Comparison of Deep Learning and Particle Smoother Expectation Maximization Methods for Estimation of Myocardial Perfusion PET Kinetic Parameters

Background: Positron emission tomography (PET) is widely used for studying dynamic processes, such as myocardial perfusion, by acquiring data over time frames. Kinetic modeling in PET allows for the estimation of physiological parameters, offering insights into disease characterization. Conventional approaches have notable limitations; for example, graphical methods may reduce accuracy due to linearization, while non-linear least squares (NLLS) methods may converge to local minima. Purpose: This study aims to develop and validate two novel methods for PET kinetic analysis of 82Rb: a particle smoother-based algorithm within an Expectation-Maximization (EM) framework and a convolutional neural network (CNN) approach. Methods: The proposed methods were applied to simulated 82Rb dynamic PET myocardial perfusion studies. Their performance was compared to conventional NLLS methods and a Kalman filter-based Expectation-Maximization (KEM) algorithm. Results: The success rates for parameters F, k3, and k4 were 46.0%, 67.5%, and 54.0% for the particle smoother with EM (PSEM) and 86.5%, 83.0%, and 79.5% for the CNN model, respectively, outperforming the NLLS method. Conclusions: The CNN and PSEM methods showed promising improvements over traditional methods in estimating kinetic parameters in dynamic PET studies, suggesting their potential for enhanced accuracy in disease characterization.
Comment: 29 pages, 7 figures, submitted to Medical Physics