Artificial intelligence for diagnosis and prognosis of thymic epithelial tumors: a systematic review

Introduction

Thymic epithelial tumors (TETs)

TETs are rare neoplasms of the thymus gland, yet represent the most common primary tumors of the prevascular mediastinum, with an incidence of 0.23–0.3 per 100,000 persons annually (1,2). The main subtypes include thymomas, thymic carcinomas (TCs), and thymic neuroendocrine tumors (NETs) (3). Diagnosis of TETs is rapidly evolving. The World Health Organization (WHO) established a TET classification system in 1999 (4), with subsequent refinements aimed at reducing interobserver variability (5-8).

Diagnostic workflow

TETs may present with clinical symptoms such as cough, chest pain, and dyspnea, or with paraneoplastic syndromes (e.g., myasthenia gravis), but are frequently asymptomatic and discovered incidentally on chest imaging (9). When a mediastinal mass is found, radiologic imaging distinguishes malignant lesions from benign conditions like thymic hyperplasia (10). Contrast-enhanced computed tomography (CT) scans provide details on tumor characteristics and local spread; magnetic resonance imaging (MRI) can further aid in differentiating benign from malignant thymic lesions, particularly for cystic lesions (11). Positron emission tomography with fluorodeoxyglucose (FDG-PET) assesses metabolic activity and aids evaluation of tumor aggressiveness (11,12).

Despite being highly informative, radiological tests often cannot provide a definitive diagnosis because imaging features frequently overlap between tumor types and non-thymic lesions. Therefore, histopathological analysis through needle biopsy, mediastinoscopy, or surgical resection is essential for diagnosis, classification, and biological assessment (13).

Histological subtyping

Following tissue sampling, subtyping per WHO classification is critical. Thymomas are categorized as A, AB, B1, B2, or B3 (Figure 1), based on epithelial cell morphology and lymphocyte proportions (13). TC is a separate, more aggressive subtype, spanning well-differentiated to poorly-differentiated categories (1,14-17).

Figure 1 Different subsets of TETs, classified according to the WHO classification (H&E, magnification 10×). TET, thymic epithelial tumor; WHO, World Health Organization.

Subtyping plays a vital role in guiding treatment decisions (10). Type A, AB and B1 thymomas are generally indolent, whereas types B2, B3 and TC are associated with more aggressive clinical courses and higher rates of invasion and recurrence (18). Histology also guides staging, as more aggressive histologies correlate with higher Masaoka-Koga or traditional tumor-nodes-metastasis (TNM) stages (19,20). Furthermore, it guides decisions regarding adjuvant therapies and postoperative surveillance. Therefore, expert pathological assessment remains essential (10).

However, subtyping remains subjective, with significant interobserver variability, notably for subtypes B1 and B2 (4,21). These challenges underscore the potential value of AI in improving diagnostic consistency.

Preoperative diagnosis

For suspected TET cases, a detailed and accurate preoperative diagnostic evaluation is essential for selecting the optimal treatment. This evaluation aims to distinguish benign (non-cancerous) growths from malignant neoplasms, which are cancerous growths that can invade nearby tissues and spread. It is also important to determine whether the tumor originates from the thymus gland or another mediastinal structure. Furthermore, classifying patients by tumor type and risk level, such as low-risk (A, AB, B1) or high-risk (B2, B3, TC), is critical for accurate staging and guiding subsequent treatment. This is especially important because thymomas and TCs are distinct tumors that require different management approaches (10,22).

Artificial intelligence (AI) in medicine

AI is now integral to healthcare, excelling at analyzing large datasets for pattern recognition, prediction, and decision support, and driving innovation in diagnostics and patient management across multiple domains.

AI methodologies

Machine learning (ML)

ML utilizes large datasets to uncover complex relationships and classify or predict outcomes without explicit rules. Techniques include support vector machines (SVM), random forests (RF), and gradient boosting (GB).

Deep learning (DL)

A subset of ML based on artificial neural networks (ANNs), particularly convolutional neural networks (CNNs), which excel at image recognition and classification, making them especially useful in radiology and pathology.

Types of data

In general, AI models in medicine utilize:

• Clinical data: a detailed list of information about patients, including details such as age, medical history, laboratory tests, and treatment outcomes.
• Pathology data: it includes whole slide images of tissue samples, details of the staining procedure, and molecular analysis results.
• Radiomics data: focuses on extracting quantitative features like texture, shape, and intensity from medical images.
• Omics data: provides molecular-level insights into diseases, aiding biomarker discovery and the development of targeted therapies.

Types of clinical goal

• Diagnostic support: AI models improve neoplasm detection and characterization through modalities such as CT and MRI. Integrating clinical, radiomics, and pathology data enhances diagnostic precision and informs personalized treatment.
• Outcome prediction: models predict disease progression, response, recurrence risk, and survival, supporting individualized therapeutic strategies.

AI for TET diagnosis and prognosis

Given the diagnostic challenges of TETs, including radiologic ambiguity, heterogeneity, and limited risk stratification, AI is well-positioned to:

• Enhance preoperative risk and stage assessment using radiologic data;
• Detect tumor invasiveness and metastatic potential;
• Predict prognosis, including recurrence, survival, and treatment response;
• Reduce variability in histopathologic classification;
• Integrate multimodal data for personalized management.

AI-enhanced diagnostic and prognostic pathways support earlier intervention, more precise treatment selection, and better overall patient outcomes and quality of life. Figure 2 illustrates the overall framework used for analyzing TETs, incorporating clinical, radiological, pathological, omics, and integrated multi-omics data for diagnosis and prognosis.

Figure 2 Pipeline illustrating TET analysis using clinical, radiology, pathology, omics, or combined multi-omics data. It includes diagnosis (tumor type, risk stratification, surgical planning, subtyping) and prognosis (recurrence, survival). TET, thymic epithelial tumor.

Objectives of the systematic review

With significant advances in oncology diagnostics and prognostics, this review assesses the current state-of-the-art AI applications in TET diagnosis and prognosis. The review is guided by questions addressing: to guide this systematic review, the following key research questions were formulated.

• AI methodologies and diagnostic accuracy (ACC):
  • Which AI methods (DL, ML, radiomics) are used for TET diagnosis?
  • How have AI-based risk tools influenced preoperative decisions, treatment planning and prognostic counseling?
  • How does AI performance compare to traditional clinical and expert methods?
• Prognostic and predictive modeling: how accurately do AI models predict clinical outcomes such as stage, invasiveness, recurrence, metastasis, and survival?
• Methodological limitations and challenges:
  • What challenges (dataset limitations, overfitting, lack of validation, interpretability) exist in current AI models?
  • Are there performance differences across TET subtypes or imaging modalities?
• Clinical integration and future directions:
  • What steps are needed for clinical adoption (validation, regulatory approval, workflow integration)?
  • How might explainable AI (XAI) and multimodal data fusion increase clinical utility and acceptance?

We present this article in accordance with the PRISMA reporting checklist (available at https://med.amegroups.com/article/view/10.21037/med-25-34/rc).

Methods

Literature review

First, the study was registered in PROSPERO (CRD420251015861). Second, we defined the reviewers for the screening step. A.S.E.D., M.D. and T.A.M. were involved in the review. A comprehensive search strategy was developed and applied across six major databases: Medline (Ovid), Embase (Embase.com), Web of Science, Cochrane Central Register of Controlled Trials, Cumulative Index to Nursing and Allied Health Literature (CINAHL) Plus, and Google Scholar (via Publish or Perish). There were no date restrictions, allowing for the inclusion of all relevant literature up to March 2025.

Eligibility criteria and study selection

Search queries combined controlled vocabulary and free-text terms focused on TETs (e.g., “thymoma”, “thymic carcinoma”, “thymic epithelial tumor”) and AI methodologies (e.g., “artificial intelligence”, “machine learning”, “deep learning”). Duplicate entries were removed, resulting in 582 unique studies. Detailed information regarding search terminology for each database and corresponding search research results is provided in Appendix 1. Studies applying AI to TETs for diagnostic or prognostic purposes were included. The following exclusion criteria were applied:

• Studies not focused on TETs;
• Studies not employing AI methods;
• Studies unrelated to diagnosis or prognosis applications;
• Abstract-only studies;
• Not a peer-reviewed research article;
• Non-English language articles;
• Studies not conducted on human subjects.

Data extraction

Screening was conducted using Rayyan software (23) by A.S.E.D., M.D. and T.A.M. First, titles and abstracts were screened by A.S.E.D. and M.D.; subsequently, full-text reviews were conducted independently by A.S.E.D. and M.D. Discrepancies were resolved through discussion with a third reviewer T.A.M.

Following the completion of the screening process, A.S.E.D. proceeded to extract data using the following variables: DOI, title, publication year, study aim (with focus on AI and TETs), diagnostic or prognostic objective, type and source of dataset, data availability, dataset link, dataset modality, staining type, sample type, number of samples, methods, AI technique (e.g., ML, DL), model type (e.g., supervised), and outcomes [e.g., area under the curve (AUC), ACC]. Limitations, findings, and contributions were also collected to prepare for the discussion.

Data synthesis

Due to substantial differences in AI applications, study designs, algorithms, patient cohorts, evaluation methods, and reported metrics, we chose a narrative synthesis over meta-analysis. This approach allows a more flexible and descriptive review of each study. It is particularly suitable for diagnostic ACC studies where patient groups and test conditions often vary, introducing heterogeneity and potential bias. Meta-analysis is generally discouraged in such cases. We also did not assess bias formally, as many studies provided insufficient methodological detail. In addition, there is no accepted reference standard for AI-based analysis of TETs.

Results

Literature search

The Preferred Reporting Items for Systematic Reviews and meta-analysis (PRISMA) flow diagram (Figure 3) illustrates the systematic search for studies on the use of TETs and AI for diagnosis and prognosis. The initial search yielded 998 records across the six databases mentioned above. After removing 416 duplicates, 582 unique records were screened by A.S.E.D. and M.D. A total of 132 full-text articles were assessed. According to the eligibility criteria, 54 studies were excluded, and 23 records resulted in disagreement between A.S.E.D. and M.D. T.A.M. selected 10 from these conflicting records. As a result, a total of 65 studies were selected for the final analysis.

Figure 3 Flowchart of the study according to the PRISMA statement. AI, artificial intelligence; CINAHL, Cumulative Index to Nursing and Allied Health Literature; PRISMA, Preferred Reporting Items for Systematic Reviews and meta-analysis; TET, thymic epithelial tumor.

Applications of AI methods for diagnosis

This section focuses on research studies utilizing AI for diagnosis. We have divided these studies into two categories: preoperative assessment and subtyping, according to the articles reviewed.

Preoperative assessment

The studies addressed significant diagnostic challenges, including the differentiation of benign from malignant lesions, the distinction between thymomas and other types of thymic or non-thymic tumors, risk stratification, and supporting surgical planning.

Malignant or benign

Ma et al. (24) designed a radiomics-based model using CT images of 100 patients (54 with benign lesions, 46 with malignant lesions). For feature extraction, the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was used. After feature selection, a Logistic Regression (LR) model was used for classification. The model achieved an ACC of 0.82 on the test set.

Mayoral et al. (25) analyzed 239 patients with 25 visual CT features and 101 radiomic features. They aimed to classify benign vs. malignant lesions, and thymoma vs. TC using three SVM models: one with conventional, one with radiomic, and one with combined features. The combined model performed best with an AUC of 0.71 for benign vs. malignant, and 0.81 for thymoma vs. TC.

Thymomas vs. other thymic tumors

Thymic cysts are rare, typically benign lesions that require no intervention, whereas TETs are primarily treated with surgical removal. Differentiating between the two on imaging can be challenging, which is clinically important. Recent studies have explored the application of AI to improve diagnostic ACC using radiomics and DL on contrast-enhanced CT scans.

Yang et al. (26,27) conducted multicenter studies using deep transfer learning (DTL) and clinical-radiomics models to distinguish thymomas from thymic cysts. Their models, those based on three-dimensional (3D) ResNet50 and Densenet169, demonstrated high diagnostic performance (AUCs up to 0.99), with XAI methods such as Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) used to enhance interpretability.

Zhao et al. (28) used radiomics and LR to differentiate TETs from cysts (AUC of 1.0) and for risk stratification of TETs (AUC of 0.76). Similarly, Wang et al. (29) applied an XGBoost model to separate mediastinal cysts from tumors in 592 patients, outperforming radiologists with an AUC of 0.97. Liu et al. (30) achieved comparable ACC (AUC of 0.94) using LR to distinguish anterior mediastinal cysts from B1/B2 thymomas.

Zhang et al. (31) created a CT-based radiomics nomogram combining LASSO-selected features with CT data to differentiate thymic cysts from TETs in 190 patients. The model achieved an AUC of 0.94, outperforming both the CT-only model (AUC of 0.85) and radiologist assessments, thereby demonstrating superior diagnostic ACC.

Differentiating thymoma from TC is critical due to their divergent tumor behavior and treatment strategies. Dai et al. (32) developed a simplified RF model using key CT features, achieving an AUC of 0.96. Ohira et al. (33) used radiomics and LASSO to reach an AUC of 0.88, while Tian et al. (34) applied radiomic feature selection and RF modeling, achieving an AUC of 0.90.

Thymic tumors vs. non-thymic tumors

Accurate differentiation between thymic tumors and non-thymic prevascular mediastinal masses, particularly lymphomas, remains challenging due to overlapping imaging features. Recent developments in AI, radiomics, and molecular profiling have provided promising tools to enhance diagnostic precision and guide treatment strategies.

Several studies have focused on distinguishing TETs from mediastinal lymphomas. Li et al. (35) developed a PET/CT-based radiomics model incorporating clinical variables [age, lactate dehydrogenase (LDH), SUV_avg] via an RF algorithm, achieving strong performance (AUC of 0.95 in training and 0.91 in internal validation). Xia et al. (36) used an ensemble ML approach combining contrast-enhanced CT radiomics and clinical data across two centers, yielding AUCs of 0.85–0.86 and boosting the performance of junior radiologists.

Addressing cases without defining clinical markers like myasthenia gravis or calcification, Huang et al. (37) implemented a radiomics-driven LR model, achieving high training ACC (AUC of 0.99), but its external validation was less robust (AUC of 0.80). Lin et al. (38) introduced dynamic contrast-enhanced MRI and decision trees, achieving 72.6% overall ACC for multiclass classification, with even stronger performance in binary tasks (e.g., Hodgkin vs. non-Hodgkin lymphoma).

Kirienko et al. (39) found clinical features to outperform non-contrast CT radiomics alone (AUC of 0.98), but adding radiomics provided complementary value. A molecular diagnostic approach by Latiri et al. (40) used DNA methylation profiling to distinguish T-lymphoblastic lymphoma from lymphocyte-rich thymoma, achieving perfect validation separation using six gene promoters in an ML model.

Dong et al. (41) differentiated mass-like thymic hyperplasia from low-risk thymoma using a radiomics nomogram (AUC of 0.90), while Agar et al. (42) applied transformer-based DL on CT, achieving near-perfect thymoma classification.

Comparative approaches underscore the potential of radiomics. Chang et al. (43) evaluated 376 patients in a multicenter study, comparing radiomics-based LightGBM models with 3D CNNs for TET classification. LightGBM outperformed CNNs with an AUC of 0.95. He et al. (44) developed radiomics, clinico-radiologic, and combined models using contrast-enhanced CT to differentiate TET from lymphoma; the radiomics model achieved an AUC of 0.96 and ACC of 0.90, outperforming traditional methods.

Shang et al. (45) applied radiomics-based ML to differentiate anterior mediastinal cysts from thymomas and stratify thymomas by risk. Their models achieved AUCs of 0.92 and 0.81, respectively.

Risk stratification

Risk stratification is central to the management of TETs, and among AI applications in this domain, it remains the most thoroughly investigated. Numerous studies have investigated a range of ML and DL techniques to classify TETs by WHO histology, clinical stage, and aggressiveness, using CT, MRI, and PET-based imaging, often in combination with clinical data.

Zhang et al. (46) achieved excellent results using a decision tree with CT radiomics and semantic features, with AUCs of 0.96 (stage) and 0.86 (WHO subtype) in a multi-center cohort. Similarly, Yoshida et al. (47) used a modified visual geometry group 16 (VGG16) CNN on CT images (AUC of 0.69); radiologists improved with AI support, although the improvement was not statistically significant.

Liu et al. (48) developed increasingly sophisticated models: a radiomics-based SVM with transfer learning (AUC of 0.95), followed by a 3D DL pipeline combining nnU-Net and ResNet50 (AUC of 0.89, sensitivity of 1.00), outperforming simpler models. Zhou et al. (49) and Chen et al. (50) combined handcrafted and deep features in radiomics nomograms; Chen’s transformer-enhanced model had the strongest external validation (AUCs of 0.85–0.87).

Other CT-based studies highlighted the value of multiphasic imaging. Liu et al. (51) combined non-enhanced CT (NECT), arterial phase CT (AECT), and venous phase CT (VECT) phase features (AUC of 0.94), and Feng et al. (52) used NECT with SVM (AUC 0.84), although sensitivity for TC detection remained limited. PET/CT studies such as Ozkan et al. (53) and Nakajo et al. (54) paired PET radiomics with DL, achieving AUCs of 0.83 and 0.94, demonstrating strong discrimination for carcinoma.

MRI-based approaches have also been effective. Xiao et al. (55) developed a logistic regression nomogram using MRI radiomics, apparent diffusion coefficient (ADC), and morphology (AUC of 0.88), consistent with earlier SVM results from Xiao et al. (56).

Other efforts focused on model interpretability or staging. Blüthgen et al. (57) applied RFs with SHAP to predict WHO subtype and TNM stage (AUCs 0.88, 0.84) and Moon et al. (58) used an end-to-end 3D DL pipeline for staging (ResNet50, AUC 0.81).

Additional radiomics-based CT studies by Yu et al. (59), Ren et al. (60), Hu et al. (61), and Shen et al. (62) confirmed the feasibility of LR and RF classifiers, typically achieving AUCs ranging from 0.84 to 0.87, with improved performance when combined with clinical staging variables.

Several radiomics studies have advanced risk stratification in TETs using CT-based modeling. Chen et al. (63) developed a nomogram integrating eight LASSO-selected radiomic features and CT traits with AUCs of 0.89 for the training set and 0.86 for the validation set. Dong et al. (64) combined contrast-enhanced CT (CECT) radiomics with clinical factors (yielding a validation AUC of 0.87).

Liang et al. (65) used multiphasic CT (NECT, AECT, VECT) in a stacked ensemble model across 131 cases, incorporating phase-difference features and demographics. Their model achieved AUCs of 0.95 for the training set and 0.90 for the validation set, emphasizing phase heterogeneity. Similarly, Liu et al. (66) classified TETs into low/high-risk and carcinoma groups using radiomics-clinical fusion, reaching AUCs of 0.88–0.94 and demonstrating superior performance compared to radiologists.

Liu et al. (67) incorporated habitat-based and peritumoral features, derived via k-means clustering, with clinical data in 140 thymoma patients. Using XGBoost, the model achieved an AUC of 0.81 and an ACC of 0.77, emphasizing the prognostic value of the tumor microenvironment.

Liufu et al. (68) showed that multiphasic CECT-based radiomics, combined with radiological features, achieved high diagnostic ACC with AUCs of 0.99 for the training set and 0.94 for the validation set in distinguishing high- from low-risk TETs. Kayi Cangir et al. (69) identified four key CT radiomics features and found that the K-nearest neighbors (KNN) and LR models reached AUCs of 0.94 in the validation set, supporting their use in preoperative risk assessment. Wang et al. (70) used texture-based radiomics from NECT and CECT images to predict the TET risk and stage, achieving AUCs up to 0.87, with CECT models outperforming radiologists (P=0.03). Similarly, Sui et al. (71) applied LASSO to derive radiomics signatures, reporting AUCs of 0.83 and 0.86 for risk and stage prediction, respectively, also surpassing radiologist performance (ACCs of 0.87 vs. 0.78).

Preoperative planning

Recent work has increasingly used AI to predict surgical complexity in TETs, focusing on resectability and vascular involvement to guide operative planning.

Li et al. (72) developed a CT-based radiomics model focused on the superior vena cava and left innominate vein. Using non-contrast CT scans of 204 patients, they extracted over 2,600 features that were filtered by LASSO and SVM. Their combined two-dimensional (2D)/3D model outperformed expert radiologists with AUCs up to 0.99.

Similarly, Onozato et al. (73) applied XGBoost to preoperative CT data from 212 patients across three institutions to predict concurrent resections. Their model achieved AUCs of 0.80 (training) and 0.82 (validation), with lung and pericardium submodels performing well (AUCs of 0.81 and 0.92).

DL models have also enhanced staging and segmentation. Yang et al. (74) introduced a 3D DenseNet to classify Masaoka-Koga stage I vs. II using CECT in 174 cases, achieving an AUC of 0.77 and identifying tumor enhancement as the key predictor.

Araujo-Filho et al. (75) built a CT-based radiomics model using elastic-net LR in 243 patients to predict incomplete resections (R1/R2) and advanced-stage disease (III/IV), reaching AUCs of 0.80 and 0.71, respectively. Texture metrics from the gray-level co-occurrence matrix, run-length matrix, and size zero matrix emerged as key discriminators.

Tumor segmentation, critical for surgical and radiologic workflow, has benefited from advanced DL architectures. Li et al. (76) introduced DSC-Net, a CNN with dense residual connections and pseudo-color CT preprocessing, trained on 310 patients. It outperformed U-Net variants, achieving 92.96% ACC, 87.86% intersection over union (IoU), and a boundary F1 score of 0.91.

Li et al. (77) further developed TA-Net, a hybrid Transformer-CNN model incorporating self-attention and attention gates, yielding a Dice score of 89.92%, an IoU of 83.80%, and 92.49% ACC. In 2023, Li et al. (78) proposed MG-Net, which extended this with global-enhanced convolution, spatial attention, and adaptive fusion to segment across WHO subtypes. Trained on the same cohort, the model achieved Dice of 91.57%, excelling in AB subtypes (Dice of 93.84%) and underperforming slightly in B3 (Dice of 88.19%). For a more detailed comparison of TET diagnosis (on retrospective) studies, please refer to Table 1.

Subtyping

Four recent studies developed AI models to subtype TETs from histopathological images. Recent advances in AI have enabled increasingly accurate thymoma subtyping using both histopathological and radiological data.

Wang et al. (79) introduced a weakly supervised, interpretable model for classifying thymoma using whole-slide images (WSIs). Leveraging attention-based multi-instance learning, the model classified 222 slides into five WHO subtypes (A, AB, B1, B2, B3), achieving an AUC of 0.94 and ACC of 0.72. Category heatmap visualization revealed subtype heterogeneity, and cell-level analyses confirmed biologically relevant features.

Lv et al. (80) explored hyperspectral imaging for thymoma typing, combining spectral data with a Res2Net-48 CNN. They converted spectral signatures from 180 tissue samples into Gramian angular field images and classified six subtypes (A, AB, B1, B2, B3, TC) with an ACC of 95%.

Zhang et al. (81) developed a vision transformer model, MC-ViT, using 323 WSIs to subtype eight TET classes—including mixed B1 + B2 and B2 + B3. The model used multi-scale patches (10×, 20×, 40×) and dual branches: local feature extraction (CAST) and subtype prediction (WT). MC-ViT achieved an ACC of 95% and an F1-score of 0.94; CAST alone achieved an F1-score of 0.94. Expanding to multimodal learning, Zhang et al. (82) proposed MHD-Net, trained on paired CT and pathology but deployable with CT alone, achieving an ACC of 89.89% on 126 patients. For a more detailed comparison of the TET subtyping studies, please refer to Table 2.

Applications of AI methods for prognosis

Survival

Tian et al. (34) performed a retrospective study using radiomic features from preoperative CT scans of 124 TET patients to predict risk-group type, TNM stage, and survival outcomes. Using 851 extracted features, RF and random survival forest (RSF) models were built. The model achieved survival prediction AUCs of 0.94 [overall survival (OS)] and 0.81 [progression-free survival (PFS)] when combining radiomic and clinical data. High-risk patients exhibited significantly worse survival outcomes.

Kim et al. (83) used Lunit SCOPE IO, a CNN-based AI tool, to analyze haematoxylin and eosin (H&E) WSIs from 35 unresectable thymic tumor cases and measure tumor-infiltrating lymphocytes (TILs). TILs increased post-neoadjuvant therapy. Patients with higher intratumoral TILs had improved survival as well as lower recurrence rates. Patients with TIL levels >10% had a 5-year survival rate of 73.8%. Desert immune phenotypes were linked to poor response (OR of 0.04).

Yang et al. (84) developed an 11-gene prognostic model based on immune microenvironment profiles from 121 thymoma patients in The Cancer Genome Atlas (TCGA). Two immunotypes with distinct immune cell patterns and survival outcomes were identified. The model, including genes like CD1C and CELF5, showed high prognostic power (AUC of 0.93) and was validated using Gene Expression Omnibus (GEO) data. High-risk scores were linked to worse survival, higher tumor mutation burden, increased stemness, and distinct immune cell infiltration. CD1C overexpression was confirmed by immunohistochemistry.

Zhang et al. (85) used clustering, gene set variation analysis (GSVA), and ML to study metabolic reprogramming in thymoma by integrating RNA sequencing (RNA-seq) and metabolomic data from 121 TCGA patients and 10 tissue samples. The lacto/neolacto-series pathway correlated with progression. An RF model and SHAP scores identified B3GNT5 as a potential prognostic biomarker.

Han et al. (86) developed a DL model based on a quasi-3D U-Net to segment TETs in 18F-FDG PET/CT scans from 186 patients. Automated parameters [SUV_max, metabolic tumor volume (MTV), total lesion glycolysis (TLG)] showed high agreement with manual values [concordance correlation coefficient (CCC) >0.92]. SUV_max was found to be an independent prognostic factor. Segmentation achieved a mean Dice score of 0.83.

Recurrence

Zhang et al. (85) analyzed proteomic data from 30 TET samples, identifying HNRNPA2B1 as a marker associated to poor survival and recurrence, and suggested ergotamine as a drug target. Su et al. (87) used RNA-seq from 114 TCGA patients to develop a 4-long noncoding (lnc) RNAs classifier via LASSO Cox regression, which demonstrated superior performance compared to clinical staging with AUCs of 0.80 (3-year) and 0.79 (5-year). Both highlight molecular models’ value in predicting TET recurrence. For more detailed comparison of TET prognostic studies, please refer to Table 3.

Discussion

AI is rapidly transforming healthcare. In medicine, its impact is especially profound for diseases that are both rare and highly variable (89). TETs are difficult to diagnose accurately. Our findings show that AI enhances diagnostic precision and speed, supporting timely, personalized care and helping patients receive optimal treatment.

AI applications in TETs are diverse, aiding diagnosis, prognosis, and treatment planning. In radiology, AI supports tumor classification, risk stratification, and surgical planning. In pathology, DL helps identify subtypes beyond human perception. Prognostically, AI predicts outcomes like metastasis and recurrence. This systematic review of 65 studies highlights AI’s growing role in enhancing diagnostic ACC and prognostic assessment in TETs, demonstrating its clinical relevance and potential impact.

The comparison between human radiologists and AI models has been a significant focus in recent research. Studies (29,70) have developed a top-performing model that provided valuable feedback to radiologists, helping them understand the logic behind radiomics textures. This feedback mechanism enhances the interpretability and clinical utility of AI tools. Similarly, Zhang et al. (31)’s radiomics nomogram method achieved higher ACC than conventional CT models and radiologist judgments, consistently yielding higher AUC values. In risk stratification, Yoshida et al. (47) developed a CT-based DL model to differentiate low- and high-risk thymomas. It achieved 71.3% ACC, outperforming radiologists (whose ACC ranged from 61.9% to 70.0%). Xia et al. (36)’s clinico-radiomics model improved TET-lymphoma differentiation, especially for less experienced radiologists by reducing diagnostic bias.

Subjectivity and interobserver variability are major challenges in both radiology and pathology, and remain significant barriers in subtyping efforts. To reduce bias, Wang et al. (79) used a consensus-labeled dataset. The issue of subjectivity in radiological assessment has been highlighted by several studies. Ohira et al. (33) demonstrated that radiomics features outperformed radiologists in diagnostic ACC. Chen et al. (63) identified drawbacks including inter-reader variability and poor repeatability. Shang et al. (45) noted that CT interpretations are influenced by radiologists’ experience, which limits their ability to distinguish mediastinal cysts from thymomas.

A solution to the subjectivity in radiological assessment is the use of automatic segmentation. Several researchers have identified this as an important application area. Wang et al. (29) suggested using DL methods with computer vision to train an automated model for lesion regions of interest (ROI) delineation, enabling clinical translation. Li et al. (77) proposed a hybrid CNN-transformer architecture, TA-Net, for effective thymoma segmentation on chest CT, achieving superior performance and consistent delineation to assist radiologists. Several studies have demonstrated that automated segmentation reduces intra-observer variations (47,86). Liu et al. (48) used a 3D segmentation model on radiomics data, which outperformed 2D segmentation, demonstrating the advantages of three-dimensional approaches in capturing the complex spatial characteristics of tumors.

Multi-approach methodologies have shown promising results in improving diagnostic ACC. Lin et al. (38) built ML models for predicting pathological subtypes of prevascular mediastinal tumors using clinical and MRI data, achieving varying sensitivity rates for detecting different tumor types. Liu et al. (48) developed a DL model combining segmentation and risk stratification to improve diagnostic consistency. Moon et al. (58) proposed a deep-learning framework for automatic segmentation and classification of high-risk TET cases, while Mayoral et al. (25) focused on predicting pathologic diagnoses of anterior mediastinal masses using ML models based on CT conventional and radiomic features. Their research showed that the best diagnostic performance was achieved by integrating both conventional and radiomic features within the predictive frameworks. Shang et al. (45) addressed two core challenges: isolating thymic cysts from thymomas and stratifying thymomas by risk. Han et al. (86) presented a two-stage DL model based on automatic segmentation of thymic tumors and recurrence prediction.

The integration of multimodal data improves diagnostic ACC. Zhang et al. (82) developed MHD-Net, a memory-aware hetero-model distillation network that transfers multimodal knowledge using only radiology data. Its spatial fusion module enhances radiology-pathology feature fusion, while the typing memory module stores pathology features to boost cross-modal learning. Liu et al. (30) explored CT-based radiomics for diagnosing anterior mediastinal cysts as well as type B1 and B2 thymomas, finding that a combined model using enhanced CT and clinical factors shows potential for differential diagnosis. Other CT radiomics studies (55,59-62) confirmed ML models’ feasibility for classifying prevascular mediastinal lesions and showed improved performance with the addition of clinical staging data.

External validation was commonly performed across numerous studies and showed promising results (26,27,29,36,37,46,49,50,62,73) (Table 1). Robust external validation is essential to ensure AI models generalize well, enhancing reproducibility and enabling reliable clinical implementation across diverse patient populations and healthcare settings.

XAI techniques have been increasingly incorporated to make AI processes more transparent (26,29,57,58,72). Used techniques such as SHAP and Grad-CAM analyses to provide insights into how their models make decisions, enhancing trust and understanding among clinical users.

These advancements in AI applications for TETs diagnosis and prognosis represent significant progress in addressing the challenges associated with these rare and heterogeneous tumors. By applying various AI techniques, from radiomics to DL and multimodal approaches, researchers are developing tools that can assist clinicians in making more accurate and early diagnoses, ultimately leading to better patient outcomes through personalized treatment strategies.

Limitations and future work

Most reviewed studies share common limitations. Many are retrospective and single-center, which limits the generalizability of their findings. A major challenge is the lack of external validation, which hinders the assessment of model robustness. Manual segmentation, which is frequently used, is prone to observer bias. Additionally, small sample sizes and the absence of standardized imaging protocols or inconsistent use across different scanners further limit model reliability. The interpretability of DL models remains limited, affecting clinical trust.

In subtyping studies, data leakage is a critical concern. WSIs or image patches from the same patient are often split across training, validation, and test sets. This leads models to learn patient-specific features rather than disease-related patterns, artificially inflating performance and reducing generalizability. Such practices compromise the true evaluation of model effectiveness and highlight the need for more rigorous data handling and validation strategies in future research. To support clinical adoption, future efforts must be directed toward filling these gaps:

• Conduct prospective, multicenter studies to capture diverse clinical populations and improve model relevance;
• Use standardized imaging protocols to reduce variability and bias, ensuring consistent, reliable results across centers;
• Secure diagnosis agreement from at least three experts to enhance ACC and reliability;
• Perform external validation on independent datasets to evaluate real-world model performance;
• Replace manual ROI segmentation with validated automated methods to minimize user bias and improve reproducibility;
• Integrate multimodal data such as radiomics, clinical, histopathological, and molecular information to enable more comprehensive feature extraction;
• Incorporate XAI techniques to enhance model transparency and clinical trust.
• Test models across different scanners and protocols to ensure cross-device robustness;
• Utilize federated learning to train models across institutions without data sharing, preserving privacy while boosting generalizability;
• Develop foundational models trained on large, diverse datasets, then fine-tune for specific tasks, improving training efficiency and adaptability;
• Expand the focus to include prognosis prediction, thereby enhancing clinical utility beyond diagnosis and subtyping.

Conclusions

Most preoperative studies integrate clinical and radiomic data to distinguish benign from malignant thymic tumors or differentiate thymomas from other tumor types. They also aim to stratify patients by risk. Feature reduction techniques, such as LASSO or deep reductions using CNNs, are commonly used. Final classifications typically rely on statistical models or ML classifiers. While the applications of DL methods for classification remain limited, such approaches are gradually emerging. Other studies support surgical planning by applying DL for tumor segmentation, followed by staging classification. Classifiers trained on segmentation outputs can estimate tumor stage and predict resectability.

Subtyping efforts primarily use H&E WSIs, which are preprocessed into small image patches for CNN input. However, many studies fail to avoid data leakage, often including patches from the same patient in both training and test sets—compromising reliability. Radiomics-based subtyping is rare and lacks strong predictive features.

Prognostic modeling remains less developed, with limited literature and a scarcity of external validation. Some models use omics data with ML to predict survival or recurrence, though they often exclude imaging and histology, thereby reducing clinical relevance.

To enhance trust in AI, newer studies increasingly compare model performance with clinical experts. There is also growing use of XAI techniques like SHAP and Grad-CAM to interpret model decisions and reduce black-box concerns.

Risk stratification is the most common application across all goals, guiding preoperative planning and supporting personalized care. However, most models still require more thorough external validation and greater multimodal integration to ensure clinical utility.

Acknowledgments

The authors thanks to the reviewers for their contributions and Wichor Bramer from the Erasmus MC Medical Library for developing the search strategies.

Footnote

Provenance and Peer Review: This article was commissioned by the Guest Editor (Malgorzata Szolkowska) for “The Series Dedicated to the 14th International Thymic Malignancy Interest Group Annual Meeting (ITMIG 2024)” published in Mediastinum. The article has undergone external peer review.

Reporting Checklist: The authors have completed the PRISMA reporting checklist. Available at https://med.amegroups.com/article/view/10.21037/med-25-34/rc

Peer Review File: Available at https://med.amegroups.com/article/view/10.21037/med-25-34/prf

Funding: This work was supported by Hanarth Fonds grant 2022 (PI: Stephanie Peeters).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://med.amegroups.com/article/view/10.21037/med-25-34/coif). “The Series Dedicated to the 14th International Thymic Malignancy Interest Group Annual Meeting (ITMIG 2024)” was commissioned by the editorial office without any funding or sponsorship. J.v.d.T. serves as an unpaid editorial board member of Mediastinum from May 2024 to December 2025. D.D. reported consulting fees from MSD, Amgen, Roche, BMS, Astra Zeneca, and Pfizer. D.D.R. reported grants or contracts from various entities, indicating institutional financial interests without personal financial gain from organizations such as AstraZeneca, BMS, Beigene, Philips, Olink, and Eli Lilly, where his involvement includes research grants, support, and advisory board participation. S.P. reported Hanarth grant financed the INTHYM project, on AI for histopathological classification and recurrence prediction of TET. The authors have no other 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.

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. International Agency for Research on Cancer, World Health Organization, International Academy of Pathology. Thoracic tumours. 5th ed. IARC; 2021.
2. Hsu CH, Chan JK, Yin CH, et al. Trends in the incidence of thymoma, thymic carcinoma, and thymic neuroendocrine tumor in the United States. PLoS One 2019;14:e0227197. [Crossref] [PubMed]
3. Travis WD, Brambilla E, Burke AP, et al. Introduction to The 2015 World Health Organization Classification of Tumors of the Lung, Pleura, Thymus, and Heart. J Thorac Oncol 2015;10:1240-2. [Crossref] [PubMed]
4. Wolf JL, van Nederveen F, Blaauwgeers H, et al. Interobserver variation in the classification of thymic lesions including biopsies and resection specimens in an international digital microscopy panel. Histopathology 2020;77:734-41. [Crossref] [PubMed]
5. Roden AC, Yi ES, Jenkins SM, et al. Reproducibility of 3 histologic classifications and 3 staging systems for thymic epithelial neoplasms and its effect on prognosis. Am J Surg Pathol 2015;39:427-41. [Crossref] [PubMed]
6. Oselin K, Girard N, Lepik K, et al. Pathological discrepancies in the diagnosis of thymic epithelial tumors: the Tallinn-Lyon experience. J Thorac Dis 2019;11:456-64. [Crossref] [PubMed]
7. Verghese ET, den Bakker MA, Campbell A, et al. Interobserver variation in the classification of thymic tumours--a multicentre study using the WHO classification system. Histopathology 2008;53:218-23. [Crossref] [PubMed]
8. Dawson A, Ibrahim NB, Gibbs AR. Observer variation in the histopathological classification of thymoma: correlation with prognosis. J Clin Pathol 1994;47:519-23. [Crossref] [PubMed]
9. Yoon SH, Choi SH, Kang CH, et al. Incidental Anterior Mediastinal Nodular Lesions on Chest CT in Asymptomatic Subjects. J Thorac Oncol 2018;13:359-66. [Crossref] [PubMed]
10. Girard N, Ruffini E, Marx A, et al. Thymic epithelial tumours: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol 2015;26:v40-55. [Crossref] [PubMed]
11. Li HR, Gao J, Jin C, et al. Comparison between CT and MRI in the Diagnostic Accuracy of Thymic Masses. J Cancer 2019;10:3208-13. [Crossref] [PubMed]
12. Strollo DC, Rosado-de-Christenson ML. Tumors of the thymus. J Thorac Imaging 1999;14:152-71. [Crossref] [PubMed]
13. von der Thüsen J. Thymic epithelial tumours: histopathological classification and differential diagnosis. Histopathology 2024;84:196-215. [Crossref] [PubMed]
14. Alqaidy D. Thymoma: An Overview. Diagnostics (Basel) 2023;13:2982. [Crossref] [PubMed]
15. Suster S, Moran CA. Thymic carcinoma: spectrum of differentiation and histologic types. Pathology 1998;30:111-22. [Crossref] [PubMed]
16. Suster S, Moran CA. Histologic classification of thymoma: the World Health Organization and beyond. Hematol Oncol Clin North Am 2008;22:381-92. [Crossref] [PubMed]
17. Suster S, Moran CA. Thymoma, atypical thymoma, and thymic carcinoma. A novel conceptual approach to the classification of thymic epithelial neoplasms. Am J Clin Pathol 1999;111:826-33. [Crossref] [PubMed]
18. Weis CA, Yao X, Deng Y, et al. The impact of thymoma histotype on prognosis in a worldwide database. J Thorac Oncol 2015;10:367-72. [Crossref] [PubMed]
19. Detterbeck FC, Nicholson AG, Kondo K, et al. The Masaoka-Koga stage classification for thymic malignancies: clarification and definition of terms. J Thorac Oncol 2011;6:S1710-6. [Crossref] [PubMed]
20. Tosi D, Damarco F, Franzi S, et al. Outcomes of extended surgical resections for locally advanced thymic malignancies: a narrative review. Gland Surg 2022;11:611-21. [Crossref] [PubMed]
21. Molina TJ, Bluthgen MV, Chalabreysse L, et al. Impact of expert pathologic review of thymic epithelial tumours on diagnosis and management in a real-life setting: A RYTHMIC study. Eur J Cancer 2021;143:158-67. [Crossref] [PubMed]
22. Drevet G, Collaud S, Tronc F, et al. Optimal management of thymic malignancies: current perspectives. Cancer Manag Res 2019;11:6803-14. [Crossref] [PubMed]
23. Ouzzani M, Hammady H, Fedorowicz Z, et al. Rayyan-a web and mobile app for systematic reviews. Syst Rev 2016;5:210. [Crossref] [PubMed]
24. Ma Y, Liu L, Liu J, et al. Prediction of benign and malignant thymic tumors based on radiomics features. In: 2019 IEEE International Conference on Mechatronics and Automation (ICMA). IEEE; 2019.
25. Mayoral M, Pagano AM, Araujo-Filho JAB, et al. Conventional and radiomic features to predict pathology in the preoperative assessment of anterior mediastinal masses. Lung Cancer 2023;178:206-12. [Crossref] [PubMed]
26. Yang Y, Cheng J, Chen L, et al. Application of machine learning for the differentiation of thymomas and thymic cysts using deep transfer learning: A multi-center comparison of diagnostic performance based on different dimensional models. Thorac Cancer 2024;15:2235-47. [Crossref] [PubMed]
27. Yang Y, Cheng J, Peng Z, et al. Development and Validation of Contrast-Enhanced CT-Based Deep Transfer Learning and Combined Clinical-Radiomics Model to Discriminate Thymomas and Thymic Cysts: A Multicenter Study. Acad Radiol 2024;31:1615-28. [Crossref] [PubMed]
28. Zhao W, Ozawa Y, Hara M, et al. Computed tomography radiomic feature analysis of thymic epithelial tumors: Differentiation of thymic epithelial tumors from thymic cysts and prediction of histological subtypes. Jpn J Radiol 2024;42:367-73. [Crossref] [PubMed]
29. Wang X, You X, Zhang L, et al. A radiomics model combined with XGBoost may improve the accuracy of distinguishing between mediastinal cysts and tumors: a multicenter validation analysis. Ann Transl Med 2021;9:1737. [Crossref] [PubMed]
30. 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]
31. Zhang C, Yang Q, Lin F, et al. CT-Based Radiomics Nomogram for Differentiation of Anterior Mediastinal Thymic Cyst From Thymic Epithelial Tumor. Front Oncol 2021;11:744021. [Crossref] [PubMed]
32. Dai H, Huang Y, Xiao G, et al. Predictive Features of Thymic Carcinoma and High-Risk Thymomas Using Random Forest Analysis. J Comput Assist Tomogr 2020;44:857-64. [Crossref] [PubMed]
33. Ohira R, Yanagawa M, Suzuki Y, et al. CT-based radiomics analysis for differentiation between thymoma and thymic carcinoma. J Thorac Dis 2022;14:1342-52. [Crossref] [PubMed]
34. Tian D, Yan HJ, Shiiya H, et al. Machine learning-based radiomic computed tomography phenotyping of thymic epithelial tumors: Predicting pathological and survival outcomes. J Thorac Cardiovasc Surg 2023;165:502-516.e9. [Crossref] [PubMed]
35. Li J, Cui N, Jiang Z, et al. Differentiating thymic epithelial tumors from mediastinal lymphomas: preoperative nomograms based on PET/CT radiomic features to minimize unnecessary anterior mediastinal surgery. J Cancer Res Clin Oncol 2023;149:14101-12. [Crossref] [PubMed]
36. Xia H, Yu J, Nie K, et al. CT radiomics and human-machine hybrid system for differentiating mediastinal lymphomas from thymic epithelial tumors. Cancer Imaging 2024;24:163. [Crossref] [PubMed]
37. Huang X, Wang X, Liu Y, et al. Contrast-enhanced CT-based radiomics differentiate anterior mediastinum lymphoma from thymoma without myasthenia gravis and calcification. Clin Radiol 2024;79:e500-10. [Crossref] [PubMed]
38. Lin CY, Yen YT, Huang LT, et al. An MRI-Based Clinical-Perfusion Model Predicts Pathological Subtypes of Prevascular Mediastinal Tumors. Diagnostics (Basel) 2022;12:889. [Crossref] [PubMed]
39. Kirienko M, Ninatti G, Cozzi L, et al. Computed tomography (CT)-derived radiomic features differentiate prevascular mediastinum masses as thymic neoplasms versus lymphomas. Radiol Med 2020;125:951-60. [Crossref] [PubMed]
40. Latiri M, Belhocine M, Smith C, et al. DNA methylation as a new tool for the differential diagnosis between T-LBL and lymphocyte-rich thymoma. J Pathol 2024;264:284-92. [Crossref] [PubMed]
41. Dong W, Xiong S, Wang X, et al. Development and validation of a contrast-enhanced CT-based radiomics nomogram for differentiating mass-like thymic hyperplasia and low-risk thymoma. J Cancer Res Clin Oncol 2023;149:14901-10. [Crossref] [PubMed]
42. Agar M, Aydin S, Cakmak M, et al. Detection of Thymoma Disease Using mRMR Feature Selection and Transformer Models. Diagnostics (Basel) 2024;14:2169. [Crossref] [PubMed]
43. 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]
44. He W, Xia C, Chen X, et al. Computed Tomography-Based Radiomics for Differentiation of Thymic Epithelial Tumors and Lymphomas in Anterior Mediastinum. Front Oncol 2022;12:869982. [Crossref] [PubMed]
45. Shang L, Wang F, Gao Y, et al. Machine-learning classifiers based on non-enhanced computed tomography radiomics to differentiate anterior mediastinal cysts from thymomas and low-risk from high-risk thymomas: A multi-center study. Front Oncol 2022;12:1043163. [Crossref] [PubMed]
46. Zhang L, Xu Z, Feng Y, et al. Risk stratification of thymic epithelial tumors based on peritumor CT radiomics and semantic features. Insights Imaging 2024;15:253. [Crossref] [PubMed]
47. Yoshida Y, Yanagawa M, Sato Y, et al. Differential diagnosis between low-risk and high-risk thymoma: Comparison of diagnostic performance of radiologists with and without deep learning model. Acta Radiol Open 2024;13:20584601241288509. [Crossref] [PubMed]
48. Liu W, Wang W, Guo R, et al. Deep learning for risk stratification of thymoma pathological subtypes based on preoperative CT images. BMC Cancer 2024;24:651. [Crossref] [PubMed]
49. Zhou H, Bai HX, Jiao Z, et al. Deep learning-based radiomic nomogram to predict risk categorization of thymic epithelial tumors: A multicenter study. Eur J Radiol 2023;168:111136. [Crossref] [PubMed]
50. Chen X, Feng B, Xu K, et al. Development and validation of a deep learning radiomics nomogram for preoperatively differentiating thymic epithelial tumor histologic subtypes. Eur Radiol 2023;33:6804-16. [Crossref] [PubMed]
51. Liu W, Wang W, Zhang H, et al. Development and Validation of Multi-Omics Thymoma Risk Classification Model Based on Transfer Learning. J Digit Imaging 2023;36:2015-24. [Crossref] [PubMed]
52. Feng XL, Wang SZ, Chen HH, et al. Optimizing the radiomics-machine-learning model based on non-contrast enhanced CT for the simplified risk categorization of thymic epithelial tumors: A large cohort retrospective study. Lung Cancer 2022;166:150-60. [Crossref] [PubMed]
53. Ozkan E, Orhan K, Soydal C, et al. Combined clinical and specific positron emission tomography/computed tomography-based radiomic features and machine-learning model in prediction of thymoma risk groups. Nucl Med Commun 2022;43:529-39. [Crossref] [PubMed]
54. Nakajo M, Takeda A, Katsuki A, et al. The efficacy of (18)F-FDG-PET-based radiomic and deep-learning features using a machine-learning approach to predict the pathological risk subtypes of thymic epithelial tumors. Br J Radiol 2022;95:20211050. [Crossref] [PubMed]
55. Xiao G, Hu YC, Ren JL, et al. MR imaging of thymomas: a combined radiomics nomogram to predict histologic subtypes. Eur Radiol 2021;31:447-57. [Crossref] [PubMed]
56. Xiao G, Rong WC, Hu YC, et al. MRI Radiomics Analysis for Predicting the Pathologic Classification and TNM Staging of Thymic Epithelial Tumors: A Pilot Study. AJR Am J Roentgenol 2020;214:328-40. [Crossref] [PubMed]
57. Blüthgen C, Patella M, Euler A, et al. Computed tomography radiomics for the prediction of thymic epithelial tumor histology, TNM stage and myasthenia gravis. PLoS One 2021;16:e0261401. [Crossref] [PubMed]
58. Moon YS, Park B, Park J, et al. Identification and risk classification of thymic epithelial tumors using 3D computed tomography images and deep learning models. Biomed Signal Process Control 2024;95:106473.
59. Yu C, Li T, Yang X, et al. Contrast-enhanced CT-based radiomics model for differentiating risk subgroups of thymic epithelial tumors. BMC Med Imaging 2022;22:37. [Crossref] [PubMed]
60. Ren C, Li M, Zhang Y, et al. Development and validation of a CT-texture analysis nomogram for preoperatively differentiating thymic epithelial tumor histologic subtypes. Cancer Imaging 2020;20:86. [Crossref] [PubMed]
61. Hu J, Zhao Y, Li M, et al. Machine-learning-based computed tomography radiomic analysis for histologic subtype classification of thymic epithelial tumours. Eur J Radiol 2020;126:108929. [Crossref] [PubMed]
62. Shen Q, Shan Y, Xu W, et al. Risk stratification of thymic epithelial tumors by using a nomogram combined with radiomic features and TNM staging. Eur Radiol 2021;31:423-35. [Crossref] [PubMed]
63. Chen X, Feng B, Li C, et al. A radiomics model to predict the invasiveness of thymic epithelial tumors based on contrast-enhanced computed tomography. Oncol Rep 2020;43:1256-66. [Crossref] [PubMed]
64. Dong W, Xiong S, Lei P, et al. Application of a combined radiomics nomogram based on CE-CT in the preoperative prediction of thymomas risk categorization. Front Oncol 2022;12:944005. [Crossref] [PubMed]
65. Liang Z, Li J, Tang Y, et al. Predicting the risk category of thymoma with machine learning-based computed tomography radiomics signatures and their between-imaging phase differences. Sci Rep 2024;14:19215. [Crossref] [PubMed]
66. Liu J, Yin P, Wang S, et al. CT-Based Radiomics Signatures for Predicting the Risk Categorization of Thymic Epithelial Tumors. Front Oncol 2021;11:628534. [Crossref] [PubMed]
67. Liu W, Wang W, Guo M, et al. Tumor habitat and peritumoral region evolution-based imaging features to assess risk categorization of thymomas. Clin Radiol 2024;79:e1117-25. [Crossref] [PubMed]
68. Liufu Y, Wen Y, Wu W, et al. Radiomics Analysis of Multiphasic Computed Tomography Images for Distinguishing High-Risk Thymic Epithelial Tumors From Low-Risk Thymic Epithelial Tumors. J Comput Assist Tomogr 2023;47:220-8. [Crossref] [PubMed]
69. Kayi Cangir A, Orhan K, Kahya Y, et al. CT imaging-based machine learning model: a potential modality for predicting low-risk and high-risk groups of thymoma: “Impact of surgical modality choice”. World J Surg Oncol 2021;19:147. [Crossref] [PubMed]
70. Wang X, Sun W, Liang H, et al. Radiomics Signatures of Computed Tomography Imaging for Predicting Risk Categorization and Clinical Stage of Thymomas. Biomed Res Int 2019;2019:3616852. [Crossref] [PubMed]
71. 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]
72. Li Z, Wang F, Zhang H, et al. The predictive value of a computed tomography-based radiomics model for the surgical separability of thymic epithelial tumors from the superior vena cava and the left innominate vein. Quant Imaging Med Surg 2023;13:5622-40. [Crossref] [PubMed]
73. Onozato Y, Suzuki H, Matsumoto H, et al. Machine learning models from computed tomography to diagnose thymic epithelial tumors requiring combined resection. J Thorac Dis 2024;16:4935-46. [Crossref] [PubMed]
74. Yang L, Cai W, Yang X, et al. Development of a deep learning model for classifying thymoma as Masaoka-Koga stage I or II via preoperative CT images. Ann Transl Med 2020;8:287. [Crossref] [PubMed]
75. Araujo-Filho JAB, Mayoral M, Zheng J, et al. CT Radiomic Features for Predicting Resectability and TNM Staging in Thymic Epithelial Tumors. Ann Thorac Surg 2022;113:957-65. [Crossref] [PubMed]
76. Li J, Sun W, Feng X, et al. A dense connection encoding–decoding convolutional neural network structure for semantic segmentation of thymoma. Neurocomputing 2021;451:1-11.
77. Li J, Sun W, Feng X, et al. A hybrid network integrated convolution and Transformer for thymoma segmentation. Intell Med 2023;03:164-72.
78. Li J, Sun W, von Deneen KM, et al. MG-Net: Multi-level global-aware network for thymoma segmentation. Comput Biol Med 2023;155:106635. [Crossref] [PubMed]
79. Wang C, Du X, Yan X, et al. Weakly supervised learning in thymoma histopathology classification: an interpretable approach. Front Med (Lausanne) 2024;11:1501875. [Crossref] [PubMed]
80. Lv Q, Liang K, Tian C, et al. Unveiling Thymoma Typing Through Hyperspectral Imaging and Deep Learning. J Biophotonics 2024;17:e202400325. [Crossref] [PubMed]
81. Zhang H, Chen H, Qin J, et al. MC-ViT: Multi-path cross-scale vision transformer for thymoma histopathology whole slide image typing. Front Oncol 2022;12:925903. [Crossref] [PubMed]
82. Zhang H, Liu J, Liu W, et al. MHD-Net: Memory-Aware Hetero-Modal Distillation Network for Thymic Epithelial Tumor Typing With Missing Pathology Modality. IEEE J Biomed Health Inform 2024;28:3003-14. [Crossref] [PubMed]
83. Kim DH, Lim Y, Kim S, et al. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes as a biomarker in locally advanced unresectable thymic epithelial neoplasm: A single-center, retrospective, longitudinal cohort study. Thorac Cancer 2023;14:3001-11. [Crossref] [PubMed]
84. Yang Y, Xie L, Li C, et al. Prognostic Model of Eleven Genes Based on the Immune Microenvironment in Patients With Thymoma. Front Genet 2022;13:668696. [Crossref] [PubMed]
85. Zhang X, Zeng B, Zhu H, et al. Role of glycosphingolipid biosynthesis coregulators in malignant progression of thymoma. Int J Biol Sci 2023;19:4442-56. [Crossref] [PubMed]
86. Han S, Oh JS, Kim YI, et al. Fully Automatic Quantitative Measurement of 18F-FDG PET/CT in Thymic Epithelial Tumors Using a Convolutional Neural Network. Clin Nucl Med 2022;47:590-8. [Crossref] [PubMed]
87. Su Y, Chen Y, Tian Z, et al. lncRNAs classifier to accurately predict the recurrence of thymic epithelial tumors. Thorac Cancer 2020;11:1773-83. [Crossref] [PubMed]
88. Zhou Z, Lu Y, Gu Z, et al. HNRNPA2B1 as a potential therapeutic target for thymic epithelial tumor recurrence: An integrative network analysis. Comput Biol Med 2023;155:106665. [Crossref] [PubMed]
89. Wojtara M, Rana E, Rahman T, et al. Artificial intelligence in rare disease diagnosis and treatment. Clin Transl Sci 2023;16:2106-11. [Crossref] [PubMed]