Direct PET Parametric Imaging Using Deep Image Prior and a Parameter Amplification Strategy

Direct parametric imaging in positron emission tomography (PET) is the technique of reconstructing pharmacokinetic parametric images directly from raw PET data. Nonlinear parametric imaging is challenging due to the high level of noise in the raw PET data and the complex mapping from kinetic parameters to the projection space. In this work, we proposed a parameter amplification strategy and incorporated deep image prior (DIP) and forward physical model to reconstruct parametric images of nonlinear compartmental model. We used a U-Net as generator to output multiple images from different channels, and each channel output was scaled by an amplification factor (AF) to obtain the parametric image. Amplification was applied on the parameters that had small values and less impact on the objective function modeled by Poisson measured data. The proposed DIP-AF method was tested on the two-tissue model using 30-minute dynamic
¹⁸
F-FDG cardiac phantom data, and an AF of 10 was applied on k
3
. Results showed that the DIP-based method outperformed than the traditional indirect method and an optimization transfer direct method (OTEM). The reconstructed k
3
and K
i
images using the proposed DIP-AF method showed less noise and better image structure compared to the DIP-only method, which demonstrated the effectiveness of the parameter amplification strategy