An ensemble machine learning model based on magnetic resonance imaging features for diagnosing deep infiltrating endometriosis.
Authors
Affiliations (1)
Affiliations (1)
- Department of General Gynecology, General Hospital of Hunan University of Medicine, Huaihua, Hunan, China.
Abstract
To address the need for more objective and reproducible preoperative diagnosis of deep infiltrating endometriosis (DIE), this study aimed to develop and validate a weighted ensemble machine learning (ML) model using multimodal magnetic resonance imaging (MRI) and clinical features, and to compare its performance against radiologists. This retrospective study enrolled 330 patients (168 in DIE group, 162 in control group) who underwent MRI examinations and surgical-pathological confirmation during January 2020 to June 2025. Following feature selection, five ML models were constructed: logistic regression (LR), support vector machine (SVM), random forest (RF), XGBoost, and CatBoost. A weighted ensemble model was developed via soft voting, weighting each model's prediction by its AUC on validation set. Model performance and clinical net benefit were assessed using ROC curves, calibration curves, and decision curve analysis. Model interpretability was achieved using a weighted SHapley additive exPlanations (SHAP) method. The ensemble model's performance was compared to two radiologists' independent readings. The ensemble model demonstrated superior performance on the independent test set, achieving an AUC of 0.938 (95% CI: 0.883-0.979; sensitivity, 0.920; accuracy, 0.859), outperforming all individual models. Weighted SHAP analysis identified C-reactive protein, cancer antigen 125, T2-weighted signal intensity, and symptom duration as top contributors. The radiologists' average AUC was 0.818 (κ = 0.42), whereas the ensemble model achieved a higher AUC on the same dataset (P = 0.002). With AI assistance, the radiologists' accuracy increased to 0.920 and interpretation time decreased by approximately 30%. The weighted ensemble ML model, integrating multimodal MRI and clinical features, enables high-accuracy DIE identification with favorable stability and interpretability. This model holds significant potential as an effective clinical decision-support tool for radiologists.