Machine learning-based differentiation of benign and malignant adrenal lesions using 18F-FDG PET/CT: a two-stage classification and SHAP interpretation study

BACKGROUND: Accurately distinguishing benign from malignant adrenal lesions remains a clinical challenge, especially in oncology patients with indeterminate imaging findings. This study aimed to develop and interpret machine learning (ML) models for classifying adrenal lesions based on 18 F-FDG PET/CT imaging and clinical parameters. METHODS: A retrospective cohort of 255 patients undergoing 18 F-FDG PET/CT was analyzed. Imaging features-including adrenal SUVmax, SUVpeak, tumor diameter, CT attenuation, and tumor-to-liver SUVmax ratio (T/L SUVmax)-along with clinical variables were extracted. Two classification tasks were constructed: (1) differentiation of benign and malignant adrenal lesions; and (2) subtyping of malignant lesions into lung cancer metastases or lymphoma. Seven ML models were trained and evaluated using 10-fold cross-validation. SHAP (SHapley Additive exPlanations) analysis was applied to elucidate feature contributions. RESULTS: For the benign/malignant classification, ensemble models (Random Forest, Bagging, XGBoost) achieved outstanding performance (AUC > 0.99), with Bagging yielding 100% recall. T/L SUVmax, adrenal SUVmax, and CT attenuation emerged as top predictors. In malignancy subtyping, the artificial neural network (ANN) attained the highest AUC (0.887) and F1-score (0.851). SHAP analysis highlighted distinct metabolic patterns, with lymphoma showing higher SUVmax and T/L ratios, and lung metastases associated with higher CT values. CONCLUSION: Machine learning models based on PET/CT-derived features enable highly accurate and interpretable classification of adrenal lesions. Integrating metabolic and anatomical parameters improves diagnostic precision, while SHAP analysis offers clinical transparency, supporting personalized decision-making in adrenal lesion management.