High Spatial-temporal Resolution PET Image Reconstruction Based On Autoencoder

PET (Positron Emission Tomography) is an advanced clinical inspection imaging technology in the domain of nuclear medicine. After injecting the radio tracer into the human body, the radioactive substances will decay and emit positron during the metabolism. The positron meets an electron and annihilates after flying for some time, which produce a pair of opposite photons with 511keV energy. Then photon detector captures the event and gets the emission data. Once we got the emission data, we can reconstruction the radio activity distribution image. For dynamic PET imaging, we collect the emission data for a certain time interval, which can reflect more physiological information with the addition of time acquisition.How to reconstruct the radioactivity distribution image started with the emission data has become a challenging problem. Indeed, the temporal resolution and spatial resolution are restricted to each other in dynamic PET imaging. In this paper, we started from combining the feature information of several frames to suppress the noise and retain the key features. Moreover, we want to realize a competitive reconstruction result with higher spatial resolution and satisfy the requirement of temporal resolution at the same time. The following points are the main contribution of this paper:(1)We proposed an autoencoder template to extract the features in PET images and discover the higher level feature by stacking several templates together. Through the existed data and labels, we train the parameters by back propagation algorithm and realize the whole SAE framework for dynamic PET images reconstruction.(2)Based on the maximum likelihood expectation maximization algorithm and state space algorithm based on Kalman Filter, we reconstruct the images in dynamic PET as our training input and the ground truth as labels, then we test our new emission data based on the model we have trained. The test data includes the brain phantom and Zubal phantom simulated by Monte Carlo algorithm as well as the real patient data.(3)Based on the stacked autoencoder, we proposed feature extraction model for different radio tracer in PET imaging and realize the PET image reconstruction for dual tracer started from the mixed emission data.