Optimized machine learning models for personalized organ dose prediction in whole-body

This study investigates the use of machine learning (ML) models to improve the accuracy of organ dose prediction in whole-body Positron Emission Tomography/Computed Tomography (PET/CT) imaging using the 18F-Fluorodeoxyglucose (18F-FDG) radiotracer. Nine ML models: Gaussian Process Regression (GPR), Ensemble Methods (ENS), Stepwise Linear Regression (SLR), Efficient Linear Regression (ELR), Support Vector Machines (SVM), Efficient Linear Regression (ELR), Tree Regression (TR), Kernel Approximation Regression (KAR), and Neural Networks (NN), were evaluated and validated based on their ability to predict organ dose based on demographic parameters such as weight, height, age, and organ function. The results revealed that the best-performing model was highly dependent on the specific organ and patient gender. The optimized Gaussian Process Regression with Bayesian Optimization (GPR+BO) model showed strong performance, particularly for the male brain (R² = 0.91) and gallbladder (R² = 0.94). However, the Ensemble (ENS) model demonstrated superior or equivalent performance for other organs in male patients, including the kidneys (R² = 0.94) and lungs (R² = 0.95). For female patients, model performance also varied, with the NN+BO model showing the best predictive accuracy for the liver (R² = 0.93). Gender-based analysis indicated that female patients absorbed higher organ doses despite similar radiotracer levels, with notable differences observed in the brain (0.36 ± 0.05 mGy for females vs. 0.33 ± 0.05 mGy for males) and lungs (0.22 ± 0.03 mGy for females vs. 0.18 ± 0.03 mGy for males). These findings show the importance of adjusting radiotracer doses based on physiological variations to ensure accurate, patient-specific dose management. This study uniquely integrates GPR+BO, significantly enhancing predictive accuracy and potentially improving patient safety by enabling personalized radiation dose in clinical PET/CT procedures