Abstract / Summary
Accurate differentiation between malignant and benign orbital tumors is essential for clinical decision-making and treatment planning. MRI-based radiomics, which extracts high-dimensional quantitative features from medical images, has emerged as a promising noninvasive approach for tumor characterization. This study aims to develop and validate a radiomics-based nomogram derived from MRI features to distinguish malignant orbital tumors from benign ones. By leveraging data from two independent centers, we aimed to provide an initial assessment of the generalizability of the radiomic signature, while acknowledging that the nomogram itself requires further external validation. A total of 152 patients from two institutions were retrospectively included. Patients from Institution 1 were randomly assigned to a training cohort ( n = 94) and a test cohort ( n = 23), while patients from Institution 2 served as an external validation cohort ( n = 35).Radiomic features were extracted from T2-weighted fat-suppressed (T2WI-FS) MRI images, and 17 features were subsequently selected using the Spearman rank correlation coefficient and the least absolute shrinkage and selection operator (LASSO) regression. Four machine learning classifiers, Logistic Regression (LR), K-Nearest Neighbors (KNN), Extremely Randomized Trees (Extra Trees), and Multilayer Perceptron (MLP) were implemented to model the selected features and identify the best-performing algorithm.Significant clinical factors were identified through multivariate logistic regression analysis. A radiomics nomogram was then constructed by integrating radiomic features with clinical variables. Model performance was assessed using the area under the curve (AUC), accuracy, sensitivity, and specificity, while clinical utility was evaluated using decision curve analysis (DCA). Among the four machine learning models, the MLP classifier demonstrated the best predictive performance and exhibited strong robustness across datasets. Consequently, the MLP-based model was selected for nomogram construction. The nomogram model achieved AUC values of 0.965 in the training cohort and 0.905 in the internal test cohort. In the independent external validation cohort, the radiomic signature (MLP model) achieved an AUC of 0.783, suggesting acceptable generalizability of the radiomic features, although the full nomogram was not externally validated. The nomogram model significantly outperformed the clinical model (DeLong test, p = 9.724 × 10⁻⁸ for the training cohort and p = 2.969 × 10⁻³ for the internal test cohort), while no statistically significant difference was observed when compared with the radiomics‑only model (DeLong test, p = 0.608 and p = 0.768, respectively). Our findings suggest that the MRI‑based radiomics nomogram is a promising internal‑tested tool for differentiating malignant from benign orbital tumors. However, its generalizability remains to be established through further external validation. This approach may assist clinicians in tumor diagnosis and management, potentially improving patient outcomes. Not applicable.