MRI-based radiomics analysis and machine learning in the differentiation of endometriomas from hemorrhagic ovarian...

Endometriomas and hemorrhagic ovarian cysts (HOCs) are common benign ovarian lesions in reproductive-aged women. Their management differs substantially, and classic MRI signs, such as T1 hyperintensity and T2 shading, might overlap or introduce uncertainty. Therefore, we investigated the diagnostic performance of MRI-based radiomics combined with machine learning (ML) algorithms for differentiating ovarian endometriomas from HOCs.

This prospective observational single-center study included patients with histopathologically confirmed endometriomas and HOCs. Lesions were manually segmented on T2-weighted MRI sequences, and radiomic features, including first-order, shape, texture, and gradient-based features, were extracted. Feature selection was performed using a nested cross-validation support vector machine recursive feature elimination (100-times per fold). Eight ML classifiers were trained and evaluated using nested ten-fold cross-validation for area under the curve (AUC), accuracy, sensitivity, specificity, and F1 score.

Out of 160 female patients (mean age: 31.3 ± 7.8 years), 89 (55.6%) were diagnosed with endometrioma. Five radiomic features were selected as the most discriminative predictors. Overall, the ensemble ML classifiers demonstrated better performance compared with the single-model classifiers (AUCs: 0.97-0.99 vs. 0.87-0.96). Extra Trees (ET) demonstrated the highest predictive performance (AUC = 0.99 [0.97-1.00], accuracy = 96%, sensitivity = 96%, specificity = 95%), which was significantly better than all of the single-model classifiers (DeLong's test P < 0.04).

MRI-based radiomics combined with ML algorithms demonstrates excellent performance in differentiating endometriomas from HOCs. This technique may hold clinical significance and, pending external validation, boost diagnostic confidence, decrease unnecessary surgeries, and aid in better patient management.