Function Magnetic Resonance Imaging (fMRI) of motor behavior and motor learning

An fMRI Compatible Robot

The Shadmehr Lab has completed the design and construction of an actuated robotic arm that can be used for fMRI experiments.  With the robot we have been investigating the neural basis of motor control, particularly for arm movements.  The first publication that described these results appeared in the Journal of Neuroscience.  The people responsible for the design and engineering of the robot are Mike Turner, Jamie Hartwell, Maneesh Dewan, Jörn Diedrichsen, and Reza Shadmehr. Machining was performed in the BME shop by Jay Burns.  A computer, situated in a room adjacent to the MRI, controls the force on the handle of the robot by controlling inputs of the servo valves. A small (2 cubic feet), extremely quiet (<60 db) air compressor provides the air pressure in the system.

The linkages are made of plastic composites. The joints of the robot house plastic composite ball bearings to facilitate smooth motion of the robot. The robot is actuated with a pair of two-way, air-driven cylinders that house pistons. Air pressure, regulated by four servo valves that are mounted very close to the cylinders, results in force on the pistons. The force produced by the pistons is bi-directional. This force is transmitted to the joints of the robot through a linkage. The resulting torques are in turn transmitted to the handle of the robot via another set of linkage. The result is a sensation of force on the subject's hand.
Optical encoders mounted parallel to the cylinders measure position of the robot's links at a resolution of better than 10 micron. This signal is sent to the computer and software calculates position of the robot's handle. This position is displayed to the subject via a video monitor. Here is a view of one of the actuators and one of the encoders.

 

Artifacts

There are number of problems that needed to be solved in order to study motor control in the MRI. The first was the design of a compatible robot. The second was the head motion resulting from the movement of the arm.  We found that head movement as small as half a millimeter could produce an increase in the noise in fMRI by a factor of 4. To stabilize the head we are using a custom-fit bite bar shown on the left.

Even if the head is completely stable, the motion of a body part of significant mass in the MRI itself causes a problem. Displacement of mass causes distortion of the magnetic field (B0 field), which causes distortions in the measured images. The figure on the left shows which portions of the functional images had a higher signal when the person held the arm stationary on the right (blue) vs. on the left (red).  That is, because the human arm has significant mass, simply placing it at the right side vs. the left side produces large artifacts.  A similar artifact can be observed when one images a water balloon and moves another balloon from one side of the bore to another. Currently we have no method to correct for the distortion effectively.  Also, the dynamic effects of a moving mass are much harder to predict and correct for than the static effects shown on the left. Thus, we monitor the influence of artifact on our results carefully and developed a method to detect and exclude noise artifacts from fMRI time series data.

The method we use to detect and adjust for artifacts is based on Weighted-least-squares. It is motivated by the observation that the variance of the images is not time-homogenous, as often assumed in fMRI-data analysis, but that the noise variance varies from image to image. The upper part of the figure shows the variance estimates of a time series of images with a movement-related artifact indicated by the arrow on the right. We developed a Restricted Maximum Likelihood (ReML) estimator for the variance of each image in a time-series and use it to weight each image by the inverse of the variance. More on this method here.

 

Surface-based analysis and inter-subject normalization

Individual anatomy of the brain and especially the folding of the neocortical sheet can differ quite dramatically from person to person. We try to respect the underlying anatomy by analyzing the functional data using surface-based methods. A high-resolution T1-weighted (anatomical) scan is used to generate a segmentation between cortical white and gray matter.  The boundary between the two is then reconstructed to form an accurate 3D-model of the individual cortical surface.  This surface then can be blown up and aligned with other participants based on cortical or functional landmarks. On the left we have an image of the flattened left cerebral hemisphere and flattened cerebellum. 

We use the software package SureFit and Caret, written and freely distributed by David van Essen's lab at Washington University. For anatomically well defined functional areas, such as primary motor cortex, this method leads to much more accurate alignment than 3D-morphing, which is often used for fMRI studies.   

As data analysis is performed on a 2D surface, we needed a way to access significance of cortical activation on that surface. One approach is to use Bonferroni-correction of statistical tests: we correct the chance over the approximately 70000 t-test that make up a cortical surface, such that our chance of getting a single significant t-test is below a level of alpha=0.05. This is called the Family-wise error correction (FWE). However, Bonferroni-correction is conservative, because it ignores the spatial dependence of neighboring t-tests. Also, we often would like to access the significance of the spatial extend of a activation.  One solution is to apply Gaussian-field theory. We have developed a set of matlab functions to implement this approach.