Iterating Registration and Activation Detection to Overcome Activation Bias in fMRI Motion Estimates Jeff Orchard ⋆ and M. Stella Atkins * Simon Fraser University, Burnaby BC V5A 1S6, Canada, {jjo,stella}@cs.sfu.ca Abstract. Most intensity-based fMRI registration methods do not ac- count for the fact that the volumes being aligned may differ: one may have blood oxygen level dependent (BOLD) contrast while the other does not. Especially in least-squares registration, this can result in motion pa- rameter errors that are correlated to the stimulus. An iterative technique to overcome this activation bias is proposed and analyzed. The method, using mostly off-the-shelf software, is able to find the least-squares solu- tion to both the registration and activation detection problems simul- taneously. The resulting motion parameters and activation maps are considerably more accurate, yielding two-thirds fewer false-positive and one-third fewer false-negative activations. 1 Introduction In functional MRI (fMRI), patient motion can have a very damaging effect on the accuracy of the resulting statistical parametric activation maps. Patient move- ments of a fraction of a millimetre or degree have the potential to cause false- negative activations [1] and false-positive activations [2]. A host of registration techniques have been proposed to combat patient mo- tion [1, 3, 4]. These methods use a variety of cost functions and optimization schemes. One commonly-used cost function is the least-squares cost function, which can be shown to be the optimal choice when the images being com- pared differ only by Gaussian noise. However, least-squares techniques are sen- sitive to outliers. In fMRI, outliers come from a variety of sources, includ- ing the blood oxygen level dependent (BOLD) signal itself. Freire et al. [5] showed that the presence of BOLD contrast can bias the motion estimates cal- culated by least-squares registration methods. This influence becomes notice- able in fMRI datasets acquired on scanners with field strengths of 3 Tesla and higher. When the dataset is motion-compensated with these inaccurate motion estimates, stimulus-correlated motion is introduced into the dataset. Stimulus- correlated motion is especially damaging [2] because it makes it impossible for the ⋆ This work was supported in part by the Natural Science and Engineering Research Council of Canada (NSERC)