This retrospective study, registered at http://clinicaltrials.gov (identifier: NCT05400304), obtained approval from the institutional review boards and the Ethics Committee of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (identifier: UHCT22749). Informed consent requirements were waived.

This study employed the following inclusion criteria: (1) preoperative CT examination indicating a lesion smaller than 3 cm, and (2) Patients with postoperative pathological confirmation of lung adenocarcinoma. The following exclusion criteria were applied: (1) pure ground glass nodules, (2) patients who underwent chemoradiotherapy, (3) patients who underwent CT-guided biopsy prior to the CT examination, (4) Preoperative CT examination was conducted within 2 weeks prior to surgery, and no lesion puncture was performed. For cases with multiple nodules, only the pulmonary nodules with definitive pathological results were included for subsequent analysis. The process of patient inclusion in this study is depicted in Figure 1.

This study collected clinical and pathological features such as age, gender, smoking history, semantic characteristics, pathological type, lymph node metastasis, neurovascular invasion, pleural invasion, and STAS of patients for analysis. Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines were followed in the study (19). The construction process of radiomics models is evaluated using the radiomics quality score (RQS) (20).

After fixation in 10% formalin, the specimens were embedded in paraffin using conventional methods for subsequent pathological diagnosis by a pathologist. Following staining with hematoxylin and eosin, the tumor section was assessed under a multiheaded microscope. The diagnosis of STAS aligns with the 2015 World Health Organization pathological classification of lung tumors (21). STAS infiltrates the alveoli of the peripheral lungs in the form of clusters of micropapillary, solid nests, or single tumor cells.

The study utilized CT scans with a layer thickness ranging from 0.625 to 1.25 mm, without the administration of contrast medium. The bone reconstruction algorithm was employed for reconstruction purposes. Supplementary S1 contains detailed information regarding the acquisition and reconstruction parameters. The chest CT images underwent retrospective analysis by CT experts, possessing 5 and 10 years of experience, using a window width of 1,600 HU and a window level of −600 HU. This analysis aimed to determine semantic characteristics. The experts performed manual layer-by-layer segmentation of the tumor outline using ITK-SNAP (version 3.8.0, available at http://www.itksnap.org/). Any differences in opinion among the experts were resolved through discussion to achieve a consensus. Subsequently, the senior radiologist assessed the quality of the region of interest (ROI) and made necessary adjustments after the primary radiologist completed tumor lesion segmentation. To increase the robustness, 50 cases were randomly selected to estimate the intraclass correlation coefficients (ICCs), with a value of ≥0.75 indicating robustness. Figure 2 provides an illustration of the overall research design.

The “PyRadiomics” package in Python software was utilized for extracting tumor ROI features. This encompassed the extraction of shape features, grayscale features, wavelet transform features, and texture features. Supplementary S2 displays the classification and quantity of extracted radiomics features.

Following feature extraction, all radiomics features were standardized using the z-score for subsequent analysis. Moreover, we performed a statistical test employing the Mann-Whitney U-test to assess the significance of the features, selecting only those with a p-value < 0.05 for further consideration. We calculated the Spearman correlation coefficient to assess the correlation between features with high repeatability. If the correlation coefficient between two features exceeds 0.9, we keep only one feature to prevent redundancy. Feature selection involves utilizing the Least Absolute Shrinkage and Selection Operator (LASSO) to reduce feature dimensionality. The final radiomics features were inputted into the machine learning model, LightGBM.

The univariable and multivariable regression analysis was employed to analyze clinical and semantic features, and the variables ultimately included in the model demonstrated significant associations with the STAS status. The LightGBM was used for clinical model construction. To enhance the predictive ability, a comprehensive model was constructed by integrating radiomics signature with clinical signature. To evaluate the stability of the constructed models, we validated both the radiomics model and the clinical model across multiple cohorts.

SHAP is a method for interpreting prediction results based on Shapley value theory, which decomposes the prediction results into the influence of each feature, providing global and local interpretability for the model. The core idea of SHAP is to allocate the contribution of feature values to different features, calculate the Shapley value of each feature, and multiply it with the feature value to obtain the contribution of that feature to the prediction result. SHAP can be used in machine learning models to generate visualized and quantitative interpretation results, helping doctors explain the decision-making process of the model. This study uses SHAP to interpret and visualize the LightGBM model.

In order to measure the accuracy of the models, receiver operating characteristic curve (ROC) were plotted to determine their diagnostic performance. Calibration curves were employed to evaluate the calibration performance of the models, and their calibration ability was further examined using the Hosmer-Lemeshow test. Additionally, the clinical utility of the predictive models was assessed through decision curve analysis (DCA). The area under the receiver operating characteristic curve (AUC), negative predictive value (NPV), positive predictive value (PPV), specificity, and sensitivity, were employed to compare the diagnostic performances of the three models in different cohorts.

The statistical analysis was conducted using SPSS and Python. The quantitative data obtained from the clinical and imaging features of the patients underwent normality tests using the Kolmogorov-Smirnov test and homogeneity of variance tests using Levene's test. Independent sample T-tests were utilized for data that met the assumptions of normality and homogeneity of variance. In cases where the assumptions were not met, Mann-Whitney U-tests were employed. For classification data, chi-square tests or Fisher tests were applied. Multivariate logistic regression analysis is employed to incorporate clinical variables into the construction of clinical models. The Spearman correlation coefficient to assess the correlation between features. The statistical significance of the results was assessed using a two-tailed p-value threshold of less than 0.05.

The study included 1,325 patients, 1,168 from hospital 1, 93 from hospital 2, and 64 from hospital 3. Within the group of patients from hospital 1, the study divided them into two cohorts: a training cohort (n = 930) and a testing cohort (n = 238). External validation 1 involved 93 patients from hospital 2, while validation 2 encompassed 64 patients from hospital 3. Clinical and semantic features of the patients are summarized in Table 1. Within the training group, consisting of 930 patients, a division was made based on the presence or absence of STAS. Statistical analysis was conducted to evaluate the clinicopathologic characteristics, as presented in Table 2. Logistic regression analyses demonstrated that gender, nodule type, tumor size, and pleural indentation independently predicted the presence of STAS, as indicated in Table 3.

A total of 833 radiomics features were extracted from the tumor ROI using PyRadiomics. 721 features were determined to be reliable after assessing their reproducibility using ICCs. Performing the Mann-Whitney U test resulted in the identification of 599 radiomic features. Subsequent analysis using the Spearman rank correlation test reduced the number of features to 196. By using the LASSO algorithm to reduce the dimensionality of features, 24 non zero coefficient radiomics features were ultimately selected. Figure 3A–C present the radiomics features with non-zero coefficients, as determined through LASSO penalized regression analysis. The performance of the radiomics model is depicted in Figure 3D, showing AUCs of 0.836 (95% CI: 0.809–0.863) and 0.755 (95% CI: 0.693–0.817) in the training cohort and the testing cohort. The corresponding DCA plot, displayed in Figure 3E, illustrated that the model consistently offers greater benefits in the majority of scenarios. Furthermore, Figure 3F showcased the confusion matrix for the radiomics model in the testing cohort, providing details such as an accuracy of 0.714, specificity of 0.645, sensitivity of 0.789, NPV of 0.769, and PPV of 0.672.

Multivariable logistic regression analyses demonstrated that gender, nodule type, tumor size, and pleural indentation independently predicted the presence of STAS. These four clinical semantic features were used to construct a clinical model integrated into the LightGBM algorithm.

Figure 3G displays the performance of the clinical model, with AUCs of 0.775 (95% CI: 0.743–0.806) in the training cohort and 0.743 (95% CI: 0.680–0.806) in the testing cohort. Figure 3H showcases the predictive capability of the models in forecasting STAS, emphasizing their net benefit. Figure 3I presents the confusion matrix for the clinical model in the testing cohort, revealing an accuracy of 0.718, specificity of 0.847, sensitivity of 0.579, NPV of 0.686, and PPV of 0.776. Additionally, the combined model achieved AUCs of 0.837 (95% CI: 0.810–0.863) in the training cohort and 0.768 (95% CI: 0.707–0.829) in the testing cohort.

In both the training and testing cohorts, the clinical model, radiomics model, and combined model were compared. The clinical model achieved an AUC of 0.775 in the training cohort and 0.743 in the testing cohort, as depicted in Figure 4A,D, respectively. Similarly, the radiomics model achieved an AUC of 0.836 in the training cohort and 0.755 in the testing cohort. The combined model, which integrated radiomics features, demonstrated an AUC of 0.837 in the training cohort and 0.768 in the testing cohort.

Figure 4B,E present the DCA of the three models, indicating that the combined model provided a net benefit in predicting STAS. Calibration curves in Figure 4C,F demonstrated agreement between the predicted and observed STAS in both cohorts. The Hosmer–Lemeshow p-values for the clinical model, radiomics model, and combined model were 0.421, 0.738, and 0.704, respectively. The DeLong test results indicated that the combined model outperformed the clinical model (p < 0.05), while no statistically significant differences were observed between the combined model and the radiomics model.

The AUC of clinical model was 0.717 (Figure 5A) in the validation 1 cohort, whereas the radiomics model (Figure 5B) and the combined model (Figure 5C) achieved AUCs of 0.758 and 0.765, respectively. In the validation 2 cohort, The AUC of the clinical model (Figure 5D) was 0.725, while the radiomics model (Figure 5E) and the combined model (Figure 5F) achieved AUCs of 0.726 and 0.746, respectively. The performance of the clinical model, radiomics model, and combined model in the training cohort, testing cohort, validation 1 cohort, and validation 2 cohort is summarized in Table 4. This study received 18 RQS points and achieved a total score of 50%. According to Supplementary S4, this signature is classified as TRIPOD 3.

SHAP is used to visualize the global interpretation of each feature contribution in the LightGBM model. In Figure 6A, Arrange the importance of features from top to bottom, and the horizontal axis of the graph displays the SHAP values of the features, with each point representing a patient. The color of a point is determined by its feature value, with the redder the color, the higher the feature value, and the bluer the color, the lower the feature value. Wavelet-LLL_glcm_ClutureShde is considered the most important feature. As shown in Figure 6B, the waterfall visualization LightGBM model generated by SHAP is used to describe the decision-making process of whether two lung adenocarcinoma patients have STAS. Based on the contribution of each feature to the decision, all features are arranged in order, and the direction of their contribution is displayed by color. The score calculation starts from E[f (x)], and then the SHAP values are added together. Red indicates an increase in the probability of STAS, blue indicates a decrease in the probability of STAS, and ends with individual prediction.