Applying deep learning-based ensemble model to [ 18 F]-FDG-PET-radiomic features for differentiating benign from malignant parotid gland diseases

To develop and identify machine learning (ML) models using pretreatment 2-deoxy-2-[ F]fluoro-D-glucose ([ F]-FDG)-positron emission tomography (PET)-based radiomic features to differentiate benign from malignant parotid gland diseases (PGDs). This retrospective study included 62 patients with 63 PGDs who underwent pretreatment [ F]-FDG-PET/computed tomography (CT). The lesions were assigned to the training (n = 44) and testing (n = 19) cohorts. In total, 49 [ F]-FDG-PET-based radiomic features were utilized to differentiate benign from malignant PGDs using five different conventional ML algorithmic models (random forest, neural network, k-nearest neighbors, logistic regression, and support vector machine) and the deep learning (DL)-based ensemble ML model. In the training cohort, each conventional ML model was constructed using the five most important features selected by the recursive feature elimination method with the tenfold cross-validation and synthetic minority oversampling technique. The DL-based ensemble ML model was constructed using the five most important features of the bagging and multilayer stacking methods. The area under the receiver operating characteristic curves (AUCs) and accuracies were used to compare predictive performances. In total, 24 benign and 39 malignant PGDs were identified. Metabolic tumor volume and four GLSZM features (GLSZM_ZSE, GLSZM_SZE, GLSZM_GLNU, and GLSZM_ZSNU) were the five most important radiomic features. All five features except GLSZM_SZE were significantly higher in malignant PGDs than in benign ones (each p