Abstract
OBJECTIVES: To evaluate whether radiomic features extracted from (18)F-NaF PET/CT scans, analyzed using machine learning (ML) methods, can improve the differentiation between true metastatic bone lesions (TP) and false-positive benign uptake (FP), thereby enhancing the diagnostic utility of (18)F-NaF PET/CT. METHODS: This retrospective study included 62 patients with known primary malignancies who underwent (18)F-NaF PET/CT. Lesions were classified as TP or FP based on consensus interpretation including follow-up. Patients were randomly split into training (n=41) and validation (n=21) groups. Radiomic features were extracted from PET images using LIFEx software. Feature selection (ANOVA, RFE) and ML model training (SVM, Random Forest, XGBoost) were performed. Model performance was evaluated using accuracy, specificity, sensitivity, and AUC, initially with a train/validation split and subsequently with 5-fold cross-validation incorporating feature engineering and hyperparameter tuning. Feature importance was assessed using SHAP. RESULTS: Significant differences in SUV(max) (p=0.006) and SUV(mean) (p=0.034) were observed between TP and FP lesions. Initial validation showed XGBoost performed best (AUC=0.78). After optimization and 5-fold cross-validation on the combined dataset (n=62), the tuned XGBoost model achieved the highest performance (Mean Accuracy: 85.7% ±2.9%, Mean AUC: 0.86), outperforming Random Forest (AUC: 0.79) and SVM (AUC: 0.74). SHAP analysis identified SUV(max), SUV(mean), Voxel Volume Num, GLRLM RLNU, and Skew. CONCLUSION: Radiomics-based machine learning classifiers, particularly XGBoost, demonstrated strong performance in distinguishing true metastatic from false-positive benign lesions on (18)F-NaF PET/CT. Integrating radiomics and ML can potentially improve the diagnostic accuracy and robustness of (18)F-NaF PET/CT for assessing bone metastases. Further validation in larger cohorts is warranted.