CT imaging models differentiate benign from malignant orbital tumors in retrospective cohort of 145 patients.

ObjectiveTo evaluate the diagnostic performance of deep learning−based radiomics (DL) and hand−crafted radiomics (HCR) in differentiating benign from malignant orbital tumors.MethodsA retrospective analysis was performed on CT data from 145 patients (48 benign, 97 malignant) diagnosed between December 2014 and March 2024. Two radiologists independently assessed conventional CT semantic features (e.g., lesion location, margin definition, internal density homogeneity, calcification, necrosis, and enhancement pattern). Deep transfer learning (DTL) extracted DL features, while traditional methods were used to obtain HCR features. Feature fusion, selection, and modeling were performed using the Least Absolute Shrinkage and Selection Operator (LASSO). Pathological diagnosis served as the gold standard. Model performance was evaluated using receiver operating characteristic (ROC) curves. A nomogram integrating clinical data and significant semantic features was constructed for visualization. The DeLong test and decision curve analysis (DCA) assessed model effectiveness.ResultsMultivariate analysis confirmed that homogeneous enhancement and ill−defined/infiltrative margins were independent CT features differentiating benign from malignant tumors. A total of 14 HCR and 30 DL features were extracted; 36 features were retained after fusion. The HCR, DL, fused, and nomogram models achieved AUCs of (0.859/0.816), (0.957/0.826), (0.986/0.811), and (0.975/0.837) in the training and test cohorts, respectively. The DeLong test showed no significant difference between the fused model and the nomogram in either cohort (P = 0.090 and P = 0.198), whereas differences for other model pairs were significant (P