics features were extracted by SlicerRadiomics version 8426cdf—an extension for 3D Slicer. CT images were resampled to a pixel spacing of 1.0 mm in all three anatomical directions, and voxel intensity was discretized with a bin width of 25. Finally, 851 features were extracted, including 14 shape features, 18 first-order statistical features, 24 gray-level co-occurrence matrix (GLCM) features, 14 gray-level dependence matrix (GLDM) features, 16 gray-level run length matrix (GLRLM) features, 16 gray-level size-zone matrix (GLSZM) features, and 5 neighboring gray-tone difference matrix (NGTDM) features. Wavelet transformation produced another 744 features with 8 kinds of filters: LLL, LLH, LHL, HLL, LHH, HLH, HHL, HHH (L and H represent ‘low pass filter’ and ‘high pass filter’ in the x-axis, y-axis and z-axis). All radiomics features are listed in Table S2.

Feature selection and model building

Feature selection was performed in the training set. At first, variance threshold and mutual information were used to remove irrelevant features, and 635 features were retained. The least absolute shrinkage and selection operator (LASSO) was then applied for dimensionality reduction (38). The final 28 selected features and their coefficients are listed in Table S3.

Two robust ensemble learning methods, bagging and boosting, were combined to build the R model. This was a gradient boosting decision tree (GBDT) using a fine-tuned random forest classifier (RFC) model’s output as its initial input, which was considered to be a nested model. Five-fold cross-validation was employed on the training set to ensure model stability. The optimization goal was to choose the best area under the receiver operating characteristic (ROC) curve in the validation set, and the optimal parameters were decided via Bayesian optimization after the parameter space was determined using learning curves. Finally, we evaluated the models with the independent internal test set to verify the results of cross-validation. All these procedures were conducted with the Scikit-Learn library and hyperparameter optimization in a Python 3.8 environment. The building process is presented in Figure 2.

Figure 2 Model-building process. 3D, three dimensional; AUC, area under the curve; CT, computed tomography; DCA, decision curve analysis; FS, frozen section; HGP, high-grade pattern; ICC, intraclass correlation coefficient; LASSO, least absolute shrinkage and selection operator; LIME, local interpretable model-agnostic explanation; PR, precision-recall.

The FS model used intraoperative FS results of 0 (non-high-grade pattern, non-HGP) or 1 (HGP) as its output, which was consistent with its probability given its practical clinical use. The R model utilized the output probability from GBDT, and the optimal cutoff point, derived from the Youden Index, served as a threshold. In practice, as long as either FS or R model identified the presence of HGPs in lesions, we considered them to contain HGPs in order to minimize the risk of recurrence. The R-FS model employed a soft voting model fusion strategy, and the two models were assigned the same weights. Radiomics and FS represent two independent sources of information at different levels. The equal weighting aligns with the parallel diagnostic strategy, in which a positive result from either modality leads to an overall positive diagnosis, thereby maximizing sensitivity. The reference standard was FP for all models.

Model interpretability

To improve user-friendliness, we determined the post hoc explanations for the R model. First, permutation importance analysis (PIA) was used to determine which features played the most significant role in the overall predictions of the R model. Second, local interpretable model-agnostic explanations (LIME) were used to explain how individual features functioned for a given a specific sample. The Explain Like I’m 5 (ELI5) library and LIME library were used in a Python 3.8 environment.

Statistical analysis

The clinical and pathological data were processed and statistically analyzed with SPSS 25.0 software (IBM Corp., Armonk, NY, USA). Quantitative variables (age and max diameter) were described as the mean ± standard deviation (SD) or interquartile range (IQR) as appropriate. Other categorical variables were described as frequencies and proportions. The differences in clinical and pathological characteristics between patients with HGPs and non-HGPs were compared using the independent samples t-test for the quantitative variable, while the Pearson Chi-squared test or Fisher exact test were used for categorical variables. The DeLong test was performed on ROC curves of the three models using MedCalc statistical software version 22.032 (MedCalc Software, Ostend, Belgium). A two-sided P value <0.05 was considered statistically significant.

Results

Baseline characteristics

The baseline characteristics of the 490 involved patients are outlined in Table 1. Significant differences in maximum diameter, sex, smoking history, CT density characteristics, lymph node metastasis, visceral pleural invasion, lymphovascular invasion, spread through air spaces (STAS), and Ki-67 expression were observed in both the training and test sets (P<0.05). HGPs showed a strong positive correlation with other pathological factors associated with the high risk of recurrence, such as STAS and lymphovascular and pleural invasion (P<0.01) (39-41). Notably, there were no pGGNs containing HGPs while the majority of PSTs contained HGPs (65/75 and 16/17 in the training and test sets, respectively).

Construction and verification of the three models

The optimal parameters of the R model are listed in Table S4. For the original 107 features, there was excellent agreement between the two observers, with a mean intraclass correlation coefficient (ICC) value of 0.93±0.08 (Tables S5,S6). The performance of the three models on both the training and test datasets was evaluated via area under the curve (AUC), precision-recall (PR) curves, sensitivity, and specificity (Table 2).

The ROC and PR curves are shown in Figure 3. The Delong test (Table S7) indicated no statistically significant differences in the pairwise comparisons of the ROC curves between the R model and R-FS model; however, there were significant differences between the R model and FS model (P<0.001), as well as between the R-FS model and FS model (P<0.001). The PR curves in Figure 3 show that the R-FS model performed best in terms of precision and recall, indicating its superior ability to identify HGPs under most thresholds. Similarly, the FS model produced the lowest sensitivity but was markedly enhanced when integrated with the R model. We calculated the probability calibration curves (Figure S1) for the R model and R-FS model to determine the agreement between individual prediction and actual observation. The Brier scores of the R model (0.094) and R-FS model (0.121) indicated that both models made accurate predictions. The calibration curve of R model demonstrated the best fit. The clinical value of these two models was further evaluated via decision curve analysis (DCA) (Figure 4).

Figure 3 Model performance evaluation: ROC and PR curve. ROC curves with corresponding AUC values for the R, FS, and R-FS models built with the (A) training set and (B) test set. PR curves with corresponding AUC values for the R, FS, and R-FS models built with the (C) training set and (D) test set. AUC, area under the curve; FPR, false positive rate; FS, frozen section; R-FS, radiomics-frozen section; PR, precision-recall; R, radiomics; ROC, receiver operating characteristic; TPR, true positive rate.

Figure 4 DCA results for R and FS models built with the (A) training set and (B) test set. DCA, decision curve analysis; FS, frozen section; R, radiomics; R-FS, radiomics-frozen section.

Furthermore, given the different clinicopathological features of PSNs, pGGNs and PSTs, we compared the prediction performance across these subgroups (42). There were no pGGNs containing HGPs, and the accuracy of the three models for this type was 100%. We subsequently compared the learning ability between the PSN and PST groups. All models performed better on PSTs than on PSNs, as the predictive performance for PSNs showed slight variations on the test set (Figure 5).

Figure 5 Subgroup analyses. ROC of the R, FS, and R-FS models with the training and test sets across different solid component ratios. (A) Part-solid nodules in the training set. (B) Pure-solid tumors in the training set. (C) Part-solid nodules in the test set. (D) Pure-solid tumors in the test set. The data in parentheses representing the 95% confidence interval. AUC, area under the curve; FPR, false positive rate; FS, frozen section; R, radiomics; R-FS, radiomics-frozen section; ROC, receiver operating characteristic; TPR, true positive rate.

Analysis of model interpretability

The PIA involves systematically permuting the values of input features to evaluate their impact on a model’s performance, which clarifies the relative importance of each feature. A higher permutation importance score indicates greater reliance on a feature, while the standard deviation shows the stability of this importance evaluation across multiple trials. In Figure 6A, +, −, and 0 indicate positive, negative, or minimal contributions to the model, respectively. LIME was used to generate interpretable and locally faithful explanations for individual samples by fitting a locally approximated model around each instance, offering insights into the decision-making process. LIME trained a simple white-box estimator to closely approximate the black-box model within the vicinity of the original instances. Figure 6B contains an illustrative example of an explanation. To further explore the association between radiomic features and HGPs, we conducted an additional analysis, which is detailed in the Supplementary Material (Figure S2).

Figure 6 Model interpretability. (A) PIA for the whole R model (top 28 features). (B) LIME for a randomly selected sample. Colors indicate each feature’s effect on the model: green means positive, red means negative. The number represents the effect of the feature value on the prediction outcome for this sample. For example, when the value range of “original_glcm_Autocorrelation” is smaller than -0.78, it indicates that within this range, the feature significantly impacts the prediction result. HGP, high-grade pattern; LIME, local interpretable model-agnostic explanation; PIA, permutation importance algorithm; SD, standard deviation.

Discussion

In this study, we developed three models based on CT-derived radiomics and FS to distinguish HGPs in patients with T1 IAC. FS, despite being performed at the histological level, has low sensitivity, which significantly limits its utility. The R model and R-FS model both exhibited good performances and did not differ significantly. Their combination could identify a greater number of HGPs lesions as compared to FS alone. It is also worth noting that the R model alone demonstrated satisfactory discriminatory performance. This suggests that in clinical scenarios where FS sensitivity may be limited, such as in lesions with low-volume HGPs or in cases where tissue sampling is technically challenging, radiomics could be independently deployed to provide diagnostic reference and clinical decision-making support. Furthermore, a positive output from the R model could prompt pathologists to consider more tailored FS sampling strategies, such as broader or multi-site tissue sampling. In cases where sufficient family consent has been obtained and the patient’s physical condition permits, a more extensive surgical approach may be considered to reduce the likelihood of false-negative FS results postoperatively. In this way, R model can support for HGPs diagnosis and resection plan. Smoking history, younger age, and male sex were associated with HGPs, which is in line with previous research (43,44). Other pathological factors such as STAS, lymphovascular invasion, and visceral pleural invasion were also associated with a high risk of recurrence.

The R model is a GBDT based on a fine-adjusted RFC, which combines their advantages of enhancing precision and reducing variability. PIA and LIME, as compared with other popular interpretation libraries, such as Shapley additive explanations (SHAP), have an advantage of not being restricted to model structures but they may have higher computing source requirements. As shown in Figure 6A, wavelet-LLH_ngtdm_Strength (specified in Table S2) was the most important positive predictive factor for the R model. NGTDM relies on differences in gray levels among neighboring pixels to capture texture variances across different regions of an image. Higher values indicate greater local gray-level differences in the image, possibly corresponding to regions of coarseness or complex texture, which indirectly support heterogeneous growth in HGPs. A single sample explanation is presented in Figure 6B. We observed that original_glcm_Autocorrelation (specified in Table S2) accounted for the largest proportion in supporting the sample’s classification as non-HGPs. GLCM describes the spatial relationships between different pixel values in an image and measures the frequency of pairs of pixels occurring at certain distances and orientations, and the autocorrelation measures the extent to which identical gray-scale values repeat in the image; this is in line with recent findings and may explain why lepidic, acinar and papillary patterns (non-HGPs) exhibit more uniform density, as they involve less heterogeneous growth and invasiveness.

When it comes to clinical practice, the approximate risk probability intersections and their odds (equal to the harm-to-benefit ratio) were calculated via linear interpolations. In the range of the highest risk probability, it was found that the R model would produce greater net benefit than R-FS model when used as a predictive model. However, for clinicians who may tend to perform aggressive resections on potentially high-risk patients in order to achieve cure, they may adopt a smaller risk threshold in which the R-FS model would produce greater net benefit. It is worth noting that, in clinical settings, resection strategies are often complex and influenced by multiple factors, ranging from the patients’ individual condition, including anesthetic tolerance and comorbidities, to broader considerations such as the feasibility of postoperative adjuvant therapy and institutional practices in real-world oncology care. Whether to proceed with segmentectomy, lobectomy, or wedge resection still depends on the specific clinical context.

Through subgroup analysis of the model performance, we found that PSNs demonstrated greater variability between the training and test datasets. The inconsistent performance on PSNs may be attributed to challenges in segmentation, particularly due to the occasionally unclear boundary between ground-glass components and normal lung parenchyma, as well as the internal heterogeneity within the nodules. Given the correlation between the solid component ratio and tumor invasiveness, a stratified modeling strategy based on solid component proportions may enhance predictive robustness in future work.

The use of any predictive model needs to take the ethical context into consideration. First, our model was developed based on patients with cT1 LUAD, which limits its applicability to patients with early stages. Diagnostic bias may result from variation or errors in pathological assessment. Measurement bias may occur due to differences in CT acquisition protocols. Beyond enhancing transparency in data and algorithms, making model prediction results accessible and understandable to both clinicians and patients is crucial for promoting transparency in clinical decision-making. A key determinant of such transparency is the model’s predictive performance in prospective studies and under complex clinical scenarios.

There are some limitations to this study which should be acknowledged. First, the model was built and validated with data from a retrospective, single-center cohort. In the future, studies employing multicenter cohorts and a prospective design could be performed to further verify the clinical value of the proposed model. Second, this study did not take the cutoff values of HGP proportion for risk stratification into consideration (17). In our study, only the presence of HGP components was considered; more quantitative model output could be achieved with regression methods to provide a more precise evaluation of risk. Future studies involving larger cohorts and more detailed pathological quantification are needed to comply with clinical grading standards. Third, although the interobserver consistency was evaluated in this work, the reproducibility from a single observer was not considered.

Conclusions

Radiomics is an effective tool that has considerable valuable in complementing intraoperative FS examination and may be particularly effective in improving the sensitivity of FS. The R-FS model represents a practical and reliable clinical tool, capable of guiding clinical decision-making and facilitating precision medicine.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-504/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-504/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-504/prf

Funding: This work was supported by the National Natural Science Foundation of China (Nos. 82071990, 81571629, 81301218, 82471977 and 82271989) and the Shanghai Anticancer Association EYAS Project (No. SACACY22C15).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-504/coif). L.K. received grants or has contracts with AstraZeneca, AMGEN and Art Tempi, receives honoraria from AstraZeneca, German Cancer Society and Art Tempi, and received support for attending meetings and travel from AstraZeneca and ELCC. 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 study was approved by the Ethics Committee of Zhongshan Hospital, Fudan University (approval No. B2021-128) and informed consent was waived from all the patients due to the retrospective nature of this study.

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. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin 2024;74:12-49. [Crossref] [PubMed]
3. de Koning HJ, van der Aalst CM, de Jong PA, et al. Reduced Lung-Cancer Mortality with Volume CT Screening in a Randomized Trial. N Engl J Med 2020;382:503-13. [Crossref] [PubMed]
4. Ginsberg RJ, Rubinstein LV. Randomized trial of lobectomy versus limited resection for T1 N0 non-small cell lung cancer. Lung Cancer Study Group. Ann Thorac Surg 1995;60:615-22; discussion 622-3. [Crossref] [PubMed]
5. Altorki N, Wang X, Kozono D, et al. Lobar or Sublobar Resection for Peripheral Stage IA Non-Small-Cell Lung Cancer. N Engl J Med 2023;388:489-98. [Crossref] [PubMed]
6. Saji H, Okada M, Tsuboi M, et al. Segmentectomy versus lobectomy in small-sized peripheral non-small-cell lung cancer (JCOG0802/WJOG4607L): a multicentre, open-label, phase 3, randomised, controlled, non-inferiority trial. Lancet 2022;399:1607-17. [Crossref] [PubMed]
7. Kamel MK, Lee B, Harrison SW, et al. Sublobar resection is comparable to lobectomy for screen-detected lung cancer. J Thorac Cardiovasc Surg 2022;163:1907-15. [Crossref] [PubMed]
8. Altorki NK, Wang X, Wigle D, et al. Perioperative mortality and morbidity after sublobar versus lobar resection for early-stage non-small-cell lung cancer: post-hoc analysis of an international, randomised, phase 3 trial (CALGB/Alliance 140503). Lancet Respir Med 2018;6:915-24. [Crossref] [PubMed]
9. Jeon YJ, Lee J, Shin S, et al. Prognostic impact of micropapillary and solid histological subtype on patients undergoing curative resection for stage I lung adenocarcinoma according to the extent of pulmonary resection and lymph node assessment. Lung Cancer 2022;168:21-9. [Crossref] [PubMed]
10. Fan X, Liang C, Ma X, et al. Clinical, Imaging, and Pathological-Molecular Characteristics Associated with Stage IA Invasive Lung Adenocarcinoma Recurrence After Sub-lobar Resection. Acad Radiol 2025;32:450-9. [Crossref] [PubMed]
11. Travis WD, Brambilla E, Noguchi M, et al. International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory Society: international multidisciplinary classification of lung adenocarcinoma: executive summary. Proc Am Thorac Soc 2011;8:381-5. [Crossref] [PubMed]
12. Murakami S, Ito H, Tsubokawa N, et al. Prognostic value of the new IASLC/ATS/ERS classification of clinical stage IA lung adenocarcinoma. Lung Cancer 2015;90:199-204. [Crossref] [PubMed]
13. Qian F, Yang W, Wang R, et al. Prognostic significance and adjuvant chemotherapy survival benefits of a solid or micropapillary pattern in patients with resected stage IB lung adenocarcinoma. J Thorac Cardiovasc Surg 2018;155:1227-1235.e2. [Crossref] [PubMed]
14. Cha MJ, Lee HY, Lee KS, et al. Micropapillary and solid subtypes of invasive lung adenocarcinoma: clinical predictors of histopathology and outcome. J Thorac Cardiovasc Surg 2014;147:921-928.e2. [Crossref] [PubMed]
15. Lee G, Lee HY, Jeong JY, et al. Clinical impact of minimal micropapillary pattern in invasive lung adenocarcinoma: prognostic significance and survival outcomes. Am J Surg Pathol 2015;39:660-6. [Crossref] [PubMed]
16. Cheng R, Hao Z, Qiu L, et al. The impact of postoperative adjuvant therapy on EGFR-mutated stage IA lung adenocarcinoma with micropapillary pathological subtypes. World J Surg Oncol 2024;22:235. [Crossref] [PubMed]
17. 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]
18. Bossé Y, Gagné A, Althakfi W, et al. Prognostic value of complex glandular patterns in invasive pulmonary adenocarcinomas. Hum Pathol 2022;128:56-68. [Crossref] [PubMed]
19. Dai C, Xie H, Kadeer X, et al. Relationship of Lymph Node Micrometastasis and Micropapillary Component and Their Joint Influence on Prognosis of Patients With Stage I Lung Adenocarcinoma. Am J Surg Pathol 2017;41:1212-20. [Crossref] [PubMed]
20. Cao S, Teng J, Xu J, et al. Value of adjuvant chemotherapy in patients with resected stage IB solid predominant and solid non-predominant lung adenocarcinoma. Thorac Cancer 2019;10:249-55. [Crossref] [PubMed]
21. Tsao MS, Marguet S, Le Teuff G, et al. Subtype Classification of Lung Adenocarcinoma Predicts Benefit From Adjuvant Chemotherapy in Patients Undergoing Complete Resection. J Clin Oncol 2015;33:3439-46. [Crossref] [PubMed]
22. Zhao K, Yang L, Liu L, et al. Real-world efficacy of adjuvant therapy for totally resected stage I lung adenocarcinoma patients with pathological high-risk factors: propensity score analysis. BMC Surg 2024;24:140. [Crossref] [PubMed]
23. Li Y, Zhao J, Zhao Y, et al. Survival benefit of adjuvant chemotherapy after resection of Stage I lung adenocarcinoma containing micropapillary components. Cancer Med 2024;13:e7030. [Crossref] [PubMed]
24. Liu S, Wang R, Zhang Y, et al. Precise Diagnosis of Intraoperative Frozen Section Is an Effective Method to Guide Resection Strategy for Peripheral Small-Sized Lung Adenocarcinoma. J Clin Oncol 2016;34:307-13. [Crossref] [PubMed]
25. Thoracic Surgery Branch of Shanghai Medical Association, Thoracic Surgeons Branch of Shanghai Medical Doctor Association, The Specialist Alliance of Clinical Skills Promotion and Enhancement for General Thoracic Surgery. Expert consensus of thoracic surgeons on guiding surgical decision-making based on intraoperative frozen sections for peripheral pulmonary nodules with diameter≤2 cm. Chinese Journal of Clinical Thoracic and Cardiovascular Surgery 2022;29:667-75.
26. Yeh YC, Nitadori J, Kadota K, et al. Using frozen section to identify histological patterns in stage I lung adenocarcinoma of ≤ 3 cm: accuracy and interobserver agreement. Histopathology 2015;66:922-38. [Crossref] [PubMed]
27. Trejo Bittar HE, Incharoen P, Althouse AD, et al. Accuracy of the IASLC/ATS/ERS histological subtyping of stage I lung adenocarcinoma on intraoperative frozen sections. Mod Pathol 2015;28:1058-63. [Crossref] [PubMed]
28. Rios Velazquez E, Parmar C, Liu Y, et al. Somatic Mutations Drive Distinct Imaging Phenotypes in Lung Cancer. Cancer Res 2017;77:3922-30. [Crossref] [PubMed]
29. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6. [Crossref] [PubMed]
30. Wang X, Ma C, Jiang Q, et al. Performance of deep learning model and radiomics model for preoperative prediction of spread through air spaces in the surgically resected lung adenocarcinoma: a two-center comparative study. Transl Lung Cancer Res 2024;13:3486-99. [Crossref] [PubMed]
31. McKinney SM, Sieniek M, Godbole V, et al. International evaluation of an AI system for breast cancer screening. Nature 2020;577:89-94. [Crossref] [PubMed]
32. Dong H, Yin L, Chen L, et al. Establishment and validation of a radiological-radiomics model for predicting high-grade patterns of lung adenocarcinoma less than or equal to 3 cm. Front Oncol 2022;12:964322. [Crossref] [PubMed]
33. Zhong Y, Cai C, Chen T, et al. PET/CT-based deep learning grading signature to optimize surgical decisions for clinical stage I invasive lung adenocarcinoma and biologic basis under its prediction: a multicenter study. Eur J Nucl Med Mol Imaging 2024;51:521-34. [Crossref] [PubMed]
34. 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]
35. Huo J, Min X, Luo T, et al. Computed tomography-based 3D convolutional neural network deep learning model for predicting micropapillary or solid growth pattern of invasive lung adenocarcinoma. Radiol Med 2024;129:776-84. [Crossref] [PubMed]
36. 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]
37. Bertoglio P, Aprile V, Ventura L, et al. Impact of High-Grade Patterns in Early-Stage Lung Adenocarcinoma: A Multicentric Analysis. Lung 2022;200:649-60. [Crossref] [PubMed]
38. Hepp T, Schmid M, Gefeller O, et al. Approaches to Regularized Regression - A Comparison between Gradient Boosting and the Lasso. Methods Inf Med 2016;55:422-30. [Crossref] [PubMed]
39. Kadota K, Nitadori JI, Sima CS, et al. Tumor Spread through Air Spaces is an Important Pattern of Invasion and Impacts the Frequency and Location of Recurrences after Limited Resection for Small Stage I Lung Adenocarcinomas. J Thorac Oncol 2015;10:806-14. [Crossref] [PubMed]
40. Yun JK, Lee GD, Choi S, et al. Comparison of prognostic impact of lymphovascular invasion in stage IA non-small cell lung cancer after lobectomy versus sublobar resection: A propensity score-matched analysis. Lung Cancer 2020;146:105-11. [Crossref] [PubMed]
41. Koike T, Koike T, Yoshiya K, et al. Risk factor analysis of locoregional recurrence after sublobar resection in patients with clinical stage IA non-small cell lung cancer. J Thorac Cardiovasc Surg 2013;146:372-8. [Crossref] [PubMed]
42. Ye T, Deng L, Wang S, et al. Lung Adenocarcinomas Manifesting as Radiological Part-Solid Nodules Define a Special Clinical Subtype. J Thorac Oncol 2019;14:617-27. [Crossref] [PubMed]
43. Fujikawa R, Muraoka Y, Kashima J, et al. Clinicopathologic and Genotypic Features of Lung Adenocarcinoma Characterized by the International Association for the Study of Lung Cancer Grading System. J Thorac Oncol 2022;17:700-7. [Crossref] [PubMed]
44. Gao Z, Liu S, Xiao H, et al. Development and validation of a clinical decision tool for preoperative micropapillary and solid pattern lung adenocarcinoma of CT ≤2 cm. Int J Surg 2024;110:7607-15. [Crossref] [PubMed]