Construction of a multimodal artificial intelligence model for differentiating benign and malignant anterior mediastinal tumors

Introduction

Background

The anterior mediastinum refers to the anatomical compartment located posterior to the sternum, anterior to the pericardium, and anterior to the trachea. It primarily contains the thymus, adipose tissue, and lymph nodes (1). Neoplastic lesions arising in this region, whether benign or malignant, are classified as anterior mediastinal tumors. Non-neoplastic conditions such as cysts or hyperplasia occurring in the anterior mediastinum are categorized as benign lesions in this study and are differentiated from malignant anterior mediastinal tumors. Although anterior mediastinal tumors are relatively rare among thoracic diseases, they account for approximately 50% of all mediastinal tumors (2,3). Thymomas represent the most common primary neoplasms of this region, followed by lipomas, lymphomas, hemangiomas, benign teratomas, and malignant germ cell tumors. Thymic cysts are the most frequently encountered non-neoplastic lesions in the anterior mediastinum (4). According to the World Health Organization (WHO) classification, thymomas are subclassified into types A, AB, and B1–B3. Given their invasive clinical behavior, all subtypes of thymoma are now generally considered malignant (5). Due to the diverse pathological types and variable morphology of anterior mediastinal tumors, accurately distinguishing between benign and malignant lesions remains a significant clinical challenge.

Rationale and knowledge gap

Currently, the differentiation of anterior mediastinal tumors primarily relies on imaging modalities such as chest computed tomography (CT) (6). However, diagnosis based solely on single-modality imaging has notable limitations. For instance, the low incidence of anterior mediastinal tumors increases the likelihood of subjective misinterpretation by radiologists. Furthermore, distinguishing between benign and malignant thymic lesions based on imaging alone can be challenging, often requiring integration with clinical information to improve diagnostic accuracy (7). Previous studies on imaging-based differentiation of anterior mediastinal lesions have largely focused on thymic diseases, such as distinguishing benign from early-stage malignant thymic lesions, differentiating thymoma subtypes, and characterizing imaging features of various anterior mediastinal pathologies (8-11).

Radiomics, which enables the extraction of high-throughput imaging features for tumor characterization, has demonstrated considerable utility in differentiating benign from malignant tumors across a range of oncologic settings. It not only improves diagnostic accuracy but also informs treatment planning and prognostic evaluation (12). Multimodal data fusion, increasingly applied in various scientific domains, is particularly vital in medical research. It allows for the comprehensive capture of disease-related characteristics and significantly enhances diagnostic accuracy over single-modality approaches such as laboratory or imaging data alone. Artificial intelligence (AI), particularly machine learning, has become a powerful tool in medical diagnostics. Multimodal machine learning models can markedly enhance diagnostic performance, assist clinical decision-making, and promote the development of precision medicine (13,14).

Objective

Constructing a highly efficient and accurate AI model that integrates multimodal data for the differentiation of benign and malignant anterior mediastinal tumors is essential for early diagnosis and personalized treatment planning. This study expands the diagnostic scope beyond thymic lesions to include a broader spectrum of anterior mediastinal diseases. By integrating clinical, laboratory, imaging, and pathological data, we developed and compared multiple machine learning models to identify the optimal approach. The proposed diagnostic model has the potential to be translated into clinical practice, providing clinicians and patients with clearer diagnostic guidance, reducing misdiagnosis rates, optimizing therapeutic strategies, and ultimately improving patient quality of life and survival outcomes. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1010/rc).

Methods

Patient selection and data collection

This retrospective study included patients diagnosed with anterior mediastinal tumors at the Department of Thoracic Surgery, the First Affiliated Hospital of Jinan University, between January 1, 2021, and January 1, 2025. All included cases were confirmed by histopathology following surgical intervention. Exclusion criteria were: (I) masses located in the middle or posterior mediastinum; (II) inflammatory lesions of the anterior mediastinum; (III) absence of a preoperative chest CT performed at our institution; and (IV) patients who received neoadjuvant chemotherapy or neoadjuvant radiotherapy. All data were retrieved from the hospital’s integrated information system. The following preoperative clinical variables were collected: age, sex, body mass index (BMI), history of non-thymic malignancies, smoking history, alcohol consumption, presence of hypertension, diabetes mellitus, myasthenia gravis (MG), reason for tumor detection (e.g., chest-related symptoms such as pain, dyspnea, or cough; incidental finding on routine physical examination; or incidental detection during chest CT performed for unrelated conditions), and surgical approach (needle biopsy, video-assisted thoracoscopic surgery, or open thoracotomy) (15). Imaging variables included tumor dimensions (long and short axes as reported in the preoperative chest CT scan). Pathological data included postoperative tumor cross-sectional diameter and the final histological diagnosis. All clinical and imaging data were obtained preoperatively, while pathological results were derived postoperatively. Baseline characteristics of the study cohort are summarized in Table 1.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Scientific Research Ethics Committee of the First Affiliated Hospital of Jinan University (No. KY-2025-146) and individual consent for this retrospective analysis was waived.

All patients were pathologically classified based on postoperative histological diagnosis according to established criteria (5,16,17), as shown in Table 2.

Several autoimmune diseases are known to be associated with thymomas, and the benefits observed after thymectomy in certain cases further support this association (18). An observational study identified Good syndrome (GS) and pure red cell aplasia (PRCA) as the most common hematological complications of thymic epithelial tumors, each accounting for 30% of hematologic abnormalities. Both conditions affect peripheral blood cell profiles to varying degrees (19). Therefore, this study included routine hematologic tests to explore potential correlations with anterior mediastinal tumors.

Systemic inflammatory markers are clinically significant in the preoperative evaluation of solid tumors. Elevations in C-reactive protein (CRP), leukocyte counts, and decreases in albumin levels may reflect disease progression. An elevated neutrophil-to-lymphocyte ratio (NLR) has been associated with poor prognosis in several cancers. Additionally, CRP, platelet-to-lymphocyte ratio (PLR), and NLR have shown diagnostic value in thymic epithelial tumors (20-22). Hence, these markers were incorporated into our analysis to assess their relevance to other anterior mediastinal pathologies.

Beyond routine blood and inflammatory indices, tumor markers such as alpha-fetoprotein (AFP), beta-human chorionic gonadotropin (β-HCG), and lactate dehydrogenase (LDH) also provide important diagnostic clues. For example, although non-seminomatous germ cell tumors (NSGCTs) predominantly occur in young patients, measurement of serum AFP and β-HCG levels remains essential in suspected cases at any age—particularly when clinical presentation is atypical, such as mediastinal masses in older patients. Significantly elevated levels carry important diagnostic value and can effectively narrow the differential diagnosis spectrum for anterior mediastinal tumors. However, these tests are not routinely performed in clinical practice (7). In our study, LDH—a routine biochemical marker—and AFP—a tumor-associated marker—were included to enhance laboratory data depth. Additional tumor-related biomarkers were also collected to improve data richness.

To avoid confounding due to perioperative fluctuations, all laboratory test data were collected preoperatively. A summary of these laboratory findings is presented in Table 3.

Imaging acquisition and processing

Preoperative non-contrast chest CT scans of all 104 patients were retrieved from the hospital’s integrated medical platform. The images were imported into three-dimensional (3D) Slicer software (version 5.7.0) for standardized preprocessing. A fixed window width of 350 Hounsfield units (HU) and a window level of 50 HU were applied uniformly. Tumor regions of interest (ROIs) in the anterior mediastinum were manually segmented on a slice-by-slice basis using the Segment Editor module. A representative example is shown in Figure 1.

Figure 1 Representative non-contrast chest CT images from two patients, with green regions indicating manually segmented ROIs. (A) A 70-year-old female diagnosed with type B2 thymoma. (B) An 81-year-old male diagnosed with a thymic cyst. CT, computed tomography; ROIs, regions of interest.

Radiomics feature extraction

The segmented ROI masks, along with the corresponding original CT images, were processed using the open-source PyRadiomics package in Python 3.7.0 (23). A total of 1,716 quantitative radiomic features were extracted, including: 14 shape features (3D morphology), 324 first-order statistics (intensity-based features), 432 gray-level co-occurrence matrix (GLCM) features (texture patterns), 252 gray-level dependence matrix (GLDM) features, 288 gray-level run-length matrix (GLRLM) features, 288 gray-level size zone matrix (GLSZM) features, 90 neighboring gray tone difference matrix (NGTDM) features.

Feature selection

To reduce dimensionality and select the most predictive features, the least absolute shrinkage and selection operator (LASSO) regression with L1 regularization was employed. A 10-fold cross-validation was performed to assess model performance across a range of regularization parameters (λ). The optimal λ was selected based on the minimum cross-validation error, as illustrated in the cross-validation error plot (Figure 2). The coefficient path plot (Figure 3) visualizes the evolution of feature coefficients across varying λ values, where each line represents the coefficient trajectory of an individual feature.

Figure 2 Cross-validation error plot. The left vertical dashed line (λ.min) indicates the value of λ corresponding to the minimum cross-validation error, while the right dashed line (λ.1se) represents the largest λ within one standard error of the minimum. As λ increases, the complexity of the model decreases by eliminating less informative features.

Figure 3 Coefficient path plot. Each colored line represents the trajectory of a feature’s regression coefficient across a range of λ values. The red vertical dashed line marks λ.min. With increasing λ, more coefficients shrink toward zero, indicating progressive feature selection and retention of the most predictive variables.

Model construction

Clinical and laboratory variables with P<0.2 were initially screened through univariate analysis. Subsequently, logistic regression was performed to identify independent clinical predictors with P<0.05. These selected variables were used to construct a nomogram (Figure 4), providing a visual representation of their individual and combined contributions to malignancy prediction.

Figure 4 Nomogram of clinical features. The points axis represents the score assigned to each predictor based on its value. The total points axis is the sum of all individual scores, which corresponds to the malignant probability axis, indicating the predicted likelihood of malignancy in anterior mediastinal tumors. CEA, carcinoembryonic antigen; CT, computed tomography; LDH, lactate dehydrogenase.

To further improve diagnostic performance, logistic regression and random forest algorithms were used to build multimodal predictive models by integrating both clinical and radiomic features. In total, four models were developed and compared: (I) clinical features only; (II) radiomic features only; (III) a logistic regression-based fusion model; and (IV) a random forest-based fusion model (24). Receiver operating characteristic (ROC) curves were plotted for all models, and the area under the curve (AUC) was calculated to evaluate and compare model performance (Figure 5).

Figure 5 ROC curves and AUC values of the four models. AUC, area under the curve; ROC, receiver operating characteristic; RF, random forest.

Model evaluation

Model performance was comprehensively evaluated using both diagnostic performance metrics and statistical comparisons of ROC curves. The metrics included AUC, accuracy, sensitivity, specificity, and F1-score. In addition, statistical tests were conducted to determine whether differences in performance between models were significant. The results of model comparisons are presented in Tables 4,5.

Statistical analysis

The Shapiro-Wilk test was used to assess the normality of continuous variables. For variables not conforming to a normal distribution, the Kruskal-Wallis test was applied; normally distributed variables were analyzed using Welch’s t-test. Categorical variables were evaluated using Fisher’s exact test. Statistical comparisons of ROC curves were performed using DeLong’s test. All statistical analyses and model development were conducted using R software (version 4.4.2) and Python (version 3.7.0).

Results

A total of 104 patients were included, comprising 58 cases of benign tumors and 46 cases of malignant tumors. The mean age was 52 years in the benign tumor group and 53 years in the malignant tumor group. Female patients were predominant in the benign group (30/58, 51.7%), while male patients were more common in the malignant group (26/46, 56.5%). The mean BMI was similar between the two groups (24.22 kg/m² in the benign group vs. 24.33 kg/m² in the malignant group). The median cross-sectional diameter of tumors was 3 cm in the benign group and 4.5 cm in the malignant group.

There were no statistically significant differences between the benign and malignant groups in terms of age, sex, BMI, tumor cross-sectional diameter, history of non-thymic malignancy, hypertension, diabetes mellitus, smoking or alcohol use, surgical approach, or reason for tumor discovery (P>0.05). However, measurements from chest CT reports revealed that both long- and short-axis tumor lengths significantly differed between benign and malignant tumors (P<0.001). The median long-axis length was 4.40 cm [interquartile range (IQR), 2.60–5.95 cm] in malignant tumors, compared to 2.20 cm (IQR, 1.60–3.72 cm) in benign tumors. Similarly, the median short-axis length was 2.50 cm (IQR, 1.90–3.98 cm) in the malignant group vs. 1.50 cm (IQR, 1.00–2.08 cm) in the benign group. A history of MG was also significantly associated with malignant tumors (P=0.002). Among the 11 patients with MG, only 1 patient (1.7% of the benign group) was diagnosed with a benign tumor, whereas the remaining 10 patients (21.7% of the malignant group) had malignant tumors. This finding suggests that clinicians should maintain a higher suspicion for malignancy when anterior mediastinal masses are accompanied by MG symptoms.

Thymic cysts (n=25) were the most common benign anterior mediastinal tumor, while thymomas (n=39) were the most frequently observed malignant tumor. These results are consistent with prior epidemiological studies of anterior mediastinal tumors. Among benign tumors, bronchogenic cysts (accounting for 19.23%) were the second most prevalent. The remaining pathologies in Table 2 were rare, with most types represented by only one case. This may be attributable to the intrinsic low incidence of these diseases or referral bias; for instance, patients suspected of lymphoma may be referred to hematology, while those with suspected thyroid lesions may be directed to endocrine or thyroid surgery departments.

As shown in Table 3, among the laboratory parameters included in this study, statistically significant differences (P<0.05) between benign and malignant anterior mediastinal tumors were observed for neutrophil percentage (NEU%), lymphocyte percentage (LYM%), NLR, and LDH. Specifically, LDH levels demonstrated a notable difference between the two groups: the median value was 176.00 U/L (IQR, 164.50–187.00 U/L) in the benign group and 194.50 U/L (IQR, 178.75–226.25 U/L) in the malignant group (P=0.003).

Using LASSO regression, a total of seven key radiomic features were ultimately selected. Among these, diagnostics-image-original-minimum (coefficient =−0.76) and diagnostics-image-interpolated-minimum (coefficient =−0.51) showed the strongest discriminatory power for differentiating between benign and malignant anterior mediastinal tumors. The optimal λ value (λ.min =0.031) yielded the lowest cross-validation error and was used to construct the final imaging-based model incorporating these seven features.

Among the four models, the logistic regression-based multimodal fusion model achieved the highest values across all evaluation metrics, indicating superior diagnostic performance. The difference between the imaging-only model and the logistic regression-based multimodal fusion model was statistically significant (P=0.03), suggesting that integrating clinical features with radiomic features significantly improved predictive capability.

Discussion

This study represents the first application of a machine learning-based multimodal fusion model for distinguishing between benign and malignant anterior mediastinal tumors. A total of 104 real-world cases were included, comprising nine benign tumor types (e.g., thymic cysts) and seven malignant tumor types (e.g., thymomas), thus broadening the diagnostic scope beyond the commonly studied thymic epithelial neoplasms. The incorporation of clinical, laboratory, and radiomic features allowed for a more comprehensive assessment of tumor heterogeneity.

Notably, prior studies have demonstrated that machine learning approaches, such as random forest-based multimodal models, can effectively differentiate between diagnostically challenging tumor subtypes—for instance, distinguishing follicular thyroid carcinoma from follicular adenoma (25). These results support the broader applicability of multimodal fusion strategies for tumor classification and underscore their potential role in precision diagnostics.

The manual segmentation of ROIs on CT images in this study, although methodologically robust, is highly time-consuming and requires substantial clinical or radiological expertise specific to anterior mediastinal tumors. Moreover, segmentation outcomes can vary between annotators, potentially affecting the reproducibility and consistency of radiomics features. This highlights the necessity of exploring more efficient and standardized segmentation strategies. Automated segmentation using deep learning techniques—such as 3D convolutional neural networks (CNNs)—could significantly streamline this process. Automated approaches not only reduce human subjectivity but also facilitate large-scale studies. However, effective implementation would require extensive and diverse training datasets to achieve accurate image normalization and reliable prediction performance. Notably, previous research has shown that 3D CNNs do not always outperform traditional radiomics-based machine learning models in analyzing anterior mediastinal tumors (26), suggesting that further refinement in AI architectures is needed for this clinical context.

All CT images in this study were acquired from the same institution using unenhanced CT (UECT) scans, effectively minimizing variability introduced by different scanners or imaging protocols. Prior evidence has demonstrated that radiomic features derived from UECT outperform those from contrast-enhanced CT (CECT) in predicting the nature of anterior mediastinal lesions (27). This supports the use of UECT, especially in settings where CECT is not routinely indicated, and aligns with real-world clinical practice.

In the nomogram, tumor cross-sectional diameter emerged as a strong predictor of malignancy. However, this metric is only available postoperatively, which limits its utility in preoperative decision-making. Therefore, developing preoperative models must account for such variables carefully to avoid overfitting and to maintain generalizability. The superior performance of our multimodal fusion model suggests that radiomic features can partially compensate for the absence of certain pathological data in preoperative settings.

Serum biomarkers such as LDH and carcinoembryonic antigen (CEA) also demonstrated notable predictive value, reinforcing the importance of including biochemical assessments in routine preoperative evaluations. In terms of imaging features, the CT-derived short-axis length of tumors was a particularly valuable metric for distinguishing benign from malignant lesions. However, anterior mediastinal tumors comprise a heterogeneous group of diseases, each with distinct imaging characteristics. For example, solid tumors may differ in shape, calcification, necrosis, hemorrhage, or invasion into adjacent structures, while cystic lesions may vary in locularity, fluid content, wall thickness, or calcification (11). These detailed features could enhance diagnostic specificity for different tumor subtypes.

In this study, we employed two widely used machine learning algorithms—logistic regression and random forest—for multimodal data fusion. This approach mitigates the potential bias associated with relying on a single machine learning model. Our results demonstrated that the logistic regression-based fusion model consistently outperformed all other models across multiple performance metrics, including AUC, accuracy, sensitivity, specificity, and F1-score. Interestingly, while the random forest fusion model yielded higher accuracy, sensitivity, specificity, and F1-score than the clinical-only or imaging-only models, its AUC was slightly lower than that of the clinical model (0.881 vs. 0.899). This indicates that although the random forest model may perform better at a specific classification threshold, its overall discriminatory power (as reflected by AUC) is less robust.

These findings suggest that multimodal fusion models do not universally outperform single-modality models and that model selection should be tailored to the specific clinical question. For instance, if a high diagnostic certainty is required at a particular threshold, the random forest model may be preferable. Conversely, if prioritizing laboratory and clinical variables for stratification, the clinical-only model may suffice. This highlights the importance of aligning model selection with diagnostic intent and practical constraints.

Previous research has underscored the limitations of chest CT in differentiating thymomas from benign anterior mediastinal conditions such as thymic cysts, hyperplasia, and lymphomas, resulting in a non-therapeutic thymectomy rate of up to 43.8% (28). The addition of magnetic resonance imaging (MRI) significantly improves diagnostic performance, raising sensitivity from 72.1% with CT alone to 97.1% with combined CT-MRI assessment (P<0.001), thereby reducing unnecessary surgeries (29). However, routine access to both imaging modalities is not always feasible in clinical practice. In this context, the diagnostic model developed in our study provides a valuable alternative for preoperative evaluation. Accurate preoperative identification of thymomas, especially at early stages (Masaoka-Koga stages I and II), has significant therapeutic implications. Complete surgical resection at these stages is associated with excellent long-term outcomes, including high recurrence-free survival rates (30). Therefore, implementing a reliable diagnostic model in clinical workflows could aid in optimizing patient selection for surgery and improving prognostic stratification.

However, the inclusion of rare tumor types—some with distinctly malignant characteristics—may have introduced selection bias, as these cases are often easier to differentiate. The limited sample size for these rarer tumors also restricts generalizability. Future research should focus on expanding the dataset and specifically including diagnostically ambiguous lesions to enhance model robustness. In the context of imaging research, the extraction and validation of more granular features specific to various anterior mediastinal tumor subtypes requires large-scale, standardized, and high-quality imaging datasets, which remains a significant challenge for future research. While radiomics undoubtedly enhances diagnostic efficacy, this study specifically leveraged high-dimensional complex features for CT image analysis. Traditional imaging metrics (e.g., HU) remain viable for investigation. Future work will prioritize constructing integrated diagnostic models that combine conventional strong predictors (such as mean HU and lesion size) with radiomics features to maximize diagnostic performance. Regarding the investigation of laboratory biomarkers, while we demonstrated the diagnostic significance of certain indicators, the remaining laboratory indicators did not show statistically significant differences, underscoring the challenge of differentiating tumor nature based solely on routine laboratory tests. These findings highlight the need for future research to identify and validate additional laboratory biomarkers that may improve diagnostic accuracy in the evaluation of anterior mediastinal tumors.

Conclusions

The logistic regression-based multimodal fusion model demonstrated the highest diagnostic performance, achieving an AUC of 0.94. This was superior to the random forest-based fusion model (AUC =0.881), the clinical-only model (AUC =0.899), and the imaging-only model (AUC =0.844). Given its robust predictive capacity, this model holds promise for future clinical implementation, such as assisting in preoperative decision-making for anterior mediastinal tumors. The clinical nomogram (Figure 4) revealed that tumor cross-sectional diameter, short-axis length on CT reports, history of MG, LDH, and CEA were significant predictors of malignancy. These findings suggest that these features should be prioritized by clinicians when evaluating anterior mediastinal lesions.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1010/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1010/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1010/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1010/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Scientific Research Ethics Committee of the First Affiliated Hospital of Jinan University (No. KY-2025-146) and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Carter BW, Tomiyama N, Bhora FY, et al. A modern definition of mediastinal compartments. J Thorac Oncol 2014;9:S97-101. [Crossref] [PubMed]
2. Araki T, Nishino M, Gao W, et al. Anterior Mediastinal Masses in the Framingham Heart Study: Prevalence and CT Image Characteristics. Eur J Radiol Open 2015;2:26-31. [Crossref] [PubMed]
3. Strollo DC, Rosado de Christenson ML, Jett JR. Primary mediastinal tumors. Part 1: tumors of the anterior mediastinum. Chest 1997;112:511-22. [Crossref] [PubMed]
4. Carter BW, Okumura M, Detterbeck FC, et al. Approaching the patient with an anterior mediastinal mass: a guide for radiologists. J Thorac Oncol 2014;9:S110-8. [Crossref] [PubMed]
5. Marx A, Chan JK, Coindre JM, et al. The 2015 World Health Organization Classification of Tumors of the Thymus: Continuity and Changes. J Thorac Oncol 2015;10:1383-95. [Crossref] [PubMed]
6. Benveniste MF, Rosado-de-Christenson ML, Sabloff BS, et al. Role of imaging in the diagnosis, staging, and treatment of thymoma. Radiographics 2011;31:1847-61; discussion 1861-3. [Crossref] [PubMed]
7. Carter BW, Marom EM, Detterbeck FC. Approaching the patient with an anterior mediastinal mass: a guide for clinicians. J Thorac Oncol 2014;9:S102-9. [Crossref] [PubMed]
8. Marom EM, Rosado-de-Christenson ML, Bruzzi JF, et al. Standard report terms for chest computed tomography reports of anterior mediastinal masses suspicious for thymoma. J Thorac Oncol 2011;6:S1717-23. [Crossref] [PubMed]
9. McErlean A, Huang J, Zabor EC, et al. Distinguishing benign thymic lesions from early-stage thymic malignancies on computed tomography. J Thorac Oncol 2013;8:967-73. [Crossref] [PubMed]
10. Liu L, Lu F, Pang P, et al. Can computed tomography-based radiomics potentially discriminate between anterior mediastinal cysts and type B1 and B2 thymomas? Biomed Eng Online 2020;19:89. [Crossref] [PubMed]
11. Nakazono T, Yamaguchi K, Egashira R, et al. Anterior mediastinal lesions: CT and MRI features and differential diagnosis. Jpn J Radiol 2021;39:101-17. [Crossref] [PubMed]
12. Liu Z, Wang S, Dong D, et al. The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges. Theranostics 2019;9:1303-22. [Crossref] [PubMed]
13. Kumar S, Rani S, Sharma S, et al. Multimodality Fusion Aspects of Medical Diagnosis: A Comprehensive Review. Bioengineering (Basel) 2024;11:1233. [Crossref] [PubMed]
14. Warner E, Lee J, Hsu W, et al. Multimodal Machine Learning in Image-Based and Clinical Biomedicine: Survey and Prospects. Int J Comput Vis 2024;132:3753-69. [Crossref] [PubMed]
15. Bayram D, Sekmek S, Kayaalp M, et al. Clinical features and prognostic factors in thymoma and thymic carcinoma. Indian J Thorac Cardiovasc Surg 2024;40:660-8. [Crossref] [PubMed]
16. Duwe BV, Sterman DH, Musani AI. Tumors of the mediastinum. Chest 2005;128:2893-909. [Crossref] [PubMed]
17. Gerber TS, Porubsky S. Benign lesions of the mediastinum. Histopathology 2024;84:183-95. [Crossref] [PubMed]
18. Shelly S, Agmon-Levin N, Altman A, et al. Thymoma and autoimmunity. Cell Mol Immunol 2011;8:199-202. [Crossref] [PubMed]
19. Rivoisy C, Besse B, Girard N, et al. Thymic Epithelial Tumor-Associated Cytopenia: A 10-Year Observational Study in France. J Thorac Oncol 2016;11:391-9. [Crossref] [PubMed]
20. Roxburgh CS, McMillan DC. Role of systemic inflammatory response in predicting survival in patients with primary operable cancer. Future Oncol 2010;6:149-63. [Crossref] [PubMed]
21. Templeton AJ, McNamara MG, Šeruga B, et al. Prognostic role of neutrophil-to-lymphocyte ratio in solid tumors: a systematic review and meta-analysis. J Natl Cancer Inst 2014;106:dju124. [Crossref] [PubMed]
22. Megyesfalvi E, Ghimessy A, Bauer J, et al. Diagnostic and prognostic relevance of inflammatory markers in surgically treated thymic epithelial tumors: An international multicenter study. Lung Cancer 2025;200:108111. [Crossref] [PubMed]
23. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
24. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44:837-45.
25. Lin AC, Liu Z, Lee J, et al. Generating a multimodal artificial intelligence model to differentiate benign and malignant follicular neoplasms of the thyroid: A proof-of-concept study. Surgery 2024;175:121-7. [Crossref] [PubMed]
26. Chang CC, Tang EK, Wei YF, et al. Clinical radiomics-based machine learning versus three-dimension convolutional neural network analysis for differentiation of thymic epithelial tumors from other prevascular mediastinal tumors on chest computed tomography scan. Front Oncol 2023;13:1105100. [Crossref] [PubMed]
27. Sui H, Liu L, Li X, et al. CT-based radiomics features analysis for predicting the risk of anterior mediastinal lesions. J Thorac Dis 2019;11:1809-18. [Crossref] [PubMed]
28. Ackman JB, Verzosa S, Kovach AE, et al. High rate of unnecessary thymectomy and its cause. Can computed tomography distinguish thymoma, lymphoma, thymic hyperplasia, and thymic cysts? Eur J Radiol 2015;84:524-33. [Crossref] [PubMed]
29. Fang W, Xu N, Shen Y, et al. Management of incidentally detected small anterior mediastinal nodules: Which way to go? Lung Cancer 2022;168:30-5. [Crossref] [PubMed]
30. Detterbeck FC. Evaluation and treatment of stage I and II thymoma. J Thorac Oncol 2010;5:S318-22. [Crossref] [PubMed]