Comparing reconstruction algorithms using a multi-variate analysis

We propose a voxel-based multi-variate analysis to evaluate the performance of tomographic image reconstruction. This technique allows simultaneous comparison of the underlying tush with resolution and noise behavior as well as other effects across the entire volume for different reconstructions. We demonstrate the idea with a 2D simulation and apply the method to compare 3D [O-15]water PET studies of a motor task reconstructed by 3D reprojection (3DRP), Fourier rebinning followed by 2D filter backprojection (FORE-FBP) and iterative filtered backprojection with median root prior (IFBP-MRP). The difference between the three reconstructions was found to be significant relative to the baseline-activation effect. Although the differences in resolution and noise characteristics between 3DRP and FORE-FBP are similar to those reported in the literature, the major pattern that separates the two reconstructions is a mean difference in the axial direction. On the other hand, IFBP-MRP shows a clear pattern of improved resolution with higher noise compared to 3DRP. All three reconstructions are capable of discriminating the baseline scans from the activation scans with high significance.