A Gradient-Informed Robust Motion Correction Method for FMRI
Cerebral activation may be a serious confounding effect during the estimation of motion correction parameters in FMRI time series. This effect, which stems from the fact that activated voxels violate the assumption of intensity conservation for perfectly aligned images, is particularly significant when using least squares-based similarity measures.
One way to deal with this problem is by down-weighting cerebral activation confounding signals during registration, which can be achieved using different metrics, other than least squares, based on robust estimators. However, this approach may lead to accuracy problems related to the increasing number of local minima, which manifest through an increased variability in motion estimates. The minimization of this problem could rely in the introduction of a pre-processing spatial smoothing step, but this strategy is likely to increase the bias between activation and motion correction parameters, due to the spatial consistency of the activation. A compromise between these two factors is obviously difficult.
In this paper, we present a different strategy, which consists of a gradient-informed robust motion correction method for FMRI time series. The robust similarity measure is least squares-based incorporating a Geman-McClure M-estimator. The cut-off power of the Geman-McClure estimator for each voxel pair is set as a linear function of the local gradient of the reference image. This strategy allows maintaining a high sensitivity relative to true intensity differences due to spatial misregistration, while minimizing activation-related confounding differences.
The robustness of the proposed method is first evaluated using a motion-free simulated time series including artificial activation-like signal changes based on a simple box-car paradigm. Results are compared with four other registration methods, which combine a least squares or a robust least squares similarity measure with a varying or non-varying strategy for spatial smoothing. These five methods are finally tested on three actual time series obtained from a 3T magnet.
KeywordsRoot Mean Square Motion Estimate Motion Correction Registration Method Spatial Smoothing
Unable to display preview. Download preview PDF.
- 2.Roche, A.: Recalage d’images médicales par inférence statistique, PhD Thesis. Université de Nice-Sophia Antipolis, Projet Epidaure, INRIA (2001)Google Scholar