Multiple Reconstructions with Different Data Quality for a Single Scan using Kernel Method in Digital PET

Abstract

In Positron Emission Tomography (PET), the data recorded in narrow energy window (EW) has the benefits of low scatter fraction (SF), which leads to a high contrast image, but it is rarely used in reconstruction because many trues are rejected. The data recorded in wide EW contains almost all trues, which leads to a noiseless image and a lot of scatters as well. To combine the characteristics of high contrast and low noise in one image, we propose a new approach to reconstruct image with both the data recorded in a narrow EW and a wide EW, which the digital PET has the ability to do. The wide EW generates a high-count, high-noise (HCHN) data and the narrow EW generates a low-count, low-noise (LCLN) data. The HCHN data is reconstructed by OSEM algorithm to obtain a prior image used in the LCLN data reconstruction. The LCLN data is reconstructed by the kernel expectation maximization (KEM) algorithm to obtain the final image. We performed a 2 step investigation on 3D brain phantom with simulated data: 1) we compare the effect of different iterations of prior image when the HCHN and LCLN data are both generated by a wide EW of 450 to 700 keV (87M trues, 41% SF); 2) we evaluate the effect of turning up the low-threshold of EW from 450 to 520 keV (17m trues, 22% SF) of LCLN data. The results show that a low iteration of prior image obtains a best final image (2 iterations 11 subsets in this work). The final image has a similar noise level but a higher contrast in the same iterations when the low-threshold of EW of LCLN data is turned up. In conclusion, the proposed method has the ability to combine the benefits of narrow EW and wide EW in one reconstructed Image,