PKiNN: a PET kinetics-informed neural network for dynamic PET kinetic modeling.

Objective.Voxel-wise PET kinetic modeling of dynamic PET data enables the generation of quantitative parametric maps reflecting tracer kinetics, but it is computationally demanding and sensitive to noise. This study investigated whether embedding the underlying kinetic model of PET tracer behavior into a neural network framework (PET kinetics-informed neural network, PKiNN) improves voxel-wise parameter estimation compared with conventional linearized approaches.Approach.PKiNN, a recurrent neural network with long short-term memory (LSTM), was trained on simulated time-activity curves generated using a one-tissue compartment model (1TCM) with randomized kinetic parameters spanning physiologically plausible ranges. A kinetics-informed loss term was incorporated into the network loss function to constrain the predicted parameters. A baseline neural network (NN-LSTM) without the kinetics-informed loss was also trained for comparison. Performance was assessed across multiple simulated noise levels and using clinical11C-UCB-J PET data. A supplementary simplified reference tissue model (SRTM) implementation was evaluated as a proof of concept.Main results.Across all simulated noise levels, PKiNN consistently outperformed NN-LSTM, yielding lower relative error (RE, %) and absolute relative error (ARE, %) for,, and. For example, at an SNR of 20 dB, the REs and AREs forestimated by NN-LSTM wereand, respectively, while for thederived from PKiNN, RE and ARE wereand, respectively. For clinical11C-UCB-J PET data, PKiNN-derived voxel-wise maps showed close agreement with VOI-based 1TCM fitting, minimal bias inand, modest overestimation of, and reduced underestimation compared with Logan graphical analysis.Significance.Overall, PKiNN provides a fast, accurate, and noise-robust alternative to conventional 1TCM kinetic modeling, demonstrating feasibility for clinical PET, and the supplementary SRTM proof-of-concept provides preliminary evidence of its potential for extension to other kinetic models.