Differentiation of Fat-poor and Atypical Adrenal Adenomas from Metastases: MRI-based Radiomic, Radiologic, and Radiomic-radiologic Machine Learning Models

INTRODUCTION: The accurate differentiation of fat-poor and atypical adrenal adenomas from metastases remains a diagnostic challenge. This study aimed to evaluate the predictive value of MRI-based radiomic, radiologic, and combined radiomic-radiologic machine learning (ML) models. METHODS: This single-center retrospective study included 37 patients with 44 adrenal masses (19 adenomas; 25 metastases). Data were split into training and testing sets (2:1). To expand the training set, data augmentation was performed by multiple sampling (56 labeled slices from 30 masses). Radiomic features were extracted from T2-weighted (T2W), in-phase, out-of-phase, and apparent diffusion coefficient (ADC) sequences, while mass size, T2W signal intensity, heterogeneity, and signal drop were assessed as radiologic features. Dimension reduction was performed by reliability analysis and wrapper-based feature selection with five algorithms. A support vector machine (SVM) was used for classification, and performance was assessed using 10-fold cross-validation and unseen testing. Friedman test and post-hoc analyses compared bootstrapped unseen test area under the curve (AUCs). RESULTS: Only 12% of radiomic features demonstrated excellent reproducibility. A significant difference was observed among the three models, χ2(2)=779.5, p