Collimator-detector response compensation in quantitative SPECT reconstruction

Abstract

The purpose of this study was to investigate the importance of collimator-detector response compensation in single photon emission computed tomography (SPECT). The compensation methods included in this study addressed three important degrading factors, which are attenuation, scatter and collimator-detector response (CDR). Geometric response (GR) and septal penetration (SP) are two most important components of collimator-detector response. The compensation work has been divided into two general categories: one is GR compensation based reconstruction (GR-RECON), which includes attenuation, scatter and geometric response compensations, and the other one is GR-SP compensation based reconstruction (GR-SP-RECON), which consists of attenuation, scatter geometric response and septal penetration compensations. Ordered-subset expectation maximization (OS-EM) method is applied as the reconstruction framework. SIMIND Monte Carlo (MC) code is incorporated in the forward projection. Multiple projection sampling convolution-based forced detection (MP-CFD) is used here to accelerate the MC code. The convolution kernels in CFD are generated by the ray-tracing (RT) method. In order to further increase the convergence speed, the models of attenuation and collimator-detector response are included in the backprojection step. In order to assess the quantitative accuracy of the collimator-detector response compensation, the reconstruction images of cylinder, four different size spheres in various medium, and NURBS-based cardiac-torso (NCAT) phantom are evaluated using 1-131 and high energy general resolution (HEGR) collimators. The reconstruction images including GR compensation appear very obvious collimator dependent Gibbs ringing artifact. The quantitative estimation of the spheres and NURBS-based cardiac-torso (NCAT) phantom has denoted the high accuracy of GR-SP-RECON, whereas, a lot of quantitative activity will be mistakenly estimated in the background using GR-RECON.