Background: Nasopharyngeal lesions encompass a broad spectrum of benign and malignant diseases with overlapping imaging features, posing diagnostic challenges. Early and accurate identification is essential to guide appropriate management and avoid unnecessary invasive procedures. In this study, radiomic-clinical diagnostic models for the identification of five nasopharyngeal lesion types via the altruistic seagull optimization algorithm (AltSOA) were developed and evaluated. Materials and methods: In total, 350 patients from three medical centers were retrospectively included. Radiomics features were extracted from magnetic resonance imaging (MRI) sequences T1-weighted images (T1WI), T2-weighted images, and contrast-enhanced (CE)-T1WI. The AltSOA with an embedded support vector machine classifier was used for feature selection. Diagnostic models (radiomic and radiomic-clinical) were constructed based on combinations of multimodal MRI sequences. Results: The radiomic models achieved large areas under the curve (AUCs) but limited performance according to other metrics. The radiomic-clinical models yielded comparable AUCs and improved precision and recall values and F1-scores. The T1WI + CE-T1WI radiomic-clinical model achieved the best performance (internal testing set: AUC = 0.909 95% CI, 0.868–0.951; external testing set: AUC = 0.825 95% CI, 0.736–0.900). SHAP analysis revealed the main contributing features, and the directions and magnitudes of their impacts on the predicted diagnostic classification. Conclusion: These findings potentially reduce unnecessary biopsy performance and advance personalized medicine.
Wang et al. (2026) studied this question.