Using Haptic fMRI to Enable Interactive Motor Neuroimaging Experiments

Abstract

Combining haptics with functional magnetic resonance imaging (Haptic fMRI) has enabled complex motor neuroimaging experiments that non-invasively map real-world motor tasks on to the human brain. The technique’s resolution, fidelity and susceptibility to scanning artifacts, however, have not yet been estimated in a quantitative manner. Here, we demonstrate that unconstrained three degree-of-freedom Haptic fMRI experiments can reliably activate brain regions involved in planning, motor control, haptic perception, and vision. We show that associated neural measurements are reliable, heterogeneous at the millimeter scale, and free from measurable artifacts, and that their anatomical localization is consistent with past neuroscience experiments. In addition, we demonstrate the feasibility of using electromagnetic actuation in Haptic fMRI interfaces to apply high fidelity open-loop three-axis haptic forces (0.5–2N; square or 0.1–65Hz sine waveforms) while maintaining negligible temporal noise in pre-motor, motor, somatosensory, and visual cortex (<1 % of signal). Our results show that Haptic fMRI is a robust and reliable technique for characterizing the human brain’s motor controller.

Appendix

1.1 MRI Protocol

All fMRI scans were conducted at Stanford University’s Center for Cognitive and Neurobiological Imaging on a GE Discovery MR750 3T MRI scanner, with a thirty-two channel Nova Medical head coil. The scan protocol was gradient echo EPI with a 16 cm field of view sampled at a $64\times 64$ resolution ($2.5\times 2.5\times 2.5$ mm$^3$ voxels), a 1.57 s repetition time, a 28 ms echo time, and a 72$^\circ$ flip angle. Each scan run was preceded by 2nd-order polynomial shimming and was sandwiched by spiral fieldmap scans ($2.5\times 2.5\times 5$ mm$^3$ voxels). Fieldmap scans were conducted within 10 s of each scan run’s start and end. After scanning, the fMRI images were slice time corrected, motion corrected (SPM [44]), spatially undistorted using fieldmaps, and analyzed to compute temporal noise-to-signal.

1.2 fMRI Analysis

Temporal noise-to-signal computations used the median neural response distribution obtained by regressing out a line from each voxel’s time series, computing the absolute value of the difference between successive time points, computing the median of these absolute differences, dividing the result by the mean of the original time series, and then multiplying by 100. Cortex segmentation used Freesurfer’s Desikan-Killiany atlas [45]. Surface registration was done using Freesurfer, and all surface images were plotted using Freeview. Freeview smoothed the surface plots while rendering (2 steps).

1.3 Estimating fMRI Impulse Response Time Series and $R^2$

fMRI measures changes in blood oxygenation induced by neural metabolic activity [1, 2], which have a slower time course than neural computation and persist long after sensory stimuli and motor tasks terminate. Such persistent responses cause raw fMRI measurements to overlap in experiments where consecutive task conditions are not be separated by large time-intervals. Separating task conditions by large time-intervals, however, makes fMRI runs very long, which can induce a variety of unwanted artifacts related to MRI scanner calibration drift, neural adaptation, or subject attention lapses, microsleep and exhaustion. Instead, we optimized our experiments to ensure reliable motor task execution [13], which caused fMRI measurements for different task conditions to overlap.

We segregated neural activation for individual tasks using a finite impulse response (FIR) model (implemented using GLMdenoise [46]). The FIR model works by associating each task type with a unique time course and segregates time courses while assuming that overlapping responses sum linearly. fMRI signal linearity, however, is an active area of research [1, 41, 47]. As such, we randomized inter-task delays and randomly ordered tasks, which made the model’s time series match anatomical expectations based on past research (see Figs. 2 and 3; read [6] for an overview). When tasks were closely spaced in time, as with planning and motion, this method was noisy. The parts of planning that overlap with motion are thus less reliable and the confidence interval for the planning time series estimates is larger after motion starts (but still above zero; see Fig. 1).

We computed $95\,\%$ confidence intervals by bootstrapping [48] runs (400 bootstraps), fitting FIR models to each, and taking the median percentile estimates across the estimated bootstrap time series. Finally, we computed $R^2$ values for each voxel by comparing the time series variance with the variance after regressing out median FIR model estimates.

1.4 Data Collection Protocols

See [13] for precise specifications of the motion protocol. Subjects executed one practice run inside the MRI scanner, and then executed at least eight scan runs (S1, 10; S2, 8). Each run was 630 s long.

The force and visual perception experiment protocol involved fixed duration stimuli instances with visual, motor, or visual and motor sensory input. The experiment was divided into runs, and each run was divided into blocks. During each block, the subject started with their hand at rest. Next, they were instructed to move their hand into free space. After a randomized delay period of 3–5 s, the subject experienced two randomly selected stimuli instances. Each stimulus instance was 3–5 s long and was separated from the other by a randomized delay 3–5 s. Finally, the subject was required to rest their hands for a random time interval (4–20 s), and then restarted the process. The subject executed four scan runs with multiple blocks. Each run was 459 s long.

Force magnitudes were set to evenly spaced directions along the x-y plane, with a magnitude of 1.2 N. The force vectors used were (1.2, 0.0), (0.0, 1.2), ($-$1.2, 0.0), (0.0, $-$1.2), (0.85, 0.0), (0.0, 0.85), ($-$0.85, 0.0), and (0.0, $-$0.85).

1.5 Haptic and Force Measurement Details

Haptic experiments were conducted with Haptic fMRI Interface [12], a three degree-of-freedom fMRI-compatible device. All motions were right handed, and the haptic control rate was 350 Hz. The reaching task spanned the entire workspace (see [13] for more details), but avoided arm motion artifacts [14].

Visual stimuli were displayed on a 30 in. diagonal (76 cm, 16 : 10 aspect ratio) flat panel display custom built by Resonance Technology. Subjects viewed visual stimuli through a dual-mirror setup. The visual distance from screen to mirror-2 is 184.4 cm, from mirror-2 to mirror-1 is 6 cm, and from mirror-1 to the eye is about 15 cm, for a total viewing distance of about 205 cm. The visual field of view is about 30$^\circ$, making each visual checkerboard square span about one and a half degrees of the visual field. The display has a native resolution of $2560\times 1600$ but stimuli were displayed at $1280\times 800$. The display has a 7 ms temporal response, and 10-bit color rendering. The maximum luminance of the display is 329 cd/m$^2$ (red is 88, green is 117, and blue is 124 cd/m$^2$).

Forces were measured using a JR3 85M35A-U560 63N4S force sensor. The raw sensor data was sampled at 1 KHz, resampled to match HFI’s control rate, and was finally filtered using a 75 Hz low pass filter to remove high frequency sensor noise.

1.6 Human Subjects

Subjects were healthy right-handed males with no history of motor disorders: S1, 29y, 185lb, 5$^\prime$9${^\prime }{^\prime }$; S2, 19y, 170lb, 6$^\prime$2${^\prime }{^\prime }$; S3, 21y, 160lb, 5$^\prime$8${^\prime }{^\prime }$. Informed consent was obtained in advance on a protocol approved by the Institutional Review Board (IRB) at Stanford University.