Quantitative Assessment in Dynamic PET/CT for Hepatocellular Carcinoma through Bayesian Prior information-based Actor-Critic algorithm

Dynamic positron emission tomography/computed tomography (PET/CT) is a promising medical imaging technique that enables the visualization of biological processes via tracer activity. Analyzing the time-activity curves (TACs) of PET using pharmacokinetic models allows for the extraction of quantifiable physiological information relevant to hepatocellular carcinoma (HCC). Estimation of the pharmacokinetic model parameters requires advanced algorithms. The challenge lies in achieving accurate parameter estimation and obtaining parameter distributions that align with physiological characteristics. This paper introduces a reinforcement learning-based Actor-Critic (AC) algorithm and proposes a Bayesian prior information-based Actor-Critic (BPAC) algorithm. A dynamic parameter estimation environment for agent interaction was established, incorporating Bayesian optimization. The preceding information obtained through Bayesian optimization was used to design the reward and penalty signals in the interaction environment, and sampled as the initial value for the agent to estimate the parameters of 24 real HCC cases. Receiver operating characteristic (ROC) analyses were performed, using P-values and areas under the curve (AUC) to evaluate diagnostic performance between normal and tumor groups. Compared to the AC, the BPAC demonstrates statistically significant differences in K1, k2, k3, fa, and vb (all P < 0.001) and aligns more closely with correct physiological characteristics in K1, k2, k3, k4, and fa. In ROC analyses, five parameters derived from the proposed method achieved higher AUC values than the AC. Results indicate that BPAC can more effectively quantitatively evaluate the physiological characteristics of HCC, such as perfusion and metabolism, resulting in superior diagnostic performance.