Abstract
Purpose: Accurate preoperative risk stratification in endometrial cancer (EC) is essential for guiding surgical and therapeutic decisions. This study aimed to evaluate the discriminative performance of [18F]-FDG PET/CT-derived radiomic features combined with machine learning models for differentiating low-risk (LRH-EC) and high-risk histology (HRH-EC) subtypes. Methods: A total of 159 patients with histopathologically confirmed EC who underwent preoperative [18F]-FDG PET/CT were retrospectively analyzed. Radiomic features were extracted using LIFEx version 7.4.0 software following IBSI guidelines. After FDR correction and Pearson correlation-based redundancy reduction (|r| > 0.80), 16 radiomic features were retained for modeling. Three feature configurations (Conventional PET parameters, Radiomics16, and Combined) were evaluated. Machine learning models were developed using stratified 5-fold cross-validation. Model performance was assessed using AUC, accuracy, sensitivity, specificity, F1-score, Wilson confidence intervals, DeLong's test, and McNemar's test. Results: Artificial Neural Network (ANN) (AUC = 0.709) and Random Forest (RF) (AUC = 0.686) achieved the highest discriminative performance within the Radiomics16 feature set. No statistically significant superiority between algorithms or feature configurations was observed by DeLong analysis. However, McNemar's test demonstrated significant patient-level classification differences for the Combined ANN model (p < 0.001). NGTDM_Coarseness and SUVmin emerged as the most influential features, reflecting tumor heterogeneity and metabolic activity. Conclusions: [18F]-FDG PET/CT-based radiomics combined with machine learning provides moderate yet consistent discrimination between LRH-EC and HRH-EC. While external validation is required, this approach may support noninvasive preoperative risk stratification in endometrial cancer.