Improved Activation Detection via Complex-Valued AR(p) Modeling of fMRI Voxel Time Series

• Activity Number: 415
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Daniel Adrian* and Ranjan Maitra and Daniel Rowe
• Companies: Grand Valley State University and Iowa State University and Marquette University
• Keywords: Cochran-Orcutt estimation procedure ; partial autocorrelation function ; Rice distribution ; signal-to-noise ratio ; bilateral finger-tapping motor experiment ; phase information

Abstract

A complex-valued model with AR(p) errors is proposed as an alternative to the more common Gaussian-assumed magnitude-only AR(p) model for fMRI time series. Likelihood-ratio-test-based activation statistics are derived for both models and are compared in terms of activation detection and false discovery rates for simulated and experimental data. For simulated data, the complex-valued AR(p) model likelihood-ratio activation statistic shows superior power of activation detection at low signal-to-noise ratios and lower false discovery rates. Also, when applied to an experimental data set, the activation map produced by the complex-valued AR(p) model more clearly identifies the primary activation regions. Our results advocate the use of the complex-valued data and the Gaussian AR(p) model as a more efficient and reliable tool in fMRI experiments over the current practice of using only the magnitude dataset.