A Gate Recurrent Unit Based Convolutional Neural Network Approach for Pharmacokinetic Parameter Reconstruction

To address the challenges of selecting regularization parameters and the high computational cost associated with traditional algorithms in reconstructing pharmacokinetic parameters of fluorescent probes, this paper introduces a Gate Recurrent Unit-based Convolutional Neural Network (GRU-CNN). Initially, a simulation program based on the two-compartment model of diffusion equation and pharma-cokinetics is employed to generate time series data of surface measurement signal distributions, which serve as the training set. A convolutional encoder with a dual-pool sampling strategy is utilized to extract crucial features from 2D image sequences. Secondly, a decoder composed of gated recurrent units processes the feature sequences, capturing temporal information and correlating features across frames to produce images that accurately match the input fluorescence probe parameters. Finally, the GRU-CNN is applied to reconstruct metabolic parameter images from the time series data of surface fluorescence distribution in the testing set. Experiment results demonstrate that GRU-CNN reconstructs pharmacokinetic parameters of fluorescent probes in various organs more rapidly and accurately than alternative methods, effectively capturing the pharmacokinetic processes of these probes.