The local Institutional Review Board approved this retrospective study and waived patient consent.

This retrospective study included consecutive cases of TET diagnosed between January 2016 and August 2022 at three hospitals: A-Shanghai General Hospital, B-Shanghai Songjiang Hospital, and C-Shanghai Jiangqiao Hospital. Histopathological findings in all cases were derived from surgical resection of the tumor, not by percutaneous biopsy. The inclusion criteria were as follows: (1) patients underwent thin-slice (< 1 mm) contrast-enhanced CT with standard or soft reconstruction kernel; (2) underwent tumor resection within two weeks after CT scanning and were diagnosed as TET by hematoxylin-eosin and immunohistochemistry staining; and (3) with complete clinical information and surgical records. Exclusion criteria were: (1) history of thymic tumor resection or tumor recurrence; (2) other malignancies; (3) chemotherapy or radiotherapy; and (4) poor-quality CT images due to artifacts. Figure 1 illustrates the inclusion flowchart.

Patient inclusion flowchart

The medical records of each patient were reviewed by two experienced oncologists to obtain a concordant judgment for Masaoka–Koga staging and WHO classification. A dichotomous classification was used based on clinical significance, i.e., early-stage (stages I and II) and advanced-stage (III and IV) invasiveness according to Masaoka–Koga staging, and low-risk (types A, AB, and B1) and high-risk (B2, B3, and thymic carcinoma) according to WHO classification. The training set consisted of two-thirds of the cases at Hospital A, with the remaining one-third assigned to the internal validation set. The external test set consisted of cases at Hospitals B and C.

Two radiologists with 15 years and 10 years of experience in chest imaging, blinded to clinical information, assessed the semantic features of each case. A senior radiologist with 30 years of experience assessed interobserver agreement. The semantic features included two demographic, 13 tumor descriptive, 14 peritumor descriptive, and two symptomatic features (Table 1).

Six CT systems (Revolution, HD750, and VCT, GE Healthcare, Milwaukee, USA; Somatom Force and Flash, Siemens, Erlangen, Germany; uCT, United Imaging, Shanghai, China) were used for scanning (Appendix Table 1). Following the administration of contrast medium (Omnipague 300 mg/mL, GE Healthcare) via the antecubital vein at a dose of 1.5 mL/kg and a flow rate of 3.0 mL/s, CT scanning was conducted to obtain submillimeter-thin images from the apex of the lung to below the diaphragm.

Two experienced radiologists with > 10 years of experience (Readers 1 and 2) segmented tumors and quantified the segmented volume of interest (VOI) using research software (Radiomics v1.2.6, Frontier, Syngo Via, Siemens Healthineers) [24]. Reader 1 segmented the VOI of all cases, while Reader 2 segmented 30 randomly selected cases. Interobserver agreement was calculated based on these 30 cases.

A semi-automated approach was used to segment the tumor VOI (Appendix Fig. 1). First, the tumor VOI was segmented using an automated tool and then manually adjusted on a slice-by-slice basis to clearly outline the tumor edges. Second, peeling VOIs extending 3 mm, 5 mm, and 8 mm outwards were automatically segmented based on the tumor region identified in the first step. The volume between the outer boundary and the tumor edge was considered the peritumor region. The tumor and peritumor regions were inspected and corrected, and the bone component was manually removed from the VOIs. The segmentation process took 30 min per case.

The construction of the radiomics model consisted of four steps: tumor and peritumor segmentation, feature extraction, feature selection, and model construction (Fig. 2). Feature extraction was conducted using the Pyradiomics library (v3.0, https://pyradiomics.readthedocs.io/en/latest/) and followed the Image Biomarker Standardization Initiative [25]. Appendix Fig. 2 illustrates the feature composition and feature classification. Four VOIs were extracted for each lesion, namely the tumor, and tumor extension (peritumor) at distances of 3 mm, 5 mm, and 8 mm. This resulted in a total of 1691 × 4 radiomics features per tumor.

Four steps to build radiomics models: tumor and peritumor segmentation, feature extraction, feature selection, and model building

Feature selection and model construction were performed for two classification systems, Masaoka–Koga staging and WHO classification. Radiomics features were selected in four steps. First, interobserver interclass coefficients (ICC) were calculated based on 30 randomly selected cases, and unstable and irreducible features with an ICC < 0.8 were excluded. Second, features that were significantly correlated (p < 0.05) with the true labels were screened by F-test. Third, the Boruta algorithm [26] was employed for 100 iterations to identify the most important features. Finally, Spearman’s correlation coefficient analysis with hierarchical clustering was applied to remove redundant features.

Three models were developed for each of the Masaoka–Koga staging and WHO classifications, including a radiomics feature model (named the radiomics model), a semantic feature model (semantic model), and a combined model that integrated radiomics scores and semantic features.

The radiomics model was constructed using a decision tree algorithm with 5-fold cross-validation parameter tuning via grid search method on the training set. This algorithm is highly interpretable and intuitively supports decision-making [27, 28]. It has been extensively used to classify large amounts of data in clinical studies [29,30,31]. The importance of features in the radiomics model was assessed by Gini impurity using the decision tree. The radiomics model generated a score between 0 and 1, indicating a trend towards bi-directional classification. A higher score indicates a higher likelihood of an advanced tumor stage.

The semantic model employed univariate logistic regression to select semantic features associated with Masaoka–Koga staging and WHO classification. Only features with a p-value < 0.05 in univariate analyses were included in the multivariate logistic regression model. Stepwise regression with the minimum Akaike information criterion was used as the model determination criterion.

The combined model was constructed using multivariate logistic regression, integrating the radiomics scores and the semantic features selected by the semantic models. All models were trained and built from the training set. For better visualization, nomograms of the combined model were generated. Each factor in the nomogram was assigned a quantitative score, and the total score was summed to calculate the risk of being highly invasive or malignant. A higher total score indicates a higher risk for the patient.

The numerical variables were described using the mean and standard deviation. Normally and non-normally distributed continuous variables were compared using the independent samples t-test and the Mann–Whitney U-test, respectively. The discriminative performance of the model was evaluated using the receiver operating characteristic (ROC) curve, area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The AUC of different models was compared using the Delong test [32]. The optimal cut-off for each model was determined by Youden’s index in the training set and applied to the test set. Calibration curves with Brier score and decision curve analysis were implemented to demonstrate the goodness of fit and clinical efficacy of different models.

A two-tailed p < 0.05 was considered statistically significant. The statistical analysis was conducted using software packages (SPSS v22.0, IBM; R v4.0.2, www.rstudio.com; Python v3.7, www.python.org) (Appendix Table 2).