A total of 1051 patients with Stage IA adenocarcinoma (T1a-cN0M0) confirmed by curative surgery between September 2015 and July 2021 in Wuhan Union Hospital were retrospectively assessed. Then, 126 patients were excluded according to the following exclusion criteria (1): previous chemoradiotherapy (n=45); (2) a history of lung operation (n=17); (3) no thin-section CT before treatment (n=44); or (4) no plain chest CT imaging (n=30, Figure 1). As shown in Figure 1, the incidence of STAS in our institution was relatively low (169/923,18.3%). To overcome potential data imbalance, we randomly divided the STAS-negative cases by 3:7 into groups and matched them with the STAS-positive group at a nearly 1:1 ratio. Such method for balancing data has been validated in previous studies (11, 18, 20). Due to the fact that the incidence of STAS negative is higher than that of positive cases, we divided the negative data by a ratio of 3:7(226 cases) rather than 2:8(150 cases) to ensure the STAS negative cases were slightly larger than the positive ones. In addition, we had reperformed these models by 2:8 ratio for divide the negative data, and found that this ratio does not affect our main conclusions. In total, 395 patients (226 STAS negative and 169 STAS positive cases) were included, which were randomly assigned into training database (316 patients) and test dataset (79 patients), with a ratio of 0.8:0.2. In the train cohort, 136 patients were presented as STAS positive while 180 were negative. In the test cohort, 33 cases were positive for STAS while 46 were negative for STAS. All included patients had single lung adenocarcinoma.

The patients’ clinical characteristics and pathological findings, including age, sex, smoking history, histologic subtypes, surgical margin, lymphatic metastasis, epidermal growth factor receptor (EGFR) mutations and anaplastic large-cell lymphoma kinase (ALK) were recorded. A history of smoking was defined as lifetime exposure to more than 100 cigarettes. TNM staging was performed according to the IASLC 8th TNM Lung Cancer Staging System (21).

Two pathologists (JF and NW, with 4 and 10 years of experience in thoracic pathology, respectively) blinded to clinical findings re-evaluated the hematoxylin and eosin-stained slides of all included patients using a multiheaded microscope. The predominant subtype of lung adenocarcinoma was assessed based on the International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory Society multidisciplinary classification of LUAD. STAS positivity was defined according to the WHO definition of STAS as tumor cells were found in the lung air spaces beyond the edge of the primary tumor, which is mainly composed of the following three forms: (1) air spaces filled by micropapillary structure without central fibrovascular cores; (2) air spaces filled by the solid component of the tumor; (3) air spaces filled by multiple discrete and single cells(1). After independent assessments, differences were resolved by consensus.

CT was performed on two multislice spiral CT scanners (SOMATOM Definition AS+ and Siemens Healthineers, Germany) at our institution, using the following parameters: detector collimation widths, 64 x 0.6 mm and 128 x 0.6 mm; tube voltages, 120 kV. The tube current was regulated by an automatic exposure control system (CARE Dose 4D). Images were reconstructed at a slice thickness of 1.5 mm or 1 mm and an interval of 1.5 mm or 1 mm. No contrast medium was used. Digital Imaging and Communications in Medicine (DICOM) images from the picture archiving and communication system (PACS) were imported to the 3D-slicer software.

Two senior radiologists (HSS and JLZ, with 31 and 25 years of experience in thoracic radiology, respectively) evaluated the CT images to determine tumor size (longest diameter in MPR images), tumor density type (solid, mix ground-glass opacities [mGGO], and pure GGO [pGGO]), and GGO ratio (GGO diameter/tumor diameter) in consensus on the PACS. These specific CT features were chosen as the most contributing risk factors for STAS in patients with LUAD (8–11). The above two radiologists blinded to clinical and histologic findings assessed CT features on both axial CT and multiplanar reconstruction (MPR) images.

The regions of interest on CT images were semi-automatically delineated layer by layer by three junior radiologists (XH, YZ and JX with 5, 3 and 2 years of experience in thoracic imaging, respectively). All the three radiologists were aware of tumor presence and location but unaware of the pathological reports and STAS status. The 3D-slicer software was used for segmenting the lesions on each slice of CT scans semi-automatically and independently. Then, the three-dimensional volumes of interest (3D-VOIs) of tumors were automatically reconstructed with the 3D-slicer software (Figure 2). Two senior radiologists (HSS and JLZ) were responsible for checking all tumor segmentations, and any deviations were addressed with additional corrections. As reported in a recent publication (22), the semi-automatic segmentation from the 3D-Slicer is a better alternative to manual segmentation, as it can produce more robust and reproducible radiomic features. In addition, we randomly selected 40 cases at a ratio of 1:10 of all sample (396 cases) for estimating intraclass correlation coefficients (ICCs) analysis, which could ensure the precision and assurance of results (23). One observer (XH) repeated the segmentation after one week for intra-observer variability analysis. The other observer (YZ) performed the segmentation on the same image set using the same method, for inter-observer variability analysis.

The Philips Radiomics Tool (Philips Healthcare, China) was used for radiomics feature extraction, and core feature calculation was based on pyRadiomics (24). A B-spline curve interpolation algorithm was used to resample each 3D CT image to a spacing of (0.7, 0.7, 1.0) mm. For each VOI, a total of 1037 3D-radiomic features, including direct, wavelet transformed, logarithm transformed and gradient filtered features, were extracted (types and numbers are shown in Figure 2, and details can be also found at pyradiomics.readthedocs.io/en/latest/features.html).

The SPSS software (SPSS, version 21, IBM, Chicago, IL, USA) and R (version 4.0.2; http://www.Rproject.org) were used for all statistical analyses. LASSO binary logistic regression was performed with the ‘glmnet’ package. Multivariate binary logistic regression was carried out with the ‘rms’ package. Decision curve analysis was performed with the ‘rmda’ package. Receiver operating characteristic (ROC) analysis was carried out with the ‘pROC’ package. Descriptive analysis was performed to describe the distribution of the variables of interest for the training and testing cohorts. All continuous data and categorical variables were expressed as mean ± standard deviation and frequency (percentage), respectively. Independent samples Student’s t test was applied to compare continuous data in two groups, and the chi-square test was used to compare categorical variables. P<0.05 (two-tailed) was considered statistically significant.

All radiomics features were normalized to the z-score. Intraclass correlation coefficient (ICC) determination and Pearson’s correlation analysis were performed to exclude redundant and unstable features (r>0.8, ICC>0.75). ICC < 0.5, between 0.5 and 0.75, between 0.75 and 0.9, and > 0.90 indicated poor, moderate, good, and excellent reliability, respectively (25). Least absolute shrinkage and selection operator (LASSO) analysis was performed to determine features for further assessment (26). Five-fold cross-validation and maximum area under the curve (AUC) were used as feature filtering criteria. A classification model based on the selected features was established by a multivariate logistic regression (LR) algorithm. Feature selection was performed on the training set. In this work, we selected no model using the one-standard-error rule as the final model, because a five-fold cross-validation LASSO analysis was performed to filter features. A temporary model was built behind the LASSO analysis as to preserve features that are significant enough for the feature coefficients. More features kept after the LASSO analysis reflect a better performance for the final multivariate logistic regression algorithm (27). Three models were built based on three sets of features: a radiomics model based on the most contributing radiomics features from the temporary model, a clinical and CT model based on clinical and CT features only, and a mixed model (MixModel) combined the Rad-score with clinical and CT characteristics. The performances of these models for predicting STAS were evaluated in the training set, then in the test set by plotting ROC curves and calculating the areas under the curves (AUCs). The accuracy, sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) were then calculated for each model. The predictive ability was also illustrated by the confounder matrix. The DeLong test was conducted to compare diagnostic efficiency among the different models. Furthermore, decision curve analysis was preformed to determine the clinical usefulness of the three models by quantifying the net benefits at different threshold probabilities in the data set.

This study included 395 eligible patients (average age, 59 ± 10 years; 207 males) in total. The clinicopathological and main CT features of the final study population are shown in Table 1. STAS was found in 169/928 (18.2%) patients. No significant differences were found between STAS-positive and STAS-negative cases in age, gender and smoking history (p=0.268, p=0.232 and p=0.053, respectively). Concerning CT features, tumors with STAS tended to be larger than STAS negative counterparts (21 ± 6.3 mm vs 18 ± 6.7 mm, p<0.001). Tumor densities differed between the STAS and non-STAS groups (p<0.001). The majority of STAS-positive tumors manifested as solid nodules (152/169, 89.9%; Figure 3), followed by the mix GGO (13/169, 7.7%) and pure GGO (4/169, 2.4%) groups. Moreover, the less the GGO ratio, the higher the possibility of STAS positivity (0.08 ± 0.21 vs 0.38 ± 0.42, p<0.001). Then, age, gender, smoking history, diameter, density and GGO ratio were selected for predicting STAS in the clinical and CT model (Clinical-CT Model) by the LR algorithm (Table 2). Based on clinical and CT features, the model had an AUC of 0.721 (a sensitivity of 69.9% and a specificity of 61.7%) in the training cohort; the AUC was 0.804 (a sensitivity of 72.7% and a specificity of 76.1%) in the test cohort (Figure 4). The distribution of the selected clinical and CT features for the patients with STAS and those without STAS in the training and test cohorts are shown in Table 3.

Pathologically, the histologic subtypes differed between the STAS and non- STAS groups (p<0.001). Specifically, STAS-positive tumors had a lower frequency of lepidic predominant subtypes (2.4% vs 15%), while STAS-negative ones had a lower rate of solid predominant (6.2% vs 17.8%). The positive rate of resection margin was 5.6% (22/395) in this cohort and tended to occur more in STAS-positive tumors (8.9% vs 3.1%, p=0.013). Pleural invasion was observed more frequently in patients with tumors positive for STAS versus the STAS-negative group (14.8% vs 5.3%, p=0.001), whereas perineural invasion showed no significant difference (5.3% vs 3.1%, p=0.266). EGFR and ALK analyses were available in 261 and 303 patients, respectively. STAS positivity was associated with reduced incidence of EGFR (p=0.039) and higher incidence of ALK (p=0.009). Lymphatic metastasis was negative in all patients.

After intraclass correlation coefficients (ICC) and Pearson’s correlation analysis, 98 radiomics features were selected for predicting STAS. Based on LASSO penalized logistic regression analysis, 18 features showed significant associations between radiomics and STAS (Figures 5A, B). Then, the top seven radiomic features with coefficients greater than 0.1 (two first-order and five second order parameters, including GLCM, GLSZM and GLDM features) were identified by the LR model (Figure 5C).

Finally, these seven radiomics features were entered into the LR model for building the radiomics model. Features contained in the radiomics model, and their coefficients are shown in Table 4. The radiomics model achieved good performance both in the training cohort(AUC: 0.812, sensitivity: 75%, and specificity: 78.3%) and the test cohort(AUC:0.850,; sensitivity:75.8%, and specificity:76.1%) (Figure 4).

The rad-score of each lesion was calculated using the following formula:

The rad scores for both the training and test sets are shown in Figure S1. STAS-positive tumors had significantly higher rad-scores than STAS-negative tumors in both the training and validation sets (p <0.001).

The MixModel of a comprehensive nomogram model was developed with the retain clinical CT characteristics(including age, gender, smoking history, diameter, density and GGO ratio) and rad-score(Figure 6), and showed AUCs of 0.822 and 0.865 in the training and test cohorts, respectively. Features contained in Mixmodel and their coefficients are listed in Table 5. Subsequently, we separately compared AUCs among MixModel, clinical-CT model and radiomics model (Figure 4). Both in the training cohort and the test cohort, the MixModel showed improvement in diagnostic ability compared with the clinical-CT model and radiomics model. In addition, the AUC of radiomics model was larger than that of the clinical-CT model in both cohorts (0.812 vs 0.721, p<0.001; 0.850 vs 0.804, p=0.228).

The results of confounder matrix analysis in the training and test sets in MixModel, and the clinical-CT and radiomics models are summarized in Table 6. The Mixmodel model showed the highest accuracy (number of correct predictions divided by that of total predictions) among the three models in the training set and in the test set.

For the radiomics model, calibration curve analysis showed P values of 0.954 and 0.792 in the training and test sets, respectively, indicating a good degree of fit for the model in both sets (Figure S2). Similarly, the clinical-CT model also showed good calibration abilities in both the training and test sets (Figure S2). The clinical usefulness of the three predictive models were examined by decision curve analysis (Figure S3). Compared with the treat-all and treat-none models, MixModel, and the clinical-CT and radiomics models could bring net benefits to patients, among which the radiomics model had the best benefit, while the clinical-CT model had the lowest.