Comparison of Deep Learning and Particle Smoother EM Methods for Estimation of Rb-82 Myocardial Perfusion PET Kinetic Parameters

Abstract

Positron emission tomography (PET) enables quantification of dynamic physiological processes through time-resolved imaging. In Rb-82 myocardial perfusion PET, kinetic compartment modeling is used to estimate physiological parameters and derive myocardial blood flow. However, conventional nonlinear least squares (NLLS) estimation is sensitive to model misspecification when not all parameters can be reliably estimated and must instead be fixed or initialized using population averages, which can degrade accuracy. This work develops and evaluates two alternative kinetic analysis approaches for Rb-82 PET: a particle smoother-based Expectation-Maximization method (PSEM) and a convolutional neural network (CNN). Both methods were evaluated using simulated Rb-82 dynamic myocardial perfusion studies and compared against NLLS and a Kalman smoother-based Expectation-Maximization (KEM) algorithm across multiple frame durations and noise levels. Across 2-10 s frames, the CNN achieved the lowest relative errors for all parameters (F: 8.78-4.98%, k3: 26.05-25.50%, k4: 34.34-22.76%), significantly outperforming NLLS, KEM, and PSEM (Holm-adjusted p < 1e-15 at 1.0x noise, 2 s frames), although performance degraded under out-of-distribution input-function conditions. Overall, the CNN provided the most accurate and robust in-distribution kinetic parameter estimates across frame durations. In contrast, PSEM exhibited parameter-dependent behavior, improving k3 estimation while underperforming for F, suggesting that further methodological refinement is needed.