Assessment of bias in experimentally measured diffusion tensor imaging parameters using SIMEX

Authors

  • Carolyn B. Lauzon,

    1. Department of Electrical Engineering, Vanderbilt University, Nashville, Tennessee, USA
    2. Institute of Imaging Science, Vanderbilt University, Nashville, Tennessee, USA
    Search for more papers by this author
  • Ciprian Crainiceanu,

    1. Department of Biostatistics, Johns Hopkins University Bloomberg School of Public Health, Baltimore, Maryland, USA
    Search for more papers by this author
  • Brian C. Caffo,

    1. Department of Biostatistics, Johns Hopkins University Bloomberg School of Public Health, Baltimore, Maryland, USA
    Search for more papers by this author
  • Bennett A. Landman

    Corresponding author
    1. Department of Electrical Engineering, Vanderbilt University, Nashville, Tennessee, USA
    2. Institute of Imaging Science, Vanderbilt University, Nashville, Tennessee, USA
    • Vanderbilt University EECS, 2301 Vanderbilt Pl., P.O. Box 351679 Station B, Nashville, Tennessee 37235-1679
    Search for more papers by this author

Abstract

Diffusion tensor imaging enables in vivo investigation of tissue cytoarchitecture through parameter contrasts sensitive to water diffusion barriers at the micrometer level. Parameters are derived through an estimation process that is susceptible to noise and artifacts. Estimated parameters (e.g., fractional anisotropy) exhibit both variability and bias relative to the true parameter value estimated from a hypothetical noise-free acquisition. Herein, we present the use of the simulation and extrapolation (SIMEX) approach for post hoc assessment of bias in a massively univariate imaging setting and evaluate the potential of a SIMEX-based bias correction. Using simulated data with known truth models, spatially varying fractional anisotropy bias error maps are evaluated on two independent and highly differentiated case studies. The stability of SIMEX and its distributional properties are further evaluated on 42 empirical diffusion tensor imaging datasets. Using gradient subsampling, an empirical experiment with a known true outcome is designed and SIMEX performance is compared to the original estimator. With this approach, we find SIMEX bias estimates to be highly accurate offering significant reductions in parameter bias for individual datasets and greater accuracy in averaged population-based estimates. Magn Reson Med, 2013. © 2012 Wiley Periodicals, Inc.

Ancillary