A multimodal MRI radiomics model for distinguishing borderline from malignant ovarian epithelial tumors.

BACKGROUND

To develop and validate a multimodal MRI radiomics machine learning model for differentiating borderline epithelial ovarian tumors (BEOTs) from malignant epithelial ovarian tumors (MEOTs).

METHODS

A total of 147 patients (72 with BEOTs and 75 with MEOTs) were retrospectively enrolled and randomly divided into training and test cohorts at a ratio of 7:3. Multivariate logistic regression identified independent clinical predictors. A radiomic model was built using multimodal MRI features, and a radiomic score (Rad‑score) was calculated. A combined model integrating clinical predictors with Rad‑score was developed. ROC curve analysis and decision curve analysis (DCA) assessed model performance. The SHAP method was used to interpret the radiomic model.

RESULTS

Human epididymis protein 4 (HE4) was an independent risk factor for MEOTs and was used to construct the clinical model. The radiomic model comprised 17 features, with Rad‑scores differing significantly between BEOT and MEOT patients (cutoff = 0.564). A combined model integrated HE4 and Rad-score. Both the radiomic and combined models outperformed the clinical model in the training cohort (AUC: 0.968/0.970 vs. 0.858) and the test cohort (0.914/0.920 vs. 0.711; all p < 0.05), with no significant difference between them. DCA confirmed their superior net benefit. SHAP analysis revealed core features and their biological implications.

CONCLUSIONS

The multimodal MRI radiomics model interpreted by SHAP offers good diagnostic performance for differentiating BEOTs from MEOTs. As a non‑invasive, interpretable tool, it holds promise for clinical translation by assisting individualized treatment decisions and reducing unnecessary surgery.