Introduction: Radiomics has not yet been widely adopted in clinical practice due to the limited understanding of the biological significance of the measured texture features and the lack of sufficient clinical evidence supporting its superiority over conventional imaging analysis methods. The present study was conducted to evaluate radiomics-based approaches for quantifying imaging features of endometriomas and hemorrhagic cysts, as well as to investigate their differentiation.

Material and Methods: This diagnostic study evaluated the performance of radiomics-based texture analysis in distinguishing ovarian endometriomas from hemorrhagic ovarian cysts using pelvic MRI data from 150 patients. Segmentation and feature extraction were performed using 3D Slicer, followed by statistical analysis of 107 radiomic features. A nested cross-validation strategy with consensus Lasso-based feature selection was applied to train SVC and KNN models. Model performance was assessed via ROC analysis, yielding metrics such as AUC, accuracy, sensitivity, specificity, and F1-score.

Results: A total of 107 features per dataset were analyzed using SVC and KNN classifiers, with feature selection guided by SHAP values and statistical testing. Shape-based features such as Sphericity, and first-order statistical features such as Skewness, were among the most important predictors. Sphericity, Skewness, and textural features such as Coarseness were frequently selected and showed statistically significant differences in T2 and T1 Fat-Sat +T2 datasets, while the T1 Fat-Sat dataset exhibited weaker discriminative power. The highest classification performance was achieved with the T1 Fat-Sat +T2 dataset, followed by T2, and then T1 Fat-Sat. Similar trends were observed with the KNN classifier.

Conclusion: The study shows that T2-weighted MRI performs well on its own, especially when data is limited, while T1 Fat-Sat alone is unreliable. Combining T1 Fat-Sat and T2 offers the best results for radiomics-based classification of some of ovarian lesions. Among classifiers, SVC had higher specificity and accuracy, whereas KNN showed better sensitivity.