Direct Deep-Learned Kinetic Parametric Map Separation for Multiplexed PET

Multiplexed positron emission tomography (PET) imaging combines information from multiple tracers to open up greater levels of relevant information for clinical decisionmaking in a single scan. Estimating kinetic parametric maps from multiplexed PET is challenging due to the high level of noise in the raw dual-tracer images obtained from the scanner, and an arterial input function needs to be provided. Deep learning now offers useful new tools for multiplexed signal separation without explicitly knowing the input functions and has been shown to implicitly denoise the separated dual-tracer images. In this work, we propose a deep learning framework for kinetic parametric map separation for multiplexed PET. In particular, we investigated the feasibility of using a trained neural network to separate dynamic [
¹⁸
F]FDG and [
¹¹
C]MET dual-tracer images into single-tracer kinetic parametric maps when two tracers are injected in close succession. The kinetic parametric maps K
1
and k
2
are learned jointly using a 2D convolutional encoder-decoder with weighted mean square error loss. The results obtained suggest that the deep learning-based method leads to better separation performance compared to the classical compartment modelling-based method for parametric maps K
1
and k
2
. Future work will focus on including k
3
and k
4
in the separation task and investigating the weighting strategy in the loss function to further improve the separation performance.