A total of 512 patients with primary lung adenocarcinoma confirmed by surgical resection and pathology were collected continuously and retrospectively at our institution from January 2017 to May 2020. The interval between CT scan and surgery is within 2 months. Patients with the following criteria were excluded: atypical hyperplasia and adenocarcinoma in situ (n = 172), lack of preoperative non-contrast CT images or obvious CT artifacts (n = 37), previous lung surgery or preoperative adjuvant chemotherapy (n = 32), pathological diameter >5 cm (n = 38), and mucinous adenocarcinoma and mucinous carcinoma (n = 17). There were 216 patients in the final cohort (males, 125), which included 56 STAS+ and 160 STAS− adenocarcinomas. This retrospective study was approved by the institutional review committee and the requirement for informed consent was waived.

To evaluate the performance of the best radiomic model, we used 46 consecutive external data from another hospital as the testing set in accordance with the aforementioned inclusion criteria. The inclusion and exclusion criteria of the external set were consistent with that of the cohort collected in our center. The external data included 15 STAS+ adenocarcinoma (eight men; mean age, 52 ± 8.7 years) and 31 STAS− adenocarcinoma(21 men; mean age, 54 ± 7.6 years).

CT images were acquired from General Electric (LightSpeed VCT; Waukesha, Wis) or Siemens (Definition Flash, Erlangen, Germany). CT parameters were as follows: tube voltage, 120 kVp; tube current, 150–200 mAs; pitch, 0.984:1 or 1.0; reconstructed thickness and interval, 1.25 mm or 1 mm. All images were reviewed and measured with a standard lung window (window width, 1500 HU; window level, −500 HU) and a mediastinal window (window width, 350 HU; window level, 50 HU).

The size of the entire lesion was defined as the average of long and short axial diameters. The consolidation tumor ratio (CTR) was defined as the proportion of the maximum consolidation diameter divided by the maximum tumor diameter. The CT morphologic features were evaluated as follows: nodule density (solid, part solid, or pure ground glass), location (upper or not), and shape (round to oval, irregular), vascular change (normal, convergent or dilated), cystic airspaces, tumor-lung interface (clear or unclear), lobulation, spiculation, and satellites around the lesion. Morphological analysis was performed by two radiologists with 10 and 20 years of experience in chest radiologic diagnosis (LQ and ML), and they were blinded to the pathological results. Any disagreements between them were resolved by reaching a consensus.

Surgically resected specimens were fixed with formalin and stained with hematoxylin–eosin. According to the 2015 WHO classification (1), STAS was defined as micropapillary clusters, solid nests, or single cells spread within the air spaces beyond the edge of the main tumor. Tumor cells that spread through the mucus were distinguished from STAS and excluded from our study.

Digital Imaging and Communications in Medicine (DICOM) images were downloaded from the picture archiving and communication system (PACS). An open-source medical image processing and navigation software (3D Slicer, version 4.8.0; National Institutes of Health; http://www.slicer.org) were used for pixel-level labeling of pulmonary nodules by one radiologist with 10 years of experience in lung cancer diagnosis; then, the labeling results were confirmed by another radiologist with 20 years of CT diagnostic experience. The structures of blood vessels, bronchi, and pleura outside the nodules were excluded as much as possible, and the three-dimensional VOIs of the tumors were constructed.

A total of 216 cases were manually labeled and reviewed, and the volume VOI of the lesion and all the CT data were exported in NII format for subsequent analysis. Each case was accompanied by a STAS label without any additional information, such as sex, age, location, morphological features. We randomly divided 80% of each of the two groups as training data sets (n = 172, 46 in STAS+ group, 126 in STAS− group) and the rest as verification data sets (n = 44, 10 in STAS+ group, 34 in STAS− group). The training set was used to train the designed group model, and the verification set was used to evaluate the accuracy of the model.

After the intranodular mask was annotated, a program was written to expand to the surrounding area within the lung tissue to capture the perinodular region up to a maximum distance of 20 mm. The modeling steps and radiomic analysis process are as follows and shown in Figure 1 :

The PyRadiomics software package ( http://www.radiomics.io/pyradiomics.html ) was used to extract imaging features from each VOI, including first-order features, morphological features, gray co-occurrence matrix (GLCM), gray run matrix (GLRLM), gray area size matrix (GLSZM), gray co-occurrence matrix (GLDM) texture features, and wavelet frequency domain texture features. Then, 851 features were extracted from each VOI, and a total of 6,808 imaging features (https://pyradiomics.readthedocs.io/en/latest/features.html#) were calculated for each CT image input, and the Z-score was standardized. To eliminate the influence of sample class imbalance on training, the synthetic minority class oversampling technique (Synthetic Minority Oversampling Technique, SMOTE) was used to maintain the balance between positive samples and negative samples. Because the dimension of the feature space is very high, we compared the similarity of each set of feature pairs. The Pearson correlation coefficient (Pearson Correlation Coefficient, PCC) was used to evaluate the correlation degrees of the extracted imaging features, which could reflect the linear correlation degree of the two variables. PCC is defined as the entropy of covariance and standard deviation between two characteristic variables. The calculation formula is as follows:

Among them, X and Y are the two characteristic variables; µ _i and σ_i were respectively.

If the PCC of a set of feature pairs was greater than 0.99, one of them was removed. After the dimension reduction of the Pearson correlation coefficient, the feature dimension of the feature space was reduced, and each feature was independent of each other. Before establishing the model, we used the recursive feature elimination method (recursive feature elimination, RFE) to evaluate the prediction efficiency of the selected features and further screened out the features with predictive value.

The goal of RFE is to select features based on classifiers by recursively considering smaller feature sets. AdaBoost was used as the classifier in this study. The basic principle of the Adaboost algorithm is to combine multiple weak classifiers into a strong classifier. AdaBoost is sensitive to noise and outliers, which can also avoid overfitting. Here, we used the decision tree as the basic classifier. To determine the superparameters of the model (such as the number of features), five-fold cross-validation was used for the training data set. Superparameters were set based on the performance of the model on the validated dataset.

Nine radiomic models were established by using the eight separate models (VOI_{2 mm}, VOI_{4 mm}, VOI_{6 mm}, VOI_{8 mm}, VOI_10 mm, VOI_{20 mm}, and VOI_tumor–lung) and a total of the eight VOIs (eight-VOI model). Then the prediction efficiencies of the nine radiomic models were compared. First, the extracted features were grouped according to different categories, including first-order features and morphological features, texture features (GLCM, GLRLM, GLSZM, and GLDM), and wavelet frequency features, and each set of features was used for modeling separately. Then, all the radiomic features were used for modeling. The grid search was used to select model hyperparameters, and the best hyperparameters were reversely selected based on the AUC in the validation set. For each model, 50% discount cross-validation was used for training. The probabilities of model prediction in the training and verification sets in each iteration were recorded, and the mean value of the probability recorded in each iteration when each data point was used as the training set or verification set was calculated as the result of the training set or verification set of the model.

The effectiveness of CT-based imaging radiomics in predicting STAS was evaluated by using the area under the receiver operating characteristic curve (ROC) metric. The area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were calculated under the critical value of maximizing the Youden index. We also estimated the 95% confidence interval of 1,000 samples by bootstrapping. Significance was defined as probability values less than 0.05 (p-value < 0.05).

Statistical analysis was performed by SPSS software (version 22.0, IBM SPSS Statistics, Armonk, NY, USA). Quantitative data were expressed as mean ± standard deviation. The clinical, pathological, and CT traditional morphological features of STAS+ and STAS− groups were compared using Shapiro–Wilk test or Mann–Whitney U test. Chi-square test was used to compare the differences of classified variables.

ROC curves were drawn to compare the diagnostic efficiencies of different CT signs. MedCalc software (Version19.0.2) was used to calculate the diagnostic efficacy of classification variables for STAS, and the ROC curve of the histological model was tested by DeLong to compare whether the efficiency difference among the models was statistically significant.

The basic clinical data and CT morphological characteristics of the two groups are shown in Table 1 . Among the 216 patients with lung adenocarcinoma (male, 58.9%; mean age, 56 ± 11 years), 56 patients were confirmed to be STAS+ pathologically (male, 62.5%; mean age 55 ± 9 years), and 160 patients were STAS− (male, 56.3%, mean age, 56 ± 10 years). Among all patients enrolled, 78 patients (36.1%) underwent sublobectomy, and 138 (63.9%) underwent lobectomy or pneumonectomy. There was no significant difference in age, sex, or the proportion of heavy smokers between the two groups. The proportion of patients who underwent sublobectomy in the STAS− group was slightly higher than that in the STAS+ group (p = 0.044).

There was a significant difference in the histological subtypes of lung adenocarcinoma between the STAS+ and STAS− groups (p < 0.0001): papillary, micropapillary, and solid types were predominant in STAS+ adenocarcinoma, while lepidic and acinar subtypes were more common in the STAS− group. In addition, the incidence of pleural invasion was higher, and the incidence of EGFR mutation in STAS+ lung adenocarcinoma was lower than that in STAS− adenocarcinoma. The incidence of pleural invasion and EGFR mutation was higher in STAS+ adenocarcinoma.

No difference was observed in the average diameter of the lesions between STAS+ and STAS− tumors, but the CTR of STAS+ tumors was greater than that in STAS− tumors (0.8 vs 0.5, p < 0.0001), and the proportion of solid nodules was higher in STAS+ cases (28.6 vs 25.6%, p = 0.01). Among the traditional morphological features, STAS+ lung adenocarcinoma showed more round or oval nodules, unclear tumor–lung interface and satellite focus (p < 0.0001), while STAS− lesions showed more cystic airspaces (p < 0.020).

CTR achieved an AUC of 0.796 (a sensitivity of 0.915 and a specificity of 0.686) for predicting STAS, and the optimal critical value was 0.88. Among the traditional morphological CT features, “unclear tumor–lung interface” and “satellite sign” achieved higher diagnostic performance to predict STAS with AUCs of 0.677 and 0.606, respectively. The AUCs of “solid nodule”, “irregular shape”, and “vacuole sign” were all less than 0.6, and the accuracy was less than 60% ( Table 2 ).

There was no significant difference between the training set and verification set of the main clinical data (i.e., age, sex, size, CTR, and pathological subtypes) in the STAS+ and STAS− groups. Therefore, it is considered that the case distribution of the training set and verification set is balanced, which is suitable for establishing and verifying the histological analysis model.

The results of the diagnostic efficiencies of the nine radiomic models are shown in Table 3 . We extracted 851 features in each model and compared the AUCs in the training and testing datasets respectively. The VOI_core model achieved the best diagnostic efficiency among the eight separate radiomic models and the AUCs of the training and testing sets were 0.843 (95% CI, 0.772–0.908) and 0.835 (95% CI, 0.682–0.963), respectively, and the diagnostic accuracy in the validation set was 0.818. The performances of eight separate radiomic models were significantly lower than that of eight-VOI model. The eight-VOI model achieved the best results by using all feature modeling ( Table E1 ). All 6,808 imaging features were modeled jointly, and 20 radiomic features were incorporated in the model after feature selection ( Figure 3 ). The AUCs of the eight-VOI model were 0.907 (95%CI, 0.862–0.947) in the training set, 0.897 (95%CI, 0.784–0.985) in the testing set, and 0.909 (95%CI, 0.863–0.949) in the external validation set, and the diagnostic accuracy in the external validation set was 0.849 ( Figure 4 ).

The modeling performance of a single texture feature was the worst, and the modeling effects of any other single feature type were relatively poor. Compared with the VOI_core model, eight-VOI model was generally improved, and the wavelet features had the greatest contribution.

Figure 5 shows the results of modeling using 6,808- and 851-dimentional features extracted in eight-VOI model and VOI_core model, respectively. The best performance of VOI_core model was obtained by using all three types of feature modeling. Compared with the VOI_core model alone, the best performance of eight-VOI modeling method was 7.4 and 2.8% higher in the AUC and accuracy of the verification set, respectively. Among the 20 radiomic features of the eight-VOI model ( Table E2 ), there were five VOI_core features, two VOI_{2 mm} and VOI_{4 mm} features, three VOI_{6 mm} and VOI_{8 mm} features, and two VOI_{20 mm} features ( Figure 4 ). The perinodular features accounted for 75% of the total effective features, which suggests that the radiomic features of perinodular tissues were crucial to predict STAS+ adenocarcinoma.