On Parameter Selection for Joint Spectral Reconstruction in Y90 SPECT

Abstract

Model-based image reconstruction for Y90 SPECT is challenging due to energy-dependent forward models including collimator-detector response (CDR) and attenuation correction factor (ACF). Since it is common to model CDR and ACF for a single energy for one energy window, there is a trade-off between the width of an energy window (number of primary counts) and the accuracy of modeling energy-dependent CDR and ACF. We previously proposed a joint spectral reconstruction (JSR) for Y90 SPECT to utilize wide energy window while accurately modeling energy-dependent forward projections for multiple windows and showed promising quantitative performance. However, the scale parameter model in JSR to relate images of different energy windows was not properly considered. In this work, we argue that scale parameters should reflect two probabilities: 1) different emission probabilities for the different windows and 2) detection probabilities in SPECT crystal that are different for the two windows. These arguments lead us to utilize Monte Carlo simulation of a point source instead of primary phantom counts for estimating scale parameters to avoid the effect of energydependent attenuations. We performed a SIMIND MC simulation study with a digital phantom similar to the NEMA PET phantom, but with larger hot spheres and a cold sphere in the center. MC simulation was done with 6 energy windows starting from 105keV to 285keV with the width of 30keV for each window using a new emission spectrum with both internal and external bremsstrahlung generated from theory and Penelope simulations. Our proposed JSR with new scale parameters yielded favorable quantitative results for both hot and cold spheres over other methods such as JSR with previous scale parameters, single spectral reconstructions with narrow and wide energy windows. Our proposed method yielded better recovery coefficients than other methods by 0.5-16.2% at the 30 iteration for all hot spheres. These differences were particularly substantial for the smallest hot sphere, 2.3-16.2%. For the cold sphere, JSR proposed yielded 3.0-15.8% lower residual count error than other methods.