Comparison of 2D and 3D radiomics features with conventional features based on contrast-enhanced CT images for preoperative prediction the risk of thymic epithelial tumors

Thymic epithelial tumors (TETs) are the most prevalent neoplasms in the anterior mediastinum. They generally exhibit low malignancy in both histological and clinical presentations.¹ The WHO classification of TETs, based on lymphocyte-to-epithelial cell ratio and epithelial cell morphology, is widely adopted. According to the 2015 criteria, TETs are categorized into thymomas (with six subtypes) and thymic carcinomas (TCs).^2,3 This classification serves as an independent prognostic factor and is simplified into low-risk (Type A, AB, B1) and high-risk (Type B2, B3, TCs) categories, influencing patient outcomes.⁴ The WHO classification reflects the tumor’s clinical and functional features, aiding preoperative diagnosis and treatment planning.⁵ Accurate, noninvasive identification and subgroup classification of TETs prior to treatment hold significant clinical value.

CT imaging is the primary diagnostic tool for TETs, revealing a wide range of biological and morphological features.^6,7 Studies have reported specific CT characteristics of TETs.^8,9 While chest contrast-enhanced computed tomography (CECT) provides general morphological parameters, distin-guishing between histological subgroups remains challenging due to significant overlap. Radiomics, leveraging radiomic signatures, extracts diverse, high-throughput imaging features, transforming medical images into mineable data.^10,11 Radiomic features can predict disease, cancer, metastasis, and prognosis.¹² Previous studies have used 2D radiomics to determine thymoma risk levels.^13,14 Wang et al. differentiated low-risk from high-risk thymomas and early from advanced tumors using contrast-enhanced CECT and non-enhanced CT15, but their study had a small sample size and the sample size was imbalanced.

Despite several CT-based radiomics analyses being used to classify TET risk, most studies rely on 2D imaging features. Further research is needed to explore 3D radiomic signatures, which may provide more comprehensive and accurate information for TET diagnosis and risk assessment. The available models still require refinement and validation.

This study aimed to develop conventional and 2D/3D radiomics signatures using the conventional and texture features extracted from CECT images for preoperative prediction of the WHO’s TET risk classification.

This study evaluates three models using conventional and radiomic signatures, revealing that radiomic features from 2D and 3D imaging serve as noninvasive biomarkers for thymic epithelial tumors (TETs). These features enable preoperative risk stratification, guiding the surgical approach and resection extent. Importantly, risk stratification aids in personalized treatment planning for high-risk patients, often needing adjuvant therapy. A Phase II studies show that preoperative chemotherapy and radiotherapy enhance R0 resection rates.¹⁶ Thus, this study highlights the significance of radiomic signatures in preoperative risk assessment and treatment planning for TETs.

In this study, three established models obtained using the conventional and radiomics signatures were examined. The results indicated that the radiomics signatures based on conventional, 2D and 3D images could be used as noninvasive biomarkers to differentiate high-risk from low-risk TETs. Specifically, these models can provide preoperative risk stratification which informs the choice of surgical approach (minimally invasive surgery or thoracotomy) and helps define the extent of resection. By accurately predicting the risk, this model could aid in minimizing unnecessary tissue damage and reducing complications, potentially improving surgical outcomes and postoperative recovery. Since high-risk TET patients generally have poorer prognosis and often require adjuvant therapy, this stratification can guide treatment planning and improve patient outcomes by facilitating more personalized interventions.

Radiomics has shown significant potential in personalized medicine by providing non-invasive insights into tumor characteristics.^10,12 It has been applied in various diseases, such as predicting thyroid cancer nodules, differentiating the benign and malignant nature of pulmonary nodules, and supporting clinical decision-making for liver cancer patients. Radiomics studies using different imaging modalities (including ultrasound, MRI, and CT) have increasingly played a crucial role in clinical decision-making, providing new technological support for prediction, diagnosis, and prognosis.^17,18,19

Radiomics can be used to determine the TET classification according to WHO criteria. Previous studies have applied radiomics to classify the risk levels of thymomas and predict their invasiveness^15,20,21, but the available models still need refinement and validation. Therefore, this study aimed to use radiomics based on CECT images to develop conventional and radiomics signatures, which may help the preoperative prediction of the risk of TET classified by WHO guideline.

We used a simplified classification that defines Type A, AB, and B1 thymomas as low-risk and Type B2 and B3 thymomas and TC as high-risk, which is related to patient outcome. High-risk thymic epithelial tumors reportedly have a much poorer prognosis compared with low-risk thymic epithelial tumors, the former need adjuvant therapy to improve survival. With complete resection of the tumor, the prevalence of recurrence of LTET seems low.

Our 3D radiomics model performed better in terms of AUC, sensitivity, and specificity compared to prior models reported in the literature. For instance, Wang et al. reported an AUC of 0.801, with a sensitivity of 75% and specificity of 77%, while Sui et al. found sensitivities ranging from 71% to 74% and specificities from 65% to 74% for differentiating low-risk and high-risk TETs. The improved performance of our model may be attributed to the enhanced information captured by 3D imaging features, which provide a more comprehensive representation of tumor morphology and texture heterogeneity than 2D features. This suggests that using 3D data could offer a substantial advantage in capturing the complex structural characteristics of TETs, which is consistent with findings from previous studies on the added value of 3D features. However, it should also be noted that 3D feature extraction and analysis increase computational complexity, highlighting a potential trade-off between accuracy and practicality. In this study, the images with 5 mm thickness in CECT were used for radiomics analysis, which not only saves time but also yields satisfactory results.

Due to the small sample size, the fluctuation of individual samples affected the results greatly. The Hosemer-Lemeshow test was greater than 0.05. Although there were small fluctuations in the calibration map images of the test set, it still showed good consistency between predicted probability and true possibility without statistical deviation.

In the clinical usefulness, the 3D radiomics model was valid from 25% to 60% and from 70% to 95% for the training set, and valid from 25% to 57% and from 75% to 90% for the test set. The DCA analysis showed the 3D model to be more effective and reliable than the other two models to determine the clinical management.

Conventional features such as contour, capsule, septum and homogeneous enhancement can be used to distinguish low-risk and high-risk TETs.^7–9 Our study also compared the parameters obtained through the conventional CT features rather than software calculations, which also has some clinical significance for distinguishing high-risk from low-risk groups.

Except CT, magnetic resonance imaging (MRI) could also be used to differentiate the grading of TETs according to WHO criteria.^22,23 Nevertheless, the use and combination of multiple imaging modalities could maximize the diagnostic value, provided that all modalities are available at a given hospital, but this is not always the case, especially in developing countries like China.

This study has several limitations. First, it is a single-center, small-sample, retrospective study, which may limit the generalizability of the findings and lead to overfitting. The small sample size also means that individual sample fluctuations could have a more significant impact on the model’s performance, thus affecting its robustness. To address this, we plan to expand the sample size and conduct a multi-center study for further investigation. Second, the absence of Masaoka staging information limits the model’s applicability in distinguishing between different stages of thymic epithelial tumors (TETs), which is crucial for clinical decision-making. In future studies, we will incorporate this information. Third, variability in CT scan parameters and reconstruction techniques could affect the consistency and reproducibility of radiomics features, further impacting model performance. Standardizing imaging acquisition and processing methods will be essential in future studies to enhance both model robustness and clinical applicability. Additionally, AI-based automated recognition methods will be employed for more in-depth tumor analysis. Although the model demonstrates excellent performance in distinguishing risk groups, additional factors, such as tumor heterogeneity or genetic information, may further enhance predictive accuracy²⁴.

In conclusion, radiomics signatures based on conventional, 2D and 3D images could be used as noninvasive biomarkers to differentiate high-risk TETs from low-risk TETs. The 3D radiomics signature can provide complementary diagnostic information and as the most useful method to determine the clinical management.