Investigation of Spatial-Temporal Kernel Method for Dynamic Imaging in Short and Long Range PET Scanners

Dynamic positron emission tomography (PET) imaging has a limited temporal resolution, and always suffers from poor signal to noise ratio (SNR) due to low count statistics from short time frames and the ill-posed nature of expectation-maximization (EM) based reconstruction algorithm. Increasing acquisition time for each frame improves image SNR, but may introduce motion artefacts to the region of heart, lung and head. So, a trade-off has to be made between temporal resolution and image quality. Previous studies have demonstrated that aided by prior information from long time frames and/or anatomical images, the kernel method can effectively improve the SNR of the dynamic PET images from very low-count data. In this study, we have investigated the performance of kernel method in long range and simulated short range scanners. We implemented a spatial-temporal kernel method (ST-KEM) and applied it to the data collected from the 194-cm PET/CT scanner (uEXPLORER). Two short range PET scanners were simulated by using partial of the total eight units of uEXPLORER. Additionally, we exploited the way to improve prior images of short range scanners through deep learning. The results showed that ST-KEM outperformed the kernelized Expectation-Maximization (KEM) and conventional Ordered Subsets Expectation Maximization (OSEM) for frames as short as 0.1 second in uEXPLORER, from which noiseless cardiac motion signal could be extracted. In simulated short range scanners, similar image quality could be achieved when the frame duration extends to 1 to 2-second.