Incorporating HYPR de-noising within iterative PET reconstruction (HYPR-OSEM)

Objectives HighlY constrained back-PRojection (HYPR) is a promising de-noising strategy originally developed for time-resolved MRI, and it has been recently applied to dynamic PET imaging and shown promising results (1-3). In this work, the HYPR-LR (constraining the back-projection to Local Regions of interest) was directly incorporated within the OSEM reconstruction for PET, and the performance of various forms of HYPR-OSEM in static imaging (i.e. single frame reconstruction) was evaluated using resolution and contrast phantom simulations.

Methods The HYPR de-noising operator was introduced (a) in the forward projection only, (b) in both forward projection and back-projection of the system model, and (c) after the OSEM update for each subset thus generating 3 forms of HYPR-OSEM: (a) HYPR-F-OSEM, (b) HYPR-FB-OSEM, and (c) HYPR-AU-OSEM, respectively. The corresponding composite images were computed as the sum of the preceding subset images and were updated for each iteration. Ideally, the composite image in HYPR needs to have high resolution with low noise, and several initial OSEM updates are required to achieve high resolution; however, noise increases as the number of updates increases. In this work, HYPR was incorporated after only one iteration of OSEM with 16 subsets. The filter function used in the HYPR operator was a Gaussian with a 5 mm FWHM (2 image voxels) except for the early updates. Since the changes between the early updates are more drastic as compared to the later updates, a narrower filter kernel (3 mm FWHM) was used in the HYPR operator during the early updates to limit the cross-talk between the composite and the target. For HYPR-FB-OSEM, a 3 mm kernel was also used for the HYPR operator in the back-projection to ensure accurate update factors can be obtained. A point source phantom (3 sources) with FWHM of 5.5 mm (including the point-spread-function of typical clinical PET scanners) and a contrast phantom with a hot-to-background ratio of 5:1 and a cold region were simulated. The point source phantom was simulated with a sufficient number of counts (1 million) while the contrast phantom was simulated under a low count condition (20 thousand counts: ~2 counts per sinogram bin). The resolution and contrast recovery coefficient (CRC) vs noise performances were evaluated for various forms of HYPR-OSEM. In addition, the data with 3 times higher activity in the contrast phantom (60k counts) were reconstructed with the conventional OSEM for comparison purposes.

Results The resolution performance of all forms of HYPR-OSEM was comparable to that of the conventional OSEM. Small variation in both FWHM and FWTM was observed for all forms of HYPR-OSEM. With regard to CRC vs noise, although the convergence in CRC for all forms of HYPR-OSEM was observed to be slower than the conventional OSEM, for a fixed CRC a 2-3 times lower noise level can be achieved by HYPR-OSEM. HYPR-AU-OSEM was observed to outperform the other HYPR-OSEM methods slightly as the number of iterations increases in this case. Furthermore, the CRC vs noise curves for HYPR-OSEM were observed to approximately follow that of the OSEM reconstruction with 3 times higher activity thus demonstrating the ability to improve the signal-to-noise ratio and to reduce the injected dose.

Conclusions Very promising results were obtained from the initial implementation and evaluations of HYPR-OSEM, and it was observed that similar image quality as compared to the conventional OSEM can be achieved by HYPR-OSEM with much lower injected dose and/or scan time (approximately 3 times in this case) in static PET imaging. Further optimization of the parameters and applications to brain and whole body imaging will be explored.