Hepatocellular carcinoma ¹⁸F-FDG PET/CT kinetic parameter estimation based on the advantage actor-critic algorithm.

Area of Science:

• Nuclear Medicine
• Oncology
• Artificial Intelligence

Background:

• Dynamic 18F-fluorodeoxyglucose (18F-FDG) PET/CT kinetic parameters aid in characterizing hepatocellular carcinoma (HCC).
• Deep reinforcement learning (DRL) shows potential for enhancing kinetic parameter estimation in medical imaging.

Purpose of the Study:

• To preliminarily assess the Advantage Actor-Critic (A2C) algorithm, a DRL method, for estimating 18F-FDG PET/CT kinetic parameters in HCC patients.
• To evaluate the A2C algorithm's ability to optimize parameter estimation using neural networks.

Main Methods:

• Prospective collection of 18F-FDG PET data from 14 liver tissues and 17 HCC tumors using an abbreviated acquisition protocol.
• Application of the A2C algorithm with a reversible double-input, three-compartment model for kinetic parameter estimation.
• Comparison of A2C results with the conventional nonlinear least squares (NLLS) algorithm, evaluating fitting errors via root-mean-square errors (RMSEs) of time activity curves (TACs).

Main Results:

• Significant differences in kinetic parameters (K1, k2, k3, k4, fa, vb) were detected between HCC and normal liver tissues using A2C (p < 0.05).
• A2C demonstrated superior diagnostic performance over NLLS for k3 and vb (p < 0.05).
• A2C yielded smaller fitting errors for both normal liver tissue (0.62 ± 0.24 vs. 1.04 ± 1.00) and HCC tissue (1.40 ± 0.42 vs. 1.51 ± 0.97) compared to NLLS.

Conclusions:

• The A2C algorithm provides more precise estimation of 18F-FDG kinetic parameters for HCC tumors compared to the conventional NLLS method.
• A2C achieves better time activity curve fitting with a lower RMSE when using a reversible double-input, three-compartment model.
• This DRL approach enhances the characterization of HCC using 18F-FDG PET/CT.