Not all endometrial tumors are cancerous. Benign conditions such as endometrial polyps, hyperplasia, and fibroids can present with similar symptoms and sometimes similar appearances on imaging to malignant endometrial cancer. Accurate preoperative differentiation between benign and malignant endometrial tumors is critical because it determines whether a patient needs aggressive cancer surgery or conservative management.
CT (computed tomography) scanning is widely available and commonly used in initial cancer staging, but its ability to characterize endometrial lesions - distinguish cancer from benign disease - has been limited when relying on standard radiological interpretation. Radiomics offers a way to extract far more information from CT images than visual inspection alone, by computing hundreds of mathematical features that describe texture, shape, and intensity patterns within and around the tumor.
This study developed and validated an explainable machine learning model based on CT radiomic features to differentiate malignant from benign endometrial tumors. A key innovation was incorporating SHAP (Shapley Additive Explanations) analysis to make the model's decisions transparent and interpretable - addressing one of the main barriers to clinical adoption of AI diagnostic tools.
The study enrolled 83 patients - 46 with malignant endometrial tumors and 37 with benign lesions - from two hospital centers. CT scans were segmented to delineate the tumor region of interest, and 1,132 radiomic features were extracted using PyRadiomics, an open-source software package for radiomics feature computation. This large initial feature set captures nearly every measurable aspect of the tumor's appearance on CT.
Patients were divided into a training set (59 patients) and a testing set (24 patients). The model was developed on the training set and then evaluated on the testing set to assess real-world performance on previously unseen cases. Six different machine learning models were trained and compared: Random Forest, Support Vector Machine, Logistic Regression, Gradient Boosting, K-Nearest Neighbors, and Decision Tree.
Feature selection was applied before training to reduce the 1,132 initial features to a smaller set of the most informative ones. Feature selection is important because machine learning models trained on too many features relative to the number of patients can memorize training data rather than learning generalizable patterns - a problem called overfitting. The final models used only the features that most reliably distinguished malignant from benign tumors.
Among all six machine learning algorithms tested, Random Forest achieved the best performance. In the training set, the Random Forest model achieved an AUROC (area under the receiver operating characteristic curve) of 1.00 - perfect discrimination. In the independent testing set, it maintained an AUROC of 0.96, with sensitivity of 100% (correctly identifying all malignant cases) and specificity of 92.31% (correctly classifying most benign cases as benign).
A sensitivity of 100% for malignant tumors is particularly important clinically: it means the model would not miss any cancer cases. A specificity of 92.31% means that fewer than 8% of benign tumors would be incorrectly flagged as malignant, which would be acceptable in a preoperative triage context where confirming a negative result can prevent unnecessary surgery.
Decision curve analysis was used to assess the clinical net benefit of the model across a range of decision thresholds. This analysis showed that the Random Forest model provided higher net benefit compared with both treating all patients (over-treating benign cases) and treating no patients (missing malignant cases). This confirms that the model adds real value to clinical decision-making beyond simple yes/no strategies.
SHAP (Shapley Additive Explanations) is a mathematical framework borrowed from game theory that quantifies how much each feature contributed to a specific prediction. For each patient, SHAP values show which radiomic features pushed the prediction toward malignant or benign, and by how much. This transforms a black-box Random Forest into a transparent, auditable decision support tool.
SHAP analysis identified the top 20 most influential radiomic features. Among these, 60% were texture features - measurements of spatial patterns of pixel intensity that reflect heterogeneity and irregularity within the tumor. The remaining 40% were first-order statistical features - basic statistical summaries of the intensity distribution such as mean, variance, and skewness. Importantly, 90% of the top features came from transformed images (such as wavelet or Laplacian of Gaussian filtered images) rather than the original CT image, suggesting that mathematical transformations reveal patterns invisible in standard images.
Visualization of radiomic feature maps on actual CT images demonstrated that malignant tumors showed higher spatial heterogeneity - more irregular patterns of tissue density - compared with benign lesions. This heterogeneity, captured mathematically by the radiomic features, corresponds to the biological reality that cancer tissue contains irregular mixtures of tumor cells, necrosis, fibrous stroma, and blood vessels, while benign tissue is more homogeneous and organized.
The combination of high diagnostic accuracy and explainability through SHAP makes this model particularly suitable for clinical translation. Clinicians are more likely to trust and use an AI tool when they can understand why it made a specific recommendation, rather than accepting a black-box output. The SHAP feature maps showing heterogeneity patterns on the actual patient CT scan provide an intuitive visual explanation that radiologists and oncologists can evaluate.
The two-center design provides initial evidence that the model generalizes beyond a single institution. Both training and testing were performed across patients from two different hospitals, meaning the model was exposed to variation in CT scanner types, imaging protocols, and patient populations. This is an important step toward demonstrating clinical generalizability, though further validation at additional centers would strengthen the evidence.
Compared with MRI-based radiomics (which requires more expensive, less widely available equipment), CT-based approaches offer the advantage of broader accessibility. In healthcare systems where MRI availability is limited, a validated CT radiomic tool could extend the benefits of precision preoperative endometrial tumor characterization to a wider patient population. Integration of this approach into routine CT reporting workflows could eventually provide radiologists with an automated second opinion to flag potentially malignant cases for prioritized clinical attention.