Differential performance of statistical versus machine learning methods in partial volume correction in oncologic F18-FDG PET/CT scanning.

Partial volume effects (PVE) in positron emission tomography (PET) imaging introduce quantification inaccuracies, particularly in small or heterogeneous lesions, necessitating precise correction methodologies. This study systematically evaluates the performance of statistical versus machine learning approaches in partial volume correction (PVC) across multiple PET reconstruction algorithms, including TrueX, TrueX + TOF, and Iterative + TOF. Phantom experiments were conducted across a broad range of lesion-to-background contrast ratios and lesion sizes to derive and validate exponential recovery coefficient (RC) fitting models. Additionally, advanced machine learning algorithms, including Random Forest, Support Vector Regression, and Gradient Boosting, were implemented to enhance PVC accuracy. The results demonstrate that Iterative + TOF reconstruction yielded the most consistent RC estimates, while machine learning-based PVC significantly outperformed traditional exponential fitting in minimizing residual errors. Among the machine learning models, Random Forest exhibited the lowest root mean square error (RMSE) and highest coefficient of determination (R²), indicating superior predictive accuracy and robustness. These findings underscore the potential of machine learning-driven PVC methodologies for standardizing PET quantification, thereby improving lesion characterization, therapy response assessment, and multi-center data harmonization.