Abstract
<title>Abstract</title> <p> <bold>Background</bold> Accurate preoperative differentiation between benign and malignant ovarian tumors remains a major clinical challenge because treatment strategy and prognosis depend substantially on tumor biology. Although magnetic resonance imaging (MRI) plays an important role in the characterization of adnexal lesions, conventional visual assessment may be limited by overlapping imaging features and observer-dependent interpretation. In addition, contrast-enhanced MRI protocols are not always feasible because of limited availability or contraindications to contrast administration. This study aimed to develop and internally validate a simplified predictive model based exclusively on routinely assessable non-contrast MRI features for preoperative differentiation of ovarian tumors. <bold>Methods</bold> This retrospective single-center study included 179 women with ovarian tumors who underwent preoperative pelvic MRI and subsequent surgical treatment between 2020 and 2025. Patients were divided into a training cohort (n = 126) for model development and an internal validation cohort (n = 53). Conventional MRI features were assessed using a structured checklist and transformed into categorical variables. Histopathological examination served as the reference standard. A multivariable logistic regression model was developed using selected non-collinear predictors, and diagnostic performance was evaluated using receiver operating characteristic (ROC) analysis. <bold>Results</bold> Multivariable logistic regression identified five independent MRI predictors of malignancy: lesion shape, wall thickness, papillary projections, regional lymph node enlargement, and pelvic fluid. In the training cohort, the model demonstrated excellent diagnostic performance with an area under the ROC curve (AUC) of 0.942 (95% CI: 0.885–0.976), sensitivity of 94.4%, specificity of 81.8%, and overall diagnostic accuracy of 87.9%. Internal validation demonstrated sensitivity of 70.6%, specificity of 89.5%, and overall accuracy of 77.4%, indicating acceptable reproducibility. <bold>Conclusions</bold> The proposed non-contrast MRI-based logistic regression model demonstrated effective differentiation between benign and malignant ovarian tumors using routinely assessable imaging features. The model may serve as an interpretable decision-support tool for preoperative risk stratification, particularly in settings where contrast-enhanced MRI is unavailable or contraindicated. Further external multicenter validation is required to confirm its generalizability and clinical utility. </p>