Robust dynamic susceptibility contrast MR perfusion using 4D nonlinear noise filters

Abstract

To investigate if 4D (simultaneous space and time) nonlinear filtering techniques can produce more robust cerebral blood flow (CBF) estimates by reducing noise in acquired dynamic susceptibility contrast (DSC) MR perfusion data. A digital anthropomorphic brain perfusion phantom was constructed to analyze filter performance by: 1) deriving anthropomorphic tissue volume fractions from a human subject and 2) simulating DSC-MR perfusion signals for voxels with mixed tissue for various signal-to-noise ratios (SNRs). DSC-MR data for 11 acute ischemic stroke patients were also acquired at 3T. CBF maps cross-calibrated so that normal white matter CBF was 22 mL/minute/100 g were produced from DSC-MR data without filtering and from 4D-Gaussian and 4D-bilateral noise-filtered DSC-MR data. The nonlinear 4D-bilateral filter yielded the lowest CBF root-mean square error (RMSE) in the phantom experiments with noise (average RMSE across all tissues regions for no filtering, 4D-Gaussian, and 4D-bilateral was 5.3 mL/minute/100 g, 6.2 mL/minute/100 g, and 4.0 mL/minute/100 g, respectively) and had the best image quality in both the phantom and patient data. Nonlinear 4D noise filters are better suited to the 4D nature of DSC-MR data. Linear spatial filters are not appropriate and can produce larger CBF errors than without filtering.