Integrating the O-RADS MRI Score with Machine Learning: Incremental Value for Predicting Malignancy in Ovarian Tumors

This study aimed to evaluate the incremental diagnostic value of integrating the Ovarian-Adnexal Reporting and Data System (O-RADS) magnetic resonance imaging (MRI) score with conventional clinical and imaging features within a machine learning (ML) framework for discriminating ovarian malignancies.

In this retrospective multicenter study of 257 patients (296 lesions), we developed and internally tested models using a development cohort. An independent external cohort was used for validation. Semantic MRI features and clinical data were extracted. After two-stage feature selection, multiple ML algorithms were used to build a traditional model (Model 1) and a combined model incorporating the O-RADS score (Model 2). Performance was evaluated using the area under the receiver operating characteristic curve (AUC).

The O-RADS score was strongly correlated with key features of malignancy (Spearman's ρ range: 0.60-0.79, all p < 0.05), including solid enhancement and elevated serum CA125. Model performance was enhanced by its continuous inclusion. An AUC of 0.942 was realized by the integrated model (Model 2) in the internal test set, greatly bigger than the traditional model's AUC of 0.799 (p = 0.005). AUC increased from 0.791 to 0.909 in external validation. According to decision curve analysis, a larger net clinical advantage for the integrated model was shown, and the ORADS score was identified as the most influential predictor through SHapley Additive exPlanations analysis.

According to this research, by including the O-RADS MRI score in a machine learning structure, great incremental diagnostic value is provided beyond traditional characteristics alone. Excellent discriminative efficiency is realized by the integrated model, increasing the AUC from 0.799 to 0.942 in internal testing (Δ = 0.143) and from 0.791 to 0.909 in external validation (Δ = 0.118). Through SHAP analysis, O-RADS is identified as the main predictor operating with core imaging characteristics in a synergistic way, realizing transparent and radiologically based decision-making. Adopting standardized semantic reporting, routinely gathered clinical information is utilized by this method without extra examination time or economic burden. These outcomes build a strong paradigm for integrating well-organized radiology lexicons with computational intelligence, guiding personalized treatment by providing precise preoperative risk stratification and stopping both overtreatment of benign lesions and delayed administration of malignancies.

By combining the O-RADS MRI score in an ML framework, great incremental diagnostic value is provided, delivering a precise and interpretable tool for preoperative risk stratification of ovarian tumors.