Preoperative prediction of cervical cancer recurrence using explainable MRI radiomics: a SHAP-Guided machine learning...

To develop and validate a machine learning model that integrated MRI radiomics features and clinical factors for preoperative prediction of postoperative recurrence in cervical cancer.

This retrospective study included 268 patients with pathologically confirmed cervical cancer (training cohort: n = 185; validation cohort: n = 83). A total of 124 radiomics features were extracted from T2-weighted images, with 12 optimal features were selected through reproducibility analysis (ICC > 0.75), minimum redundancy maximum relevance (mRMR), and least absolute shrinkage and selection operator (LASSO) regression. Four machine learning (ML) models (logistic regression [LR], naïve Bayes [NB], gradient boosting machine [GBM], random forest [RF]) were trained and evaluated for their discrimination performance (area under the curve [AUC], sensitivity, specificity), calibration accuracy, and clinical utility (via decision curve analysis [DCA]). SHapley Additive exPlanations (SHAP) were employed to elucidate the importance of each feature.

The LR model demonstrated the highest performance in the validation cohort (AUC = 0.818, sensitivity = 0.629, specificity = 0.806), which significantly outperformed clinical factors alone (AUC = 0.681, P = 0.005). SHAP analysis identified tumor heterogeneity features such as original_firstorder_Minimum and GLCM-correlation, as the top predictors. The combined radiomics-clinical model further improved AUC to 0.844. DCA confirmed a net benefit across clinically relevant risk thresholds.

An interpretable ML radiomics model that leveraged preoperative MRI and clinical data showed potential to stratify the risk of recurrence in cervical cancer, thereby offering potential to guide personalized decisions regarding surgical and adjuvant therapies.