Accelerated Deep Image Prior-based PET Image Reconstruction Using Two-Step Optimization

Convolutional neural networks (CNNs) are a powerful tool for deep learning-based positron emission tomography (PET) image reconstruction to enhance PET image quality. Recent unsupervised learning-based PET image reconstruction methods have shown a better generalization performance with different PET tracers because they do not need to use any training dataset, such as pairs of high- and low-quality PET images. However, these reconstruction methods require very time consuming calculations. In this work, we proposed an accelerated deep image prior (DIP)-based unsupervised PET image reconstruction using two-step optimization to improve the computational speed as well as the reconstruction performance. The proposed reconstruction method consists of two steps: a pre-training step and a reconstruction step. In the pre-training step, the pre-training is performed using the magnetic resonance image as input data and the reconstructed image from measured PET data as target data. In the reconstruction step, the CNN weights are transferred from the first step network. And then, we adjust (or fine-tune) the network weights using measured data, and get the final reconstructed image in a few iterations. The proposed reconstruction method was evaluated using Monte Carlo simulation data of a human brain [
¹⁸
F]FDG PET, and its results were compared with those of the maximum likelihood expectation maximization and the other DIP-based one-step reconstruction method. The simulation results showed that the proposed two-step reconstruction method has a faster computational time than the previous one-step method by substantially reducing the number of iterations while enhancing the reconstruction performance. Therefore, we concluded that our proposed two-step reconstruction method is a practical algorithm for DIP-based PET image reconstruction.