Abstract
The subtle imaging features of thymic epithelial tumors (TETs), which comprise multiple pathological subtypes of thymoma and thymic carcinoma, are of great significance for the identification of high-risk patients. Finding the radiomics features related to the immunohistochemical markers of TETs may provide a non-invasive method for the construction of a prediction model. This retrospective study analyzed non-enhanced computed tomography (NECT) images of 307 patients with TETs from two institutions. The radiomic features were extracted, clustered, and used to develop the models with machine learning algorithms. In general, the radiomics of TET patients were profiled and clustered into three clusters, which showed differences in correlation between clinicopathological characteristics, including histological type, Masaoka stage, and immunohistochemical results. Moreover, the “original-shape-flatness” and “wavelet-LHL-first-order-Median” were the most strongly correlated with CD117 and TDT expression, and the combined model of the two demonstrated predictive efficacy for CD117/TDT expression and risk groups in training and validation cohorts. This study highlights that radiomics and biomarker-associated features can serve as a non-invasive predictive biomarker for TET patients.
Subject terms: Biomarkers, Cancer, Computational biology and bioinformatics, Oncology
Introduction
Thymic epithelial tumors (TETs) represent the most common primary tumors of the anterior mediastinum, accounting for nearly 50% of all mediastinal neoplasms in adults1,2. According to the World Health Organization (WHO) classification revised in 2015, TETs are categorized into thymomas (A, AB, B1, B2, and B3 types) and thymic carcinoma3,4,5. Currently, chest non-enhanced computed tomography (NECT) imaging serves as the main diagnostic modality for evaluating suspected TETs. NECT effectively assesses tumor localization, local invasion, and distant metastasis, guiding the selection of surgical or neoadjuvant therapeutic strategies6–8. However, NECT imaging alone is limited in its ability to distinguish between specific histologic subtypes of TETs, particularly between type B3 thymoma and thymic carcinoma9,10. Clinicians are in need of new methods for identifying and predicting high-risk thymomas or TETs.
Although several markers, including CD117, are strongly positive on thymic squamous cell carcinoma, and TDT is used to mark thymoma with immature lymphocytes, the biomarker system for prognostic prediction of TETs remains lacking11,12. Moreover, the bleeding or infection risks associated with biopsy is extremely high due to the proximity of the thymus to the pericardium. Recently, radiomics has emerged as a powerful technique capable of extracting high-dimensional, quantitative images from routine medical imaging modalities13. By capturing subtle phenotypic information at both macroscopic and microscopic scales, radiomics, combined with advanced machine learning algorithms, has shown significant promise in predicting molecular and immunological biomarkers across various cancers14,15. Moreover, deep learning-based radiomics studies have successfully predicted PD-L1 expression, immune infiltration, and genetic mutations in different tumor types, underscoring the technique’s translational potential for clinical decision-making16.
The thymus is located in the mediastinal region of the chest, which can be conveniently imaged by chest CT screening. Herein, we extracted the radiomics features of 307 TET cases from two treatment centers, and developed a CT-based radiomics classification model for the prediction of histological type and Masaoka stage in TETs. Moreover, we describe the radiomics indicators with the strongest correlation between the pathological indicators, CD117 and TDT; a prediction model based on these features shows great diagnostic potential for TETs (Fig. 1a).
Fig. 1. Study design and workflow.
a Radiomic feature extraction and correlation analysis. (I) Chest NECT images were acquired for each patient. (II) Primary tumors were segmented. (III) Radiomics features were extracted from the entire tumor volume. (IV) The high-dimensional feature set was reduced to a low-dimensional feature set, then compared with clinical data. b The TET patients were enrolled from two centers and divided into three cohorts.
Results
Clinical and pathological characteristics of the patients
The detailed characteristics of the study population are presented in Table 1. A total of 307 TET patients and 100 HCs were enrolled from two centers. The median ages of the TET patients were 55 (17–83), 56 (21–79), and 57 (23–80) years across the three cohorts, with 149 males (48.5%) and 158 females (51.5%). In terms of pathological classification, the two cohorts collectively included 140 cases of the low-risk group (A, AB, and B1 types), and 167 cases of the high-risk group (B2, B3, and C types). Regarding the Masaoka stages, the majority of patients were in stages I and II (n = 246 cases, 80.1%). Additionally, 29 patients (9.4%) exhibited myasthenia gravis (MG) prior to surgery. An immunohistochemical (IHC) pathological diagnosis, including expression of CD1a, EMA, CD117, and TDT, was also collected. TET patients from the ZCH were randomly divided into a training cohort and a validation cohort at a ratio of 7:3, and the TET patients from the JCH were applied as an expanded test cohort (Fig. 1b and Table 1).
Table 1.
Summary of patient demographics and clinical characteristics in this study
| Training ZCH cohort (cases) | Validation ZCH cohort (cases) | Test JCH cohort (cases) | ||
|---|---|---|---|---|
| Thymic epithelial tumors | ||||
| 307 cases | 143 | 65 | 99 | |
| Age (years) | ||||
| median | 55 (17–83) | 56 (21–79) | 57 (23–80) | |
| Gender | ||||
| male | 68 | 35 | 46 | |
| female | 75 | 30 | 53 | |
| WHO histology | ||||
| Low-risk | A type | 21 | 6 | 9 |
| AB type | 25 | 15 | 23 | |
| B1 type | 23 | 11 | 7 | |
| High-risk | B2 type | 26 | 17 | 29 |
| B3 type | 23 | 10 | 11 | |
| C type | 25 | 6 | 20 | |
| Masaoka stage | ||||
| I | 62 | 35 | 54 | |
| II | 53 | 19 | 23 | |
| III | 25 | 10 | 17 | |
| IV | 3 | 1 | 5 | |
| MG | ||||
| Yes | 18 | 8 | 3 | |
| No | 51 | 23 | 23 | |
| N/A | 74 | 34 | 73 | |
| IHC | ||||
| CD1α | + | 40 | 20 | 21 |
| - | 19 | 5 | 12 | |
| N/A | 84 | 40 | 66 | |
| EMA | + | 56 | 21 | 7 |
| - | 36 | 18 | 31 | |
| N/A | 51 | 26 | 61 | |
| CD117 | + | 35 | 11 | 16 |
| - | 69 | 34 | 48 | |
| N/A | 39 | 20 | 35 | |
| TDT | + | 84 | 41 | 51 |
| - | 34 | 14 | 24 | |
| N/A | 25 | 10 | 24 | |
| Healthy human thymus | ||||
| 100 cases | 73 | 27 | / | |
| Age (years) | ||||
| Median | 16 (14–21) | 17 (15–19) | ||
| Gender | ||||
| Male | 30 | 11 | ||
| Female | 43 | 16 | ||
Profiling and classifying TET patients by radiomics
A total of 852 radiomics features of thymic or mediastinal mass regions were extracted from 307 TET patients and 100 HCs, and 812 features with an ICC value greater than 0.75 were then selected for future analysis (Fig. S1a). The principal component analysis (PCA) based on these features shows that the top seven PCs account for 96.8% of the variance between the TETs and NCs (Fig. 2a and Fig. S1b, c). To comprehensively evaluate the relationship between the radiomics features and pathological characteristics in TET patients, we performed K-means clustering analysis, which revealed clusters within the patient population that exhibited similar radiomic expression patterns, with a Hopkins index of 0.863 and a Jaccard index of 0.61 (Cluster 1 = 131, Cluster 2 = 136, Cluster 3 = 40; Fig. 2b, c). We compared these clusters with the pathological characteristics and found significant correlation with the histopathological classification and Masaoka stage (P < 0.001; Fig. 2c and Table 2). Specifically, Cluster 1 was associated with low-risk pathological subtypes, and nearly all patients in this cluster were diagnosed at an early stage. In contrast, Cluster 2 was linked to high-risk pathological subtypes, with patients in this cluster representing a mix of early and late stages. Cluster 3 was associated with thymic carcinomas, and most patients were diagnosed with TET at a late stage (Fig. 2d). In addition, these clusters demonstrated significant correlation with CD117, TDT, EMA, and CD1a expression (P = 0.041), particularly CD117 (P = 0.008) and TDT (P = 0.014) (Fig. 2e and Table 2).
Fig. 2. Profiling and clustering the radiomics of TET patients.
a Radiomic feature-based PCA score plots between TET patients and HCs. b Clustering projection. The feature-based clustering result of each sample is depicted as a point. Each color represents a cluster group (k = 3, Hopkins = 0.863). c Radiomics-based heat map. Correspondence of radiomic feature groups with the clustered expression patterns is at the top of the figure. d Clinical patient parameters showing significant association of the radiomic expression patterns with Masaoka stage (P < 0.001, w2 test), and histological type (P < 0.001, w2 test). e Bar chart showing the IHC results for each marker among the three clusters.
Table 2.
Statistical analysis of three TET clusters based on the radiomics and clinicopathological characteristics
| Cluster 1 | Cluster 2 | Cluster 3 |
P value (Chi-square test) |
||
|---|---|---|---|---|---|
| N (cases) | 131 | 136 | 40 | ||
| Age (years) | |||||
| Median | 55 (23–82) | 54 (17–76) | 63 (29–83) | ||
| Gender | |||||
| Male | 56 | 76 | 21 | 0.191 | |
| Female | 72 | 63 | 19 | ||
| WHO histology | |||||
| Low-risk | A type | 19 | 17 | 0 | <0.001 |
| AB type | 39 | 22 | 2 | ||
| B1 type | 22 | 18 | 1 | ||
| High-risk | B2 type | 28 | 42 | 2 | |
| B3 type | 13 | 23 | 8 | ||
| C type | 10 | 14 | 27 | ||
| Masaoka stage | |||||
| I | 83 | 60 | 8 | <0.001 | |
| II | 39 | 47 | 9 | ||
| III | 8 | 26 | 18 | ||
| IV | 1 | 3 | 5 | ||
| MG | |||||
| Yes | 12 | 16 | 1 | 0.329 | |
| No | 46 | 41 | 11 | ||
| N/A | 73 | 79 | 28 | ||
| IHC | |||||
| CD1α | + | 35 | 38 | 8 | 0.041 |
| - | 17 | 10 | 9 | ||
| N/A | 79 | 88 | 23 | ||
| EMA | + | 31 | 39 | 15 | 0.047 |
| - | 44 | 35 | 6 | ||
| N/A | 56 | 62 | 19 | ||
| CD117 | + | 17 | 33 | 12 | 0.008 |
| - | 68 | 72 | 11 | ||
| N/A | 46 | 31 | 17 | ||
| TDT | + | 87 | 77 | 12 | 0.014 |
| - | 21 | 38 | 13 | ||
| N/A | 23 | 21 | 15 | ||
Correlation analysis of radiomics and immunohistochemical biomarkers and establishment of the FM model
To obtain the detailed features associated with CD117 and TDT expression, we grouped TET patients with IHC detection of CD117 (62 CD117 + , 151 CD117−) and TDT (176 TDT + , 72 TDT−) (Table 1). We next used LASSO regression to identify four significant features associated with CD117 expression (Fig. 3a, b). The importance was ranked according to their coefficients, which showed that the highest correlation feature was “original-shape-flatness” (Fig. 3c). While six features exhibited high correlation with TDT expression, the “wavelet-LHL-first-order-Median” feature found to be the most highly correlated with TDT expression (Fig. 3d–f).
Fig. 3. Correlation analysis of radiomics features and biomarkers of CD117 and TDT expression.
a Cross-validation curves of the LASSO regression model based on CD117 expression. b Coefficient curves for the radiomic features based on CD117 expression. c Coefficients in the LASSO model based on CD117 expression. d Cross-validation curves of the LASSO regression model based on TDT expression. e Coefficient curves for the radiomic features based on TDT expression. f Coefficients in the LASSO model based on TDT expression. g–i Box plots showing the value of the related features between the CD117-TDT+ group and the CD117 + TDT− group. j CT image, HE, TDT, and CD117 results from two of the TET patients. k ROC curves of the FM model in the three cohorts.
We validated the value of the two above features in the groups of typical IHC results of thymic carcinoma (CD117 + , TDT−) and thymoma (CD117−, TDT + ), the only 7 cases with atypical IHC results (CD117 + TDT+ and CD117− TDT−) were excluded from the analysis to avoid interference (Table S2). The values of “original shape-flatness value” (F value) and “wavelet-LHL-first-order-Median” (M value) of the CD117 + TDT− group were significantly higher than those the CD117− TDT+ group (P < 0.001, respectively; Fig. 3g, h, j). Moreover, the FM model (Score = Log2 (F × (M + 5)), which combined the above two most IHC biomarkers-associated features, was also significantly different between the two groups (P < 0.001; Fig. 3i).
We next investigated whether the FM model can serve as a predictor of CD117/TDT expression in TET patients. To this end, we divided TET patients with both CD117 and TDT IHC analysis into an internal training cohort, an internal validation cohort, and an external test cohort (Table S2). The machine learning algorithms (LG, DF, RF, SVM, and Ada) were used to evaluate the AUC values of the FM model in the different cohorts; all showed relatively stable prediction efficiency (Fig. S2 and Table S3). The optimal results showed that the AUC values of the FM model in the internal training cohort, the internal validation cohort, and the external test cohort were 0.882, 0.853, and 0.844, respectively (RF algorithm; Fig. 3k). Additionally, the results of calibration metrics and DCA suggest that there is a favorable net benefit in the target population and a reliable probability calibration effect (Fig. S7).
Evaluation and comparison of the prediction performance of the radiomics-based models in TET patients
To expand the clinical applications of the FM model in predicting the clinicopathological features of TET patients, we next evaluated the predictive performance of the FM model in the risk group, Masaoka stage, MG, gender, and age status. The AUC values of the FM model in predicting risk were 0.788, 0.779, and 0.765 in the training cohort, validation cohort, and test cohort, respectively (Fig. 4a). The AUC values of the FM model for predicting Masaoka stage were 0.762, 0.709, and 0.645 in the training cohort, validation cohort, and test cohort, respectively (Fig. 4d). In addition, the FM model showed low predictive value for predicting MG, gender, and age (AUC < 0.61, respectively; Fig. S3).
Fig. 4. ROC curves of the different models in the TET cohorts.
a ROC curves for predicting the risk group based on the FM model. b ROC curves for predicting the risk group based on the optimal CD117/TDT-associated signature model. c ROC curves for predicting the risk group based on the optimal risk-associated signature model. d ROC curves for predicting the Masaoka stage based on the FM model. e ROC curves for predicting the Masaoka stage based on the optimal CD117/TDT-associated signature model. f ROC curves for predicting the Masaoka stage based on the optimal risk-associated signature model.
We also constructed prediction models using LASSO regression to screen out the most relevant signatures with CD177/TDT expression (Fig. S4), risk (Fig. S5a–c), and Masaoka stage (Fig. S6a–c). In comparison with the optimal CD117/TDT-associated signature model, the FM model showed better prediction efficiency in predicting risk (Fig. 4a, b) and Masaoka stage (Fig. 4d, e). Among the three types of models, the optimal prediction model constructed according to risk and Masaoka stage showed better AUC values (Fig. 4c, f); but the signature combination of the optimal models of risk and Masaoka stage was completely different, and there was no predictive value between the two (Figs. S5a, S6h, and S8).
Discussion
In the present study, we demonstrated that radiomics can serve as a non-invasive predictive biomarker for TET patients. Clustering analysis based on radiomics features revealed three distinct subtypes of TETs. Cluster 1 predominantly consisted of A and AB-type thymomas, characterized by spindle-shaped cells. Cluster 2 mainly comprised B1, B2, and B3-type thymomas, with lymphocytes being the predominant cell type. Cluster 3 is primarily composed of thymic carcinomas, characterized by a higher proportion of atypical cells. Recent single-cell profiling studies have shown that different subtypes of TETs have different cells of origin. Taken together, our results suggest that the heterogeneity of the cellular composition of TET tissue structure may affect CT and/or MRI imaging results.
To comprehend the biologic rationale behind radiomics as an imaging biomarker, we explored the relationship between the molecular characteristics of the radiomics subtypes and the expression of specific immune molecules. We observed a significant correlation between the three imaging subtypes and the expression of CD117 and TDT. Subsequently, LASSO regression was applied to further analyze the data, revealing a strong association between “first-shape-flatness” and CD117 expression. In radiomics, “flatness” refers to the balance of amplitudes among the frequency components in a spectrum. Higher flatness values indicate a flatter spectrum with balanced amplitudes, while lower values reflect prominent peaks17. In our study, the flatness values in the CD117 cohort were significantly higher than those in the TDT cohort. Previous studies have consistently reported that CD117 is frequently overexpressed in thymic carcinoma but absent in thymoma18, suggesting that thymic carcinoma may exhibit more uniform signal intensity. This difference may be related to the fact that the cells in thymic carcinoma are more closely packed and have a greater ability to proliferate, whereas thymoma is often infiltrated by immature lymph nodes and forms different grid areas19. On the other hand, the “wavelet-LHL-first-order-Median” was found to be most correlated with the TDT expression. Wavelet-LHL-first-order-Median refers to the grayscale median based on the wavelet transform20. TDT is mostly expressed in thymomas with immature lymphocytes19. It has been reported that undifferentiated lymphocytes have a higher nuclear cytoplasmic ratio than other cells21, and lymphocytes are highly accumulated in thymoma, which may be related to the lower gray-level intensity in the CT imaging. Additionally, regardless of the MRMR + LASSO or T test + LASSO method used, the features most relevant to CD117 and TDT expression that were selected were original-shape-flatness and wavelet-LHL-first-order-Median (Table S4).
Recently, Gao et al.22, Xiao et al.23, and Mayoral et al.24 reported that CT radiomics can distinguish low-risk from high-risk thymomas, showing a positive predictive performance. Araujo-Filho JAB et al.8 and Christian Blüthgen et al.25 demonstrated that a CT-based radiomics model could predict TNM staging with AUCs of 0.801 and 0.708 in the testing sets. Wenzhang He et al.26 combined conventional imaging and radiomics features, achieving respective AUCs of 0.715 and 0.810 in internal and external validation for distinguishing TET and lymphomas in the anterior mediastinum. Herein, we established a novel model based on the ratio of the two CD117 and TDT-associated radiomic features. The FM model exhibited a robust ability to predict CD117/TDT expression (AUC value up to 0.825 in the test cohort) and was cross-validated in five different machine learning models in the three cohorts. Furthermore, among the clinical factors closely related to TETs, the FM model demonstrates potential predictive application value in predicting Masaoka staging. However, its predictive performance for MG, gender, and age is not satisfactory. This suggests that radiomics associated with IHC results may have better tumor phenotype-related predictive capabilities, and its predictive ability is not universally generalized. Focusing on radiomics models associated with the core biological characteristics of tumors may have greater clinical application value in assisting with diagnosis and prognosis prediction.
Although various strategies have been employed to control overfitting in this study, performance degradation was still observed in the external validation cohort. In the future, standardized and domain-adaptive studies on multi-center data need to be established to reduce bias and enhance the robustness and applicability of the FM model. In conclusion, our study developed a novel model for identifying imaging biomarkers associated with molecular expression in TET patients. Mining tumor-specific, expression-related pair features in radiomics may be of particular importance for designing novel and clinically beneficial tests to predict risk in patients with TETs.
Methods
Patient cohort and pathologic evaluation
We retrospectively screened patients with TET who underwent surgical resection and healthy controls (HCs) at the Departments of Thoracic Surgery at Zhejiang Cancer Hospital (ZCH) and Jiangxi Cancer Hospital (JCH), between June 2019 and January 2024. The inclusion criteria were as follows: (1) pathologically confirmed TET after sternotomy, thoracoscopic surgery, or biopsy; (2) NECT scan before any treatment or operation; (3) clinical and IHC data available in medical records. The exclusion criteria were as follows: 1) unavailability of preoperative CT images or images with inadequate quality for radiomic analysis; 2) TETs that were not classified into one of the six main WHO histological subtypes; 3) incomplete clinical information. This study was approved by the Institutional Review Board of Zhejiang Cancer Hospital (IRB-2024-78) and the Institutional Review Board of Jiangxi Cancer Hospital (2023ky237) and in accordance with the Declaration of Helsinki. The requirement for informed consent was waived, due to the retrospective nature of the study. Patient confidentiality was strictly maintained, and all clinical data were anonymized to ensure patient privacy and compliance with ethical standards.
Surgical specimens and puncture biopsies from all patients were subjected to histopathological analysis and staging confirmation. Histologic subtypes were classified into low-risk (A, AB, B1) and high-risk (B2, B3, C) groups according to the WHO classification17,18. Stage classification was divided into early (I/II) and advanced (III/IV) stages based on the Masaoka staging system27.
CT data acquisition
CT examinations were performed using three different instruments: Scanner A (GE Optima CT680, n = 125), Scanner B (Siemens Definition Flash, n = 183), and Scanner C (Philips Ingenuity CT, n = 99), the detail information of the equipment parameters was described in Table S1. The slice thickness for all scans was 5 mm. During the CT examination, the patient was positioned supine on the examination table with arms raised above the head. After taking a deep breath, the patient held their breath for the duration of the scan. Axial images were reconstructed from the source data, with coronal and sagittal reconstructions available for all patients. After acquisition of the CT images, the analysis workflow consisted of three steps: tumor segmentation, feature extraction, analysis, and comparison (Fig.1a).
Lesion segmentation and feature extraction
Images were imported into 3D Slicer (v5.0.3) for lesion segmentation and extraction of radiomic features. Following initial delineation by two observers, the inter-observer reliability of radiomic features was assessed using the ICC (2,1) model. All segmentations were then verified by a senior radiologist, with a third expert resolving any disputes. The images were resampled to a pixel spacing of 1 × 1 × 1 (millimeter) to offset the interference caused by inconsistent spatial resolution. Subsequently, wavelet filters with various parameters were employed to preprocess the original NECT images, emphasizing the texture at different scales.
For each lesion, 852 radiomic features were computed. These features can be classified into eight categories: first-order statistics (18 features), shape-based features (14 features), grey-level co-occurrence matrix (GLCM; 24 features), grey-level run length matrix (GRLRM; 16 features), neighboring grey tone difference matrix (NGTDM; 5 features), grey-level difference matrix (GLDM; 14 features), grey-level size zone matrix (GLSZM; 16 features), and 745 wavelet features.
Unsupervised machine learning
Unsupervised k-means clustering was conducted to identify homogeneous imaging-based patient clusters without a priori assumptions in the derivation cohort (n = 307). This clustering was performed in R (v4.4.2) using the Cluster package. A range of 1 to 10 was explored to select the optimum number of clusters, with the point of inflection of inertia, or within-cluster sum of squares, identified as the optimal value. K-means clustering was performed with 10 iterations with random initialization to ensure the robustness of clustering.
Feature selection and model building
To select the most informative features, a multistep approach was employed. Radiomic features from the ZCH training cohort were pre-screened using the Maximum Relevance and Minimum Redundancy (mRMR) algorithm, yielding the top 100 discriminative features, where K = 100 was the optimal value identified by 10-fold cross-validation (search range: 50–150). The Mann-Whitney U test was used to compare features between groups, retaining those with p-values less than 0.05. The least absolute shrinkage and selection operator (LASSO) method was subsequently used to further refine the feature set based on its coefficients. This feature selection process resulted in fewer than 20 radiomics features.
A total of five classifiers were tested in this study: Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Adaptive Boosting (Adaboost)28–30. Predictive models were built using the selected radiomics signature and generated through stratified cross-validation31,32, evaluating 10 iterations of each model with a dataset partition of 70% training and 30% validation. The standard SMOTE algorithm (K = 5 and dup_size=0) was applied to achieve complete class balance when the imbalance ratio of training cohorts exceeded 3, with strict prevention of data leakage under a 5-fold cross-validation framework33–37. These models were then tested in the external validation datasets. The predictive performance of each model was quantified using the receiver operating characteristic (ROC) curve, calculating the area under the ROC curve (AUC), decision curve analysis (DCA) and the calibration curve.
Statistical analysis
Z-score normalization was used to transform data with different scales or units into a standardized scale. Continuous variables were analyzed using the Wilcoxon rank sum test, and the test was used depending on the normality of the distribution. The chi-squared test was used to identify significant differences in categorical variables. Statistical analyses were performed with R (v4.4.2) software. All statistical tests were two-sided, and P < 0.05 was considered indicative of a statistically significant difference.
Supplementary information
Acknowledgements
This work was supported by the National Natural Science Foundation of China (82172567), the Ministry and the province of Zhejiang Medical and Health Science and Technology (No. WKJ-ZJ-26030), the Key R&D Plan of Jiangxi Province (2021BBG71006), and the Key Project of Science and Technology Innovation of Health Commission of Jiangxi Province (2023ZD005 and 2024ZD008).
Author contributions
YT.Z., YZ.G., JY.L., and HT.J. collected the data and wrote the manuscript. A.Z. and WM.M. initiated the study and revised and finalized the manuscript. YT.Z., YZ.G., YY.H., WH.S., Y.X., JH.W., C.Y., CC.W., QL.L., and BJ.F. participated in the data processing and statistical analysis of the manuscript. All authors read and approved the final manuscript.
Data availability
The image datasets generated and analyzed during the current study are not publicly available due to patient privacy concerns but are available from the corresponding author An Zhao (zhaoan@zjcc.org.cn) upon reasonable request.
Code availability
The bioinformatics analyses were performed using open-source software, including 3D Slicer (v5.0.3)38, GraphPad Prism (v10.4.1)39, and R (v4.4.2) with the following packages: cluster, pROC, glmnet, rmda, ROSE, ggplot2, mRMRe, and writexl. All aforementioned R packages are publicly available via the Comprehensive R Archive Network (CRAN), with specific access links provided on their respective CRAN pages (e.g., cluster package: https://CRAN.R-project.org/package=cluster).
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
These authors contributed equally: Yutian Zhang, Yangzhong Guo, Junyu Li, Haitao Jiang.
Contributor Information
An Zhao, Email: zhaoan@zjcc.org.cn.
Weimin Mao, Email: maowm@zjcc.org.cn.
Supplementary information
The online version contains supplementary material available at 10.1038/s41698-026-01286-4.
References
- 1.Wang J. et al. Insights into the characteristics of cancer incidence in China. Sci. China Life Sci.68, 1512–1515 (2025). [DOI] [PubMed]
- 2.Bray, F. et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J. Clin.74, 229–263 (2024). [DOI] [PubMed] [Google Scholar]
- 3.Carbone, M. et al. Advances in the Basic Sciences in Thoracic Oncology in the Last 20 Years and Their Translational Impact. J Thorac Oncol. 21(1):41-76 10.1016/j.jtho.2025.11.002 (2026). [DOI] [PMC free article] [PubMed]
- 4.Den Bakker, M. A., Roden, A. C., Marx, A. & Marino, M. Histologic classification of thymoma: a practical guide for routine cases. J. Thorac. Oncol.9, S125–S130 (2014). [DOI] [PubMed] [Google Scholar]
- 5.Engels, E. A. Epidemiology of thymoma and associated malignancies. J. Thorac. Oncol.5, S260–S265 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Benveniste, M. F. K. et al. Role of imaging in the diagnosis, staging, and treatment of thymoma. RadioGraphics31, 1847–1861 (2011). [DOI] [PubMed] [Google Scholar]
- 7.Gao, J. et al. Clinical significance of positron emission tomography-computed tomography in the classification of thymic tumors. Interdiscip. Cardiovasc. Thorac. Surg.40, ivaf065 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Araujo-Filho, J. A. B. et al. CT radiomic features for predicting resectability and TNM staging in thymic epithelial tumors. Ann. Thorac. Surg.113, 957–965 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Nakagawa, K. et al. Thymoma: a clinicopathologic study based on the new World Health Organization classification. J. Thorac. Cardiovasc. Surg.126, 1134–1140 (2003). [DOI] [PubMed] [Google Scholar]
- 10.Mayerhoefer, M. E. et al. Introduction to radiomics. J. Nucl. Med.61, 488–495 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Zhang, Y. et al. Thymoma and thymic carcinoma: surgical resection and multidisciplinary treatment. Cancers15, 1953 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Scorsetti, M. et al. Thymoma and thymic carcinomas. Crit. Rev. Oncol. Hematol.99, 332–350 (2016). [DOI] [PubMed] [Google Scholar]
- 13.Horvat, N., Papanikolaou, N. & Koh, D. M. Radiomics beyond the hype: a critical evaluation toward oncologic clinical use. Radiol. Artif. Intell.6, e230437 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Guiot, J. et al. A review in radiomics: making personalized medicine a reality via routine imaging. Med. Res. Rev.42, 426–440 (2022). [DOI] [PubMed] [Google Scholar]
- 15.Shur, J. D. et al. Radiomics in oncology: a practical guide. RadioGraphics41, 1717–1732 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Bera, K., Braman, N., Gupta, A., Velcheti, V. & Madabhushi, A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat. Rev. Clin. Oncol.19, 132–146 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Abbasian Ardakani, A., Bureau, N. J., Ciaccio, E. J. & Acharya, U. R. Interpretation of radiomics features–a pictorial review. Comput. Methods Prog. Biomed.215, 106609 (2022). [DOI] [PubMed] [Google Scholar]
- 18.Pan, C., Chen, P. C. & Chiang, H. KIT (CD117) is frequently overexpressed in thymic carcinomas but is absent in thymomas. J. Pathol.202, 375–381 (2004). [DOI] [PubMed] [Google Scholar]
- 19.Silverstone, A., Cantor, H., Goldstein, G. & Baltimore, D. Terminal deoxynucleotidyl transferase is found in prothymocytes. J. Exp. Med.144, 543–548 (1976). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Zhang, W., Guo, Y. & Jin, Q. Radiomics and its feature selection: a review. Symmetry15, 1834 (2023). [Google Scholar]
- 21.Sebastian J. A., Moore M. J., Berndl E. S. L. & Kolios M. C. An image-based flow cytometric approach to the assessment of the nucleus-to-cytoplasm ratio. L’Imperio V., ed. PLoS ONE. 16, e0253439 (2021). [DOI] [PMC free article] [PubMed]
- 22.Gao, C. et al. Differentiating low-risk thymomas from high-risk thymomas: preoperative radiomics nomogram based on contrast enhanced CT to minimize unnecessary invasive thoracotomy. BMC Med. Imag.24, 197 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Xiao, G. et al. MR imaging of thymomas: a combined radiomics nomogram to predict histologic subtypes. Eur. Radio.31, 447–457 (2021). [DOI] [PubMed] [Google Scholar]
- 24.Mayoral, M. et al. Conventional and radiomic features to predict pathology in the preoperative assessment of anterior mediastinal masses. Lung Cancer178, 206–212 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Blüthgen C. et al. Computed tomography radiomics for the prediction of thymic epithelial tumor histology, TNM stage and myasthenia gravis. PLoS ONE16, e0261401 (2021). [DOI] [PMC free article] [PubMed]
- 26.He, W. et al. Computed tomography-based radiomics for differentiation of thymic epithelial tumors and lymphomas in anterior mediastinum. Front. Oncol.12, 869982 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Markowiak, T., Hofmann, H. S. & Ried, M. Classification and staging of thymoma. J. Thorac. Dis.12, 7607–7612 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Greener, J. G., Kandathil, S. M., Moffat, L. & Jones, D. T. A guide to machine learning for biologists. Nat. Rev. Mol. Cell Biol.23, 40–55 (2022). [DOI] [PubMed] [Google Scholar]
- 29.Erickson, B. J., Korfiatis, P., Akkus, Z. & Kline, T. L. Machine learning for medical imaging. RadioGraphics37, 505–515 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Wang, Y., Carter, B. Z., Li, Z. & Huang, X. Application of machine learning methods in clinical trials for precision medicine. JAMIA Open5, ooab107 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Baron, G. & Stańczyk, U. Standard vs. non-standard cross-validation: evaluation of performance in a space with structured distribution of datapoints. Procedia Comput. Sci.192, 1245–1254 (2021). [Google Scholar]
- 32.Wainer, J. & Cawley, G. Nested cross-validation when selecting classifiers is overzealous for most practical applications. Expert Syst. Appl.182, 115222 (2021). [Google Scholar]
- 33.Muntasir Nishat M. et al. A comprehensive investigation of the performances of different machine learning classifiers with SMOTE-ENN oversampling technique and hyperparameter optimization for imbalanced heart failure dataset. Sci. Program2022, 1–17 (2022).
- 34.Seo, J. H. & Kim, Y. H. Machine-learning approach to optimize SMOTE ratio in class imbalance dataset for intrusion detection. Comput. Intell. Neurosci.2018, 1–11 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Mahani A. & Riad Baba Ali A. Classification problem in imbalanced datasets. In Recent Trends in Computational Intelligence (eds Sadollah, A. & Sinha, T. S.) Ch. 11 (IntechOpen, 2020). 10.5772/intechopen.89603 (2020).
- 36.Noorhalim N., Ali A., Shamsuddin S. M. Handling imbalanced ratio for class imbalance problem using SMOTE. In: Kor L. K., Ahmad A. R., Idrus Z., Mansor K. A., eds. Proceedings of the Third International Conference on Computing, Mathematics and Statistics (iCMS2017). Springer Singapore, 19–30 (2019).
- 37.Chawla, N. V., Bowyer, K. W., Hall, L. O. & Kegelmeyer, W. P. SMOTE: synthetic minority over-sampling technique. J. Artif. Intell. Res.16, 321–357 (2002). [Google Scholar]
- 38.Huie, J. M., Summers, A. P. & Kawano, S. M. SegmentGeometry: a tool for measuring second moment of area in 3D slicer. Integr. Org. Biol.4, obac009 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Mavrevski, R. Approaches to modeling of biological experimental data with GraphPad Prism software. Procedia Comput. Sci.192, 4312–4320 (2021).
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The image datasets generated and analyzed during the current study are not publicly available due to patient privacy concerns but are available from the corresponding author An Zhao (zhaoan@zjcc.org.cn) upon reasonable request.
The bioinformatics analyses were performed using open-source software, including 3D Slicer (v5.0.3)38, GraphPad Prism (v10.4.1)39, and R (v4.4.2) with the following packages: cluster, pROC, glmnet, rmda, ROSE, ggplot2, mRMRe, and writexl. All aforementioned R packages are publicly available via the Comprehensive R Archive Network (CRAN), with specific access links provided on their respective CRAN pages (e.g., cluster package: https://CRAN.R-project.org/package=cluster).