Preoperative prediction of thymoma risk classification with machine learning-based computed tomography radiomics features Authors: Kai Zhao, Yiming Liu, Honghao Xu, Wenhan Cai, Jiamei Jin, Leilei Shen, Zirui Zhu, Herui Han, Mingchuan Hu, Qiang Lu, Zhiqiang Xue Published: 2026 ABSTRACT AbstractBackgroundThymomas are primarily assessed using chest computed tomography (CT); however, this method has inherent limitations such as an overlap between the semantic features of thymomas presenting as localized lesions and other histological subtypes. Therefore, this study aimed to develop a combined radiomics model, based on non-contrast CT, and investigate its potential clinical utility in preoperatively differentiating between the thymoma risk classification.MethodsWe retrospectively analyzed the clinical and imaging data of 436 patients with pathologically confirmed thymomas between January 2010 and October 2023. The cohort comprised 272 low-risk cases (types A, AB, and B1) and 164 high-risk cases (types B2 and B3), which were randomly divided into training (n = 306) and validation (n = 130) sets in a 7:3 ratio. Radiomic features were extracted from the volume of interest on non-contrast CT images. Feature selection was performed using principal component analysis, correlation analysis, least absolute shrinkage, and selection operator algorithms to identify the most discriminative features. A combined radiomics nomogram was developed by integrating significant clinical factors with radiomics scores. The discriminative performance of the model was assessed using receiver operating characteristic curve analysis.ResultsTwo clinicoradiological and 11 radiomics features were identified and used to construct a radiomics nomogram. The diagnostic performance of the nomogram for thymoma risk stratification surpassed that of any single model. The nomogram yielded an area under the curve of 0.753 (accuracy, 71.2%; sensitivity, 62.4%; specificity, 76.7%) in the training cohort and 0.735 (accuracy, 70.8%; sensitivity, 58.8%; specificity, 78.5%) in the validation cohort.ConclusionThe nomogram model integrating clinical factors and radiomics features accurately differentiated the histological subtypes of thymoma. This tool may be helpful in formulating personalized treatment plans in clinical practice and is worthy of clinical promotion and application.Keywords: Thymoma, Histopathological classification, Radiomics analysis, Computed tomography, Machine learning Introduction Among anterior mediastinal masses, thymomas are the most common primary tumors accounting for 47% of all mediastinal tumors [1]. The updated 2021 World Health Organization classification stratifies thymomas into five histological subtypes (types A, AB, B1, B2, and B3) based on the morphology of epithelial cells and the ratio of lymphocytes to epithelial cells [2]. Specifically, type B2 and B3 thymomas are more invasive and associated with higher postoperative recurrence rates and lower survival rates than type A, AB, and B1 thymomas. Furthermore, A, AB, and B1 thymomas are completely surgically resected more frequently than B2 and B3 thymomas [3, 4]. Based on these distinct clinical behaviors, thymomas are often classified into two risk groups: low-risk (types A, AB, and B1) and high-risk (types B2 and B3) [5–7]. Surgical resection is the primary treatment for thymoma, with complete resection being the goal for optimal survival. However, histological subtype is a key factor in planning therapeutic approaches. Patients with low-risk thymomas typically achieve high rates of complete resection and are often managed with surgery alone without neoadjuvant or adjuvant therapy. By contrast, high-risk thymomas are less amenable to complete resection and often require multimodal treatment [8–10]. Consequently, accurate preoperative risk stratification is crucial for guiding surgical planning, informing decisions regarding neoadjuvant or adjuvant therapies, and ultimately improving patient prognosis. Radiomics has been widely used for the early diagnosis, screening, and prognostic assessment of tumors, facilitating the development of personalized treatment strategies. Currently, radiomics utilizes high-dimensional quantitative features extracted from medical images, such as computed tomography (CT) images, to noninvasively quantify tumor heterogeneity and elucidate the underlying malignant characteristics [11]. Non-contrast chest CT is the primary imaging modality for thymoma assessment due to its widespread availability, high spatial resolution, and ability to provide information on tumor vascularity [12, 13]. However, chest CT images of patients with thymomas require evaluation by radiologists, leading to uncertainties that may affect the accurate classification of thymomas. For instance, the semantic features of thymomas presenting as localized lesions, especially those with no obvious invasion of the surrounding structures on CT images, may overlap with those of different histological subtypes. This overlap is unfavorable for preoperative thymoma risk classification [14]. Additionally, owing to the rarity of thymomas, clinicians (including thoracic surgeons and radiologists) in some centers may have limited experience in evaluating thymomas. This highlights the need for a quantitative method to classify thymomas with high accuracy and objectivity, and to reduce reliance on radiologists’ experience. Previous studies suggest that CT-based radiomics can aid in thymoma risk stratification; however, the clinical translation of this technology has been limited, primarily due to small sample sizes (often fewer than 200 patients), which have constrained the robustness and generalizability of the findings [9, 15, 16]. Therefore, our study used a large cohort to develop a radiomics model based on non-contrast CT images using machine learning, which integrated radiomics features with clinical and radiological characteristics. This approach is expected to enhance the accuracy and clinical applicability of preoperative risk prediction for thymomas. Materials and methods Patient recruitment This retrospective study was approved by the Institutional Review Board of our hospital, and the requirement for informed consent was waived. This cohort study has been reported in line with the STROCSS 2024 guidelines [17]. The study cohort comprised patients with pathologically confirmed thymomas who underwent surgical resection at our institution between January 2010 and October 2023. The inclusion criteria were as follows: (1) pathologically confirmed thymoma following surgical resection; (2) no prior anticancer treatment before surgery; (3) availability of a non-contrast chest CT scan performed within two weeks prior to surgery; and (4) complete clinical records. The exclusion criteria were as follows: (1) small tumor size (maximum diameter ≤9 mm); (2) poor image quality due to significant artifacts; (3) a history of other malignancies; and (4) incomplete clinical or imaging data. In total, 436 patients with thymomas were enrolled (Fig. 1). The cohort comprised 223 men and 213 women, aged 25–80 years with a mean age of 55.85 ± 10.99 years. Based on the postoperative pathology, the patients were categorized into a low-risk group (n = 272), consisting of type A (n = 32), AB (n = 155), and B1 (n = 85) thymomas, and a high-risk group (n = 164), consisting of type B2 (n = 115) and B3 (n = 49) thymomas. Flowchart of the patient selection process. CT, computed tomography Histopathological classification According to the 2021 World Health Organization classification, thymomas are categorized into five histological subtypes (A, AB, B1, B2, and B3) based on the morphology of epithelial cells and the lymphocyte to epithelial cell ratio. Type A thymoma is characterized by a population of spindle-shaped epithelial cells with minimal atypia and a notable absence of immature T-lymphocytes. Atypical type A thymoma meets the criteria for type A thymoma but also exhibits focal necrosis, mitotic activity (> 4 mitoses per 2 mm²), and nuclear crowding. Type AB thymoma demonstrates the morphological features of type A thymoma, but also contains abundant immature T cells, either locally or diffusely distributed. Type B1 thymomas most closely resemble the normal thymic architecture, featuring large areas of immature T cells, medullary differentiation (medullary islands), and scant aggregates of polygonal or dendritic epithelial cells (fewer than three adjacent cells). Type B2 thymomas show an increased proportion of polygonal or dendritic epithelial cells, appearing alone or in clusters, against a background of numerous immature T cells. Type B3 thymomas are composed of sheets of epithelial cells exhibiting mild to moderate atypia, with absent or inconspicuous intercellular bridges and a paucity of mature T lymphocytes. CT examination and image analysis All patients underwent non-contrast CT using either a Philips (Amsterdam, the Netherlands) Brilliance iCT scanner or a Siemens (Erlangen, Germany) Somatom Definition scanner. The standard scanning range was extended from the lung apex to the costophrenic angle. In patients with ectopic neck thymomas, the range was adjusted to cover the upper edge of the aortic arch from the inferior border of the hyoid bone. The scanning parameters were as follows: tube voltage of 120 kV, automatic tube current modulation, matrix of 512 × 512, rotation time of 0.6 s, slice thickness of 5 mm for acquisition, and 1.25 mm for reconstruction. Images were reviewed using a mediastinal window setting (window width, 350 HU; level, 50 HU). Two radiologists, each with over 10 years of experience, independently assessed all images. Any discrepancies were resolved by a consensus discussion between the two radiologists. The following CT features were evaluated: tumor size (defined as the maximum axial diameter); location (left-sided, central, or right-sided); morphology (regular or irregular); margin (well-defined or ill-defined); and presence or absence of calcification, necrosis/cystic degeneration, invasion of adjacent structures, and distant metastasis. Image segmentation and feature extraction The patients were randomly allocated to the training (n = 306) or validation sets (n = 130) in a 7:3 ratio. Original CT images in a DICOM format were retrieved from the Picture Archiving and Communication System. A two-dimensional image slice showing the largest tumor diameter was selected from each scan. All images were then resampled to a uniform voxel spacing of 1 × 1 × 1 mm³ to standardize the spatial resolution, which is the default and most widely used method in PyRadiomics to preserve feature stability. Manual delineation of the volume of interest (VOI) was performed using the open-source software ITK-SNAP (version 3.6.0; Fig. 2). The initial contouring was performed by an experienced radiologist. A second radiologist independently reviewed all segmentations to ensure reliability. Any discrepancies were resolved through a consensus discussion between the two readers to finalize the VOIs, thereby enhancing the accuracy and reproducibility of subsequent feature extraction. During manual segmentation, caution was exercised to encompass the entire tumor volume while meticulously avoiding lesion boundaries, adjacent vascular structures, and internal cystic or necrotic areas. Image segmentation. A the outlined ROI ; (B) the combined three-dimensional VOI. ROI, region of interest; VOI, volume of interest Subsequently, 1032 traditional radiomics features were extracted from each VOI using the PyRadiomics platform (version 3.1.0). The extracted feature set encompassed first-order statistics, shape-based features, texture features (derived from the gray-level co-occurrence matrix, gray-level run-length matrix, and gray-level size zone matrix), and higher-order features obtained through wavelet and Laplacian Gaussian filter transforms. The intraclass correlation coefficient was calculated for all features based on the segmentation to ensure feature stability and reproducibility. Only the features demonstrating excellent consistency (intraclass correlation coefficient > 0.75) were retained for subsequent analyses. ICC was calculated based on independent double segmentations by two experienced radiologists before consensus. Each reader segmented the tumor independently; ICC was computed between these two sets of segmentations to assess feature reproducibility. Only features with ICC > 0.75 were retained. This process ensures robustness against inter-observer variation. Radiomics feature selection The 1,032 radiomics features extracted from each VOI underwent feature selection. Briefly, Z-score normalization was applied to all features to eliminate scale differences. Class imbalance in the training set was addressed using the synthetic minority oversampling technique to generate synthetic samples and balance class distribution. Subsequently, feature dimensionality reduction was conducted using least absolute shrinkage and selection operator regression with a 10-fold cross-validation to determine the optimal penalty parameter (λ). The L1 regularization inherent in the least absolute shrinkage and selection operator shrinks the coefficients of irrelevant features to zero, resulting in the selection of a subset of features that are most predictive of the outcome. Model construction and diagnostic efficacy evaluation Three distinct models were developed to differentiate between low- and high-risk thymomas. A clinicoradiological model was constructed using independent predictors identified using univariate and multivariate logistic regression analyses. Concurrently, a radiomics model was built and a radiomics score (Rad-score) was calculated for each patient. Finally, a combined nomogram integrating significant clinicoradiological factors and Rad-scores was established. The diagnostic performance of each model was evaluated and compared using the accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Calibration of the combined nomogram was assessed using a calibration plot, and its clinical net benefit was quantified using decision curve analysis (Fig. 3). Flowchart of the model construction process Statistical analysis All statistical analyses were performed using the SPSS software (version 26.0). The normality of continuous variables was assessed using the Shapiro–Wilk test. Data conforming to a normal distribution were compared using the independent samples t-test, whereas non-normally distributed data were compared using the Mann–Whitney U test. Categorical variables were compared using the chi-squared test or Fisher’s exact test, as appropriate. The independent predictors were identified using univariate and multivariate logistic regression analyses. Spearman’s correlation analysis, least absolute shrinkage, and selection operator regression were employed for dimensionality reduction for radiomics feature selection. Predictive models were constructed using machine-learning algorithms. Model performance was evaluated using the AUC, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. The DeLong test was used to compare the AUCs of different models. Calibration curves and decision curve analysis were used to assess the calibration and clinical utility of the models, respectively. Statistical significance was set at a two-sided P < 0.05. Results Clinical and radiological features This study enrolled 436 patients with thymomas, who were randomly allocated to a training cohort (n = 306) or a validation cohort (n = 130) in a 7:3 ratio. Baseline characteristics, including age and sex, were well balanced between the two cohorts (all P > 0.05, Table 1). The cohort comprised 272 low-risk thymomas (types A, AB, and B1) and 164 high-risk thymomas (types B2 and B3). The distributions of these risk categories were consistent between the training (low-risk, n = 189; high-risk, n = 117) and validation (low-risk, n = 83; high-risk, n = 47) cohorts. The distributions of histological subtypes were well balanced between the training and validation cohorts. In the training cohort (n = 306), the low-risk group comprised 22 cases of type A, 108 cases of type AB, and 59 cases of type B1 thymomas, while the high-risk group included 81 cases of type B2 and 36 cases of type B3 thymomas. In the validation cohort (n = 130), the low-risk group consisted of 10 cases of type A, 47 cases of type AB, and 26 cases of type B1 thymomas, and the high-risk group contained 34 cases of type B2 and 13 cases of type B3 thymomas. Comparative analysis revealed that tumor morphology, margin definition, and invasion of the surrounding tissues were significantly different between the low- and high-risk groups (all P < 0.05; Table 2). Table 1Comparison of clinical and imaging features of thymomas between the training and validation cohortsBaseline CharacteristicsTraining cohort (n = 306)Validation cohort (n = 130)P-valueSex (case)0.836 Male15865 Female14865Age (years)0.332 ≤ 5014956 > 5015774Diameter (cm)0.174 ≤ 513969 > 516761Location0.073 Left11844 Middle5938 Right12948Morphology0.745 Regular18174 Irregular12556Margin0.594 Clear18382 Unclear12348Calcification0.119 Yes4712 No259118Necrosis and cystic degeneration0.689 Yes5621 No250109Invasion of surrounding tissues0.536 Yes14155 No16575Distant metastasis0.768 Yes3316 No273114 Comparison of clinical and imaging features of thymomas between the training and validation cohorts Table 2Comparison of clinical and imaging features between low-risk and high-risk thymomasBaseline characteristicsTraining cohortValidation cohortLow-risk group (n = 189)High-risk group (n = 117)P-valueLow-risk group (n = 83)High-risk group (n = 47)P-valueSex (case)0.4670.273 Male94644520 Female95533827Age (years)0.1240.230 ≤ 5085643224 > 50104535123Diameter (cm)0.9470.767 ≤ 586534425 > 5103643922Location0.1760.328 Left70482816 Middle32273417 Right87422114Morphology< 0.0010.644 Regular130514925 Irregular59663422Margin0.0030.417 Clear126575527 Unclear63602820Calcification0.8630.597 Yes281993 No161987444Necrosis and cystic degeneration0.7810.125 Yes3620174 No153976643Invasion of surrounding tissues< 0.001< 0.001 Yes70712431 No119465916Distant metastasis0.9980.691 Yes201397 No1691047420 Comparison of clinical and imaging features between low-risk and high-risk thymomas Clinical feature selection and model construction Univariate and multivariate logistic regression analyses of clinical and radiological features identified tumor morphology and invasion of surrounding tissues as independent predictors of thymoma risk (all P < 0.05; Table 3). The clinicoradiological model, constructed using these independent factors, demonstrated an AUC of 0.668 (95% confidence interval [CI]: 0.613–0.721), accuracy of 68.3%, sensitivity of 46.2%, and specificity of 82.0% in the training cohort, and an AUC of 0.646 (95% CI: 0.557–0.728), accuracy of 63.8%, sensitivity of 38.3%, and specificity of 78.3% in the validation cohort (Table 4). Table 3Univariate and multivariate analyses used to identify independent predictors of thymoma riskCharacteristicsUnivariate analysisMultivariate analysisOR95%CIP-valueOR95%P-valueSex0.9950.675–1.4670.981NANANAAge0.6520.442–0.9630.650NANANADiameter1.0090.685–1.4880.962NANANALocation1.0790.835–1.3940.561NANANAMorphology2.2291.500–3.312< 0.0012.0421.198–3.4800.009Margin1.8941.274–2.8160.0020.5650.300–1.0620.076Calcification1.0160.576–1.7920.956NANANANecrosis and cystic degeneration1.4120.834–2.3910.200NANANAInvasion of surrounding tissues3.1152.083–4.659< 0.0013.3662.014–5.626< 0.001Distant metastasis0.8590.469–1.5750.624NANANAOR Odds ratio, CI Confidence interval, NA Not applicable Univariate and multivariate analyses used to identify independent predictors of thymoma risk OR Odds ratio, CI Confidence interval, NA Not applicable Table 4Predictive performance of the training and validation cohorts for different modelsModelsAUC95%CIAccuracySensitivitySpecificityPPVNPVTraining cohort Clinicoradiological model0.6680.613–0.7210.6830.4620.8200.6140.711 Radiomics model0.7510.698–0.7980.7250.7350.7200.6190.814 Nomogram model0.7530.700–0.8000.7120.6240.7670.6240.767Validation cohort Clinicoradiological model0.6460.557–0.7280.6380.3830.7830.5000.691 Radiomics model0.7250.640–0.8000.6770.7660.6270.5370.825 Nomogram model0.7350.651–0.8090.7080.5880.7850.6380.747AUC Area under the receiver operating characteristic curve, CI Confidence interval, PPV Positive predictive value, NPV Negative predictive value Predictive performance of the training and validation cohorts for different models AUC Area under the receiver operating characteristic curve, CI Confidence interval, PPV Positive predictive value, NPV Negative predictive value Construction of radiomics prediction model A total of 1032 radiomics features were extracted from each VOI, capturing comprehensive information on the geometry, intensity, and textural heterogeneity of the tumors. Dimensionality reduction was performed using PCA, least absolute shrinkage, and selection operator regression to mitigate the risk of overfitting from high-dimensional data. This process identified 11 key predictive features (Fig. 4) that were used to construct the radiomics model. In the training cohort, the model achieved an AUC of 0.751 (95% CI: 0.698–0.798), accuracy of 75.2%, sensitivity of 73.5%, and specificity of 72.0%. In the validation cohort, the AUC was 0.725 (95% CI: 0.640–0.800), with accuracy of 67.7%, sensitivity of 76.6%, and specificity of 62.7% (Table 4). The 10-fold cross-validation process of LASSO regression. A Determination of the optimal hyperparameter λ, (lambda) for the LASSO model through 10-fold cross-validation. B Curve showing the changes in feature coefficients with varying λ values; each colored line represents the variation in the coefficient of a specific feature. C Selected optimal features with their corresponding weights Model efficacy evaluation A combined radiomics nomogram was developed using logistic regression analysis to preoperatively predict thymoma risk by integrating significant clinical features (tumor morphology and invasion of surrounding tissues) with the radiomics signature (Fig. 5). In the training cohort, the nomogram achieved an AUC of 0.753 (95% CI, 0.700–0.800), with accuracy of 71.2%, sensitivity of 62.4%, and specificity of 76.7%. When evaluated in the independent validation cohort, it yielded an AUC of 0.735 (95% CI: 0.651–0.809), with accuracy of 70.8%, sensitivity of 58.8%, and specificity of 78.5% (Table 4). Notably, the combined radiomics nomogram demonstrated higher AUC values than the clinicoradiological and standalone radiomics models in the training and validation cohorts. The DeLong test revealed that in the training cohort, the nomogram performed significantly better than the clinicoradiological model (P < 0.001) and the standalone radiomics model (P = 0.012). Although the AUC of the nomogram was higher than that of the standalone radiomic model, the difference was not statistically significant (P = 0.863). In the validation cohort, the superior performance of the nomogram compared to the clinicoradiological model remained significant (P = 0.016). However, the differences between the nomogram and standalone radiomics model and between the standalone radiomics model and clinicoradiological model were not statistically significant (both P > 0.05). Collectively, these results indicate that the combined nomogram offers a more robust predictive advantage for thymoma risk stratification compared to other models (Table 5; Fig. 6). Combined radiomics nomogram for thymoma risk stratification. A A radiomics nomogram for thymoma risk stratification was developed in the training cohort, incorporating clinical features, (tumor morphology and surrounding tissue invasion) and radiomics features. B, C Calibration curve analyses of the clinicoradiological model, radiomics model, and radiomics nomogram in the training and validation cohorts, (B: training cohort, n = 306; C: validation cohort, n = 130). The 45° line represents a perfect match between the actual outcomes, (Y-axis) and the probabilities from differential diagnosis, the clinicoradiological model, the radiomics model, and the radiomics nomogram, (X-axis) Table 5Comparison of prediction with the combined radiomics nomogram, clinicoradiological model, and radiomics modelsGroupModel 1Model 2P-valueTraining cohortNomogramRadiomics0.863NomogramClinicoradiological< 0.001RadiomicsClinicoradiological0.012Validation cohortNomogramRadiomics0.509NomogramClinicoradiological0.016RadiomicsClinicoradiological0.109 Comparison of prediction with the combined radiomics nomogram, clinicoradiological model, and radiomics models The AUC values and DCA of the clinicoradiological model, radiomics model, and radiomics nomogram in the training cohort and validation cohort. (A\C: training cohort; B\D: validation cohort). AUC, area under the receiver operating characteristic curve; DCA, Decision curve analysis The reliability and clinical value of the prediction models were assessed. The calibration curve indicated excellent agreement between the predicted probabilities of the combined radiomics nomogram and the actual observed outcomes in the training and validation cohorts, which was confirmed by the non-significant Hosmer–Lemeshow test results (all P > 0.05) (Fig. 5). Decision curve analysis further demonstrated that the application of this combined nomogram provided a higher net benefit across a wide range of threshold probabilities than the clinicoradiological and standalone radiomics models (Fig. 6). These findings collectively confirm the favorable calibration and clinical utility of the combined radiomics nomogram for thymoma risk stratification. Discussion Radiomics enables the quantitative analysis of medical images and captures disease-specific features that may not be discernible through conventional visual assessment. The increasing application of radiomics in oncology has demonstrated a promising potential for augmenting traditional imaging analysis and supporting personalized treatment strategies. In this study, we developed a radiomics model based on non-contrast CT images to predict the histological risk of thymomas preoperatively. The standalone radiomics model showed superior predictive performance compared with the model based solely on clinicoradiological features. Furthermore, the combined nomogram that integrated radiomics and clinicoradiological factors demonstrated the best overall performance in differentiating high-risk thymomas from low-risk ones, with an efficacy comparable to that of the standalone radiomics model, while incorporating clinically relevant variables. The proposed nomogram shows potential as a practical tool to aid clinical decision-making for the individualized management of patients with thymoma. The results of this study indicate that high-risk thymomas predominantly exhibit an irregular shape, unclear boundaries, and invasion of surrounding tissues, whereas low-risk thymomas are more frequently round or oval, with clear boundaries and no invasion. These findings were consistent with those of previous reports [18–20]. The irregular morphology commonly observed in high-risk tumors may be attributed to their higher malignant potential, greater cellular atypia, and heterogeneous growth patterns, leading to asymmetrical expansion. Thus, tumor shape serves as a key radiological indicator for risk stratification. Although the univariate analysis indicated an association between blurred margins and high-risk thymoma, this feature was not retained as an independent predictor in the multivariate model. This suggests that, while margin characteristics possess some discriminatory value, their predictive power may be subordinate to other stronger indicators, such as invasion or shape, when evaluated collectively. Consequently, the diagnostic utility of tumor margins might be better elucidated through future studies with larger cohorts or by analyzing enhanced CT images that provide superior soft tissue contrast. Consistent with the existing literature [21–23], our study confirmed that invasion into adjacent structures is a hallmark of high-risk thymomas, with a significantly higher incidence than that of low-risk tumors. This reinforces the value of peritumoral invasion as a critical imaging marker for evaluating the biological aggressiveness of thymomas. CT is an established first-line imaging modality for the evaluation of anterior mediastinal masses. Consequently, multiple studies have investigated CT-based strategies for distinguishing between low-and high-risk thymomas [19, 24]. Liang et al. [25] found that radiomics based on enhanced CT could be used to identify the histological subtypes of thymomas. The AUC of the radiomics nomogram and radiomics model for distinguishing between low-and high-risk thymomas was higher than that of the clinical model. Thus, the enhanced CT-based radiomics model and radiomics nomogram, integrating clinical factors, CT features, and radiomics features, are useful for identifying thymoma histological subtypes. Liu et al. [26] evaluated 30 patients with high-risk thymomas and 117 patients with low-risk thymomas to identify the key CT features for risk stratification. The results confirmed that high-risk thymomas had more irregular shapes and contours and a higher incidence of invasion into surrounding tissues and organs on CT images compared to low-risk thymomas. Han et al. [27] constructed and validated an imaging model combining clinical factors and non-contrast CT features to distinguish between high-risk and low-risk thymomas. They found that CT imaging features showed no significant differences between patients of different sexes or ages; however, multiple statistically significant differences in the imaging features were observed between the high- and low-risk groups. Sadohara et al. [20] proposed that combining the features of CT and magnetic resonance imaging (such as tumor contour, capsule, septation, and enhancement uniformity) is helpful in distinguishing low-and high-risk thymomas. Furthermore, Xiao et al. [28] showed the good performance of a combined radiomics nomogram that integrated magnetic resonance imaging with tumor shape, ADC coefficient, and radiomics features. Nakajo et al. [29] reported the good predictive efficacy of machine-learning methods based on deep-learning features from 18 F-FDG-PET radiomics. The experimental results of this study demonstrate that radiomics based on non-contrast CT exhibits favorable predictive efficacy for distinguishing between low- and high-risk thymomas. In the training cohort, the combined radiomics and standalone radiomics models achieved significantly higher AUC values than the clinicoradiological model (P < 0.05). In the validation cohort, the superiority of the combined radiomics model over the clinicoradiological model remained significant (P = 0.016). Although the combined model also yielded a higher AUC than the standalone radiomics model (which outperformed the clinicoradiological model), the differences were not statistically significant (P > 0.05). Given that traditional radiological parameters (e.g., tumor capsule integrity and degree of necrosis) remained clinically valuable and considering that the combined model demonstrated a superior performance compared to the individual models, we propose the integrated radiomics nomogram as the final predictive tool for clinical translation. Nonetheless, this study has some limitations. First, the single-center retrospective study design may have introduced selection bias and limited the generalizability of our findings, as patient demographics and clinical practices can vary across institutions. Second, the exclusive use of non-contrast CT images lacks the functional information available from contrast-enhanced CT, magnetic resonance imaging (particularly diffusion-weighted imaging sequences), or positron emission tomography-CT, despite providing anatomical details. The incorporation of multiparametric data in future studies may yield more discriminative features. Third, the absence of genomic data represents a significant limitation as it could offer crucial insights into the underlying biology of thymomas. Consequently, multicenter prospective studies are required in the future to validate and refine the model as our results demonstrate that integrating radiological, clinical, and genomic data is essential for developing a more comprehensive understanding of thymomas and for building robust predictive models to guide personalized treatment. In conclusion, this study developed and validated a non-contrast CT-based radiomics model as a noninvasive, reliable, and reproducible tool for predicting the histological subtypes of thymoma. The radiomics model demonstrated a superior predictive performance compared with the traditional clinicoradiological model. These results suggest that the proposed approach can assist clinicians in formulating personalized treatment strategies for patients with early stage thymoma, thereby optimizing therapeutic efficacy and improving clinical outcomes, and showing promise for broader clinical applications. Acknowledgements N\A. Abbreviations Computed tomography Volume of interest Area under the receiver operating characteristic curve Confidence interval Authors’ contributions K.Z. participated in research design, data collection, data analysis, and writing of the paper. Y.L. participated in research design and data analysis. H.X. participated in research design and data analysis. W.C. participated in research design and data analysis. J.J., L.S., Z.Z., and H.H. participated in research design and proofreading of the article. M.H. participated in research design. Q.L. participated in research design and data analysis. Z.X. participated in research design, data analysis, contribution of new analytic tools, writing, and study supervision. Funding This study was supported by the Research and Development Project on Medical Big Data and Artificial Intelligence of the PLA General Hospital (2019MBD-017) and Shaanxi Province Key Core Technology Tackling Project (2024SF-GJHX-27). These funding agencies were not involved in study design, analysis, or reporting. Data availability The data supporting the findings of this study were retrieved from the clinical database of PLA General Hospital, a secure institutional database housing de-identified clinical records of patients treated at the hospital. Declarations Ethics approval and consent to participate This retrospective study was conducted in accordance with the ethical standards of the Declaration of Helsinki and its later amendments. Ethical approval was obtained from the Institutional Review Board (IRB) of Hainan Hospital of Chinese PLA General Hospital (IRB approval number: S2025-11-01). Human Ethics and Consent to Participate declarations: The requirement for written informed consent was waived by the IRB due to the retrospective nature of the study. All patient data were anonymized and de-identified prior to analysis, and no interventions were performed that would affect clinical care or patient outcomes. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Kai Zhao and Yiming Liu contributed equally to this work and share first authorship. Contributor Information Qiang Lu, Email: luqiang@fmmu.edu.cn. Zhiqiang Xue, Email: xuezhiqiang301@126.com. References Associated Data Data Availability Statement The data supporting the findings of this study were retrieved from the clinical database of PLA General Hospital, a secure institutional database housing de-identified clinical records of patients treated at the hospital.