Detector modeling in PET list-mode reconstruction: comparison between pre-calculated and on-the-flight computed system matrix

Objectives In list-mode PET iterative reconstruction two major approaches are used to model the detector response. The first one is based on measured or estimated system response matrix (SRM), subsequently stored for use. The second approach is based on on-the-flight computation of the SRM coefficients. In this work we compare these approaches in terms of image qualitative/quantitative accuracy and computation times.

Methods The NEMA NU-4 2008 phantom together with the preclinical Siemens-INVEON PET scanner were considered. All data were simulated with the GATE platform (back-to-back emission) and reconstructed with a LM-EM approach. Three SRM simulated with GATE and different statistics were used. Three on-the-flight projectors were also used: the Siddon projector (Siddon,Medical Physics,1985), a projector with a Gaussian PSF modeling (Ortuno et al, NSS-MIC, 2011) and a multi-ray approach called IRIS (iterative random IDRF sampling). The recovery coefficient (RC) in the 5, 4, 3, 2 and 1mm diameter rods and the NoiseSD in the warm homogeneous background body were measured for the 100 firsts iterations. Reconstructions with pre-calculated and on-the-flight computed system matrices were performed on CPU and CPU/GPU respectively.

Results For the 4 and 5 mm rods, no major differences were noticeable. The trade-off between noise and RC was low for the smallest rods using the Siddon and low statistics SRM based reconstruction. The Gaussian projector had the advantage of associated low noise levels but the RC was higher than expected for the 3 largest rods. The IRIS based reconstruction gave similar or superior results to the pre-calculated high statistics SRM. As expected, the reconstruction using projectors computed on-the-flight on a GPU ran between one and two orders of magnitude faster than the pre-calculated SRM reconstructions.

Conclusions The IRIS reconstruction leads to equivalent or superior results in terms of image quality and computational speed compared to a pre-calculated SRM based reconstruction.