Abstract
OBJECTS: Explore value of multiparametric MRI texture features in differentiating O-RADS MRI 4 ovarian-adnexal lesions. METHODS: A retrospective study was conducted, involving the clinical and MRI data of patients with ovarian - adnexal tumors that were surgically and pathologically confirmed and had undergone preoperative MRI examinations. Lesions with an O - RADS MRI score of 4 were included and then divided into benign and malignant groups (with borderline tumors categorized as malignant). Texture features of the entire lesion were extracted from the solid region of the tumor in T2WI, ADC, and CE - T1WI using the MATLAB image texture analysis toolbox. Feature selection was performed using a sparse representation - based feature selection method. Classification training and testing were carried out using five machine learning models: SVM, RF, DT, AB, and SR. Evaluated performance with ROC curve. Constructed MRI quantitative texture combination model, compared with MRI quantitative and texture feature models. Calculated AUC, etc. for each model, compared AUCs with Delong test, internally validated best model with ten - fold cross - validation. RESULTS: 58 patients with 60 lesions (25 benign, 35 malignant). Extracted 1,053 texture features. Selected 7 optimal texture features. AUCs of machine learning models for predicting lesion nature were 0.803, 0.884, 0.765, 0.849, 0.894 (SR best). AUC, etc. for MRI quantitative, texture feature, and combination models were as follows: MRI quantitative model: 0.811, 80.0%, 74.3%, 88.0%, 89.7%, 70.9%; texture feature model: 0.894, 78.4%, 86.7%, 66.7%, 78.8%, 77.8%; MRI quantitative texture combination model: 0.952, 92.2%, 96.7%, 85.7%, 90.6%, 94.7%. Differences in AUC between combination model and others were significant. Average AUC of combination model after cross - validation was 0.931. CONCLUSIONS: The texture features based on multiparametric MRI images can improve the diagnostic efficacy for distinguishing benign from malignant O-RADS MRI 4 lesions. The MRI quantitative texture combined model showed promising performance, highlighting its potential value in guiding clinical decision-making. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12880-025-02112-2.