One pair of hands is not like another: caudate BOLD response in dogs depends on signal source and canine temperament

Participants

Participants were dogs (n = 13) from the Atlanta community (Table 1). All were pets and/or released service dogs whose owners volunteer their time for fMRI training and experiments. All had previously completed an fMRI session in which two hand signals were presented by their primary trainer, one indicating forthcoming food reward, the other indicating no reward. Accordingly, all dogs had demonstrated an ability to remain still during training and scanning for periods of 30 s or greater. However, one dog exhibited excessive motion during this experiment and was subsequently excluded from analysis because of insufficient observations after motion censoring (see below).

This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The study was approved by the Emory University IACUC (Protocol #DAR-2001274-120814BA), and all dogs’ owners gave written consent for participation in the study.

Training

For participation in previous experiments (details in: Berns, Brooks & Spivak, 2013; Berns, Brooks & Spivak, in press), dogs took part in a training program using behavior shaping, desensitization, habituation, and behavior chaining to prepare them to be comfortable with the physical confines of the MRI bore and the loud noise produced by scanning. The end goal of initial training was that each dog remain comfortable and motionless for up to 30 s at a time during live scanning.

In the current experiment, dogs received reward and no reward signals from a familiar human (their primary experimental trainer), an unfamiliar human, and from computer-generated stimuli on a projection screen. As mentioned above, each dog in this experiment was highly trained on two hand signals, an upright raised left hand signifying reward, and both hands held sideways and inward-pointing, with the back of each hand toward the dog, signifying no reward (as described in Berns, Brooks & Spivak, 2012). No further training for these signals was required.

The current experiment involved two new experimental stimuli: computer-generated images to be presented on a projection screen. These were to serve the same signaling function as the previously learned hand signals (reward and no-reward). The computer reward signal consisted of an illustrated image of a raised hand, oriented so it would appear as a raised left hand facing the dog. The computer no-reward signal consisted of two illustrated inward pointing hands with thumbs up, oriented to appear as if the backs of the hands were facing the dog (Fig. 1).

For maximum discriminability, the two signals were presented on black backgrounds, and each at one of the wavelengths to which dogs are most sensitive (Neitz, Geist & Jacobs, 1989), a yellow–green for the reward signal (555 nm), and a medium blue (429 nm) for the no-reward signal. Wavelengths were converted RGB values (http://rohanhill.com/tools/WaveToRGB/). Because we were not making predictions about visual processing, but rather wanted two stimuli that were maximally discriminable, we did not attempt to normalize the contrast and luminance of the two computer stimuli.

Training for the current experiment involved biweekly instruction at our training facility overseen by core project staff and regular training at home overseen by each dogs’ primary experimental trainer. Dogs were trained to associate the green raised hand signal with reward and the sideways blue hand signal with no reward via the same method previously employed to teach them the human hand signals (Berns, Brooks & Spivak, 2013). During training sessions, each dog cooperatively stationed him- or herself in a custom-made foam chinrest placed inside of a mock MRI coil. Presentation of signals was contingent on the dog’s being able to station calmly and with minimal motion. The signals were displayed on a laptop screen placed approximately 20 inches in front of the dog, and were presented in short (<20 trials) sessions. Presentation order was semi-randomized (each signal was presented no more than 4 times consecutively, and reward to no-reward ratio was between 1:1 and 1:3, with the reward signal becoming less frequent once each dog became comfortable and experienced with the protocol). If the dog was able to hold still for the duration of the green reward hand signal, a food reward was dispensed immediately after. Following the blue sideways hand signal, no reward was dispensed. If the dog left the station during either signal, the stimulus was turned off, and the handler waited until the dog returned, at which point they presented the same stimulus again.

Association learning with the two computer stimuli was conducted over multiple sessions and for an amount of time comparable to initial training with the human hand signals (approximately 1 month).

When dogs were accustomed to viewing the computer signals, further training was conducted at our training facility. First, dogs were acclimated to a new food-delivery system (dubbed the “treat-kabob”). This was necessary because, while previously food reward had been delivered directly by hand, in the current experiment the projection screen required for presentation of the computer stimuli made this difficult. Furthermore, we wished to evaluate only the context variability of the signals. Therefore, we needed to keep the food delivery methodology constant across the three signal source states. Two variants of the food-delivery system were used—one in which a long wooden dowel with a small treat placed on a dull skewer at the end was slid up a PVC pipe and presented to the dog, and another in which the same dowel was used to bring the treat to the dog’s mouth without the PVC guide tube (the specific delivery method was adjusted for the dog’s comfort). In both cases, dogs were first exposed to the treat kabob in their chin stations on the ground until it was determined they were not nervous and were competent at removing the treats from the skewer. Training was then transitioned into the mock MRI scanner (Fig. 2A, see Berns, Brooks & Spivak, 2013). In 1–3 sessions with each dog’s primary experimental trainer presenting the familiar reward and no-reward hand signals, food reward was delivered via the treat kabob by a project staff member sitting below and to the side of the bore. Each dog also received one or two sessions with an unfamiliar human delivering the reward and no-reward hand signals, to assess their willingness to remain stationary in the presence of an unfamiliar human. Note that the human used in this behavioral assessment was different than the unfamiliar human who delivered signals at testing. When dogs were judged comfortable with this setup, training commenced for acclimation to a screen being placed in the bore. A computer monitor was placed at the end of the bore, directly in front of the dog’s chin station, to simulate the projection screen used in live scanning. These training sessions were aimed at making the dogs comfortable with stationing in the bore with a visual barrier in front of them.

We did not have a strict behavioral metric to determine when this training was complete—rather a practical measure was used. Dogs were cleared for testing when they had completed at least one month of regular practice and were able to do an extended “dress rehearsal” session in the mock scanner without showing anxiety or escape behavior. This session involved stationing with the monitor in place, recorded scanner noise playing, and the primary handler out of view. Dogs also had to be wearing their ear protection, and treats were delivered via the treat kabob. Total time in training, from first introduction to the computer stimuli to finishing “dress rehearsal” ranged from 8 weeks to 16 weeks, with the duration affected by both the aptitude of the dog and the diligence of the owner in implementing homework sessions.

MRI scanning

All scanning for the current experiment was conducted with a Siemens 3 T Trio whole-body scanner. Dogs were stationed in their custom chin-rests, which had been placed in a standard neck coil as previously described in Berns, Brooks & Spivak (2013) (Fig. 2B). All participants wore ear protection during scanning, either Mutt Muffs™ or ear plugs with wrap, depending on dog and owner preference.

Each scan session began with a 3 s, single image localizer in the sagittal plane (SPGR sequence, slice thickness = 4 mm, TR = 9.2 ms, TE = 4.15 ms, flip angle = 40°, 256 ×256 matrix, FOV = 220 mm).

A T2-weighted structural image was previously acquired during one of our earlier experiments using a turbo spin-echo sequence (25–30 2 mm slices, TR = 3,940 ms, TE = 8.9 ms, flip angle = 131°, 26 echo trains, 128 × 128 matrix, FOV = 192 mm), which lasted ∼30 s.

Functional scans used a single-shot echo-planar imaging (EPI) sequence to acquire volumes of 24 sequential 3 mm slices with a 10% gap (TE = 28 ms, TR = 1,400 ms, flip angle = 70°, 64 × 64 matrix, 3 mm in-plane voxel size, FOV = 192 mm). Slices were oriented dorsally to the dog’s brain (coronal to the magnet, as, in the sphinx position, the dogs’ heads were positioned 90° from the usual human orientation) with the phase-encoding direction right-to-left. Sequential slices were used to minimize between-plane offsets from participant movement, and the 10% slice gap minimized the crosstalk that can occur with sequential scan sequences.

Six runs of up to 300 functional volumes were acquired, each lasting approximately 6 min. For the first dog (Zen), three runs of up to 600 functional volumes were acquired, each lasting approximately 12 min. This was subsequently split into 6 runs in a counterbalanced fashion for the remainder of the dogs. As part of a separate experiment, an additional functional run was acquired during the session in which the dog was presented with different types of visual stimuli on the screen; however, these data are not analyzed or reported here.

During functional scanning, reward was delivered via the aforementioned treat-kabob, operated by a project staff member sitting below and to the side of the scanner bore (out of sight of the dog).

Experimental design

Stimuli for this experiment were as described above: the two natural hand signals representing reward and no-reward delivered by a familiar and an unfamiliar human, and the two illustrated computer signals representing reward and no-reward, projected on a screen at the head of the bore.

Each subject dog received 15 reward and 15 no-reward signals with the familiar human, the unfamiliar human, and with the computer images, for 90 trials overall across the three source conditions. Each stimulus was presented for approximately 10 s, regardless of source. These 90 trials were broken into six runs of 15 trials each. An event-based design was used, with reward and no-reward trials presented semi-randomly within each run (either 7 reward trial and 8 no-reward trials or vice versa, and with no more than three of either stimulus type presented consecutively). The six runs for each dog were always in the same order: familiar human, unfamiliar human, computer, computer, unfamiliar human, familiar human (as part of a different experiment, an additional run was included in the middle of this sequence in which different visual stimuli were presented on the computer screen and are analyzed and reported elsewhere). Low sample size didn’t allow for different counterbalancing across participants or randomizing order, and running each source condition on a separate day to avoid effects from in-session habituation was not practical. This ABCCBA pattern of presentation controlled, at least partially, for effects in the BOLD signal from habituation, sensitization, and scanner drift becoming confounded with signal source. In addition, we determined that placing the computer runs first or last might be difficult for some dogs—in practice we had observed that dogs tended to move more and show more signs of anxiety when receiving computer signals as opposed to receiving signals directly from their handlers. By placing the computer signals in the middle of the experimental sequence, each dog had a chance to “warm up” to the task with the more familiar source conditions before the computer runs, but also had the more familiar (potentially easier) source conditions at the end of the experimental sequence, when stress and fatigue may have made them less inclined to continue participating. The first dog’s 90 trials were broken instead into three runs of 30 trials each (15 reward and 15 no-reward), in the following order: familiar, unfamiliar, computer. Thus, he received the same amount of total trials and trials for each condition as the other dogs, but without the pyramidal order provided the other dogs.

Behavioral criteria during testing for all dogs were the same as during training—each dog was required to hold still for the duration of each signal. Following a reward signal, they received a small piece of hot-dog via the treat-kabob. No reward was given following a no-reward signal. Following each 15-trial run, the dog was taken out of the scanner and allowed to walk around, drink water, etc.

For live scanning, some dogs did show some anxiety when first placed in the bore with the projection screen in place and their handler out of sight. Basic training mechanisms were employed to work around this, reprising the initial acclimation approach taken in the mock scanner at our training facility discussed previously. In brief, prior to the beginning of these scanner runs, the dogs were stationed with the screen present and their owner’s face in view. They were rewarded a number of times as the owner was approximated back and to the side, and then live scanning began.

Event recording

Trial events (onset and offset of reward and no-reward signals) were recorded by an observer via a four-button MRI-compatible button-box. The observer stood next to the experimental trainer and unfamiliar person respectively on runs 1, 2, 5, and 6, such that they could see the dog’s head in the bore of the magnet. On these runs, the observer signaled the experimental trainer and unfamiliar person when to present which signal. On runs 3 and 4, the computer runs, the observer moved to the side so that the dog was unable to see them, but the they could just see the very tip of the dog’s nose (to ascertain that the dog was still appropriately stationed/not moving). On the computer runs, the observer used the button-box to present and advance the reward and no-reward stimuli (while simultaneously recording onset/offset).

A laptop computer running Matlab (MathWorks) and Cogent (FIL, University College London) was connected to the button-box via serial port, and recorded both the button-box responses by the observer and scanner sequence pulses.

C-BARQ

Because we hypothesized that variability in participant disposition toward handlers and strangers could have a strong impact on the relative reward value of the two signal types across different source conditions the dogs’ handlers completed the Canine Behavioral Assessment & Research Questionnaires (C-BARQ). These have been used on over 20,000 dogs and represent a standardized and validated (Duffy & Serpell, 2012) tool for obtaining behavioral measures from owner report. The C-BARQ consists of 101 questions asking the respondent to report on how the dog typically responds to common events.

Scores between 0 (minimum) and 5 (maximum) are then computed for 14 behavioral categories. Because of the small number of dogs relative to the number of C-BARQ factors, there was a high potential for factors to be correlated with each other. This collinearity, combined with the number of factors, would be problematic in any modeling of the neural data. Therefore, we performed principal component analysis (PCA) on the 14 C-BARQ scores from our subjects, with a limit of four factors. By limiting to four factors, we were able to include these factors in the neural model without overfitting, while still being able to identify the major groupings of C-BARQ dimensions that were responsible for differences in neural activation.

Functional data preprocessing and analysis

Preprocessing was conducted using AFNI (NIH) and its associated functions, and most steps were identical to those listed in Berns, Brooks & Spivak (in press). In brief, 2-pass, 6-parameter affine motion correction was used with a hand-selected reference volume for each dog. Because dogs moved between trials (and when rewarded), aggressive censoring was carried out, relying on a combination of outlier voxels in terms of signal intensity and estimated motion. Censored files were inspected visually to be certain that bad volumes (e.g., when the dog’s head was out of the scanner) were not included. The majority of censored volumes followed the consumption of food. On average, 51% of total EPI volumes were retained for each subject (ranging from 38% to 60%). This was in line with previous experiments using reward and no-reward signals (Berns, Brooks & Spivak, 2013; Berns, Brooks & Spivak, in press). In addition, we computed the scan-to-scan movement from the AFNI motion files in all three principal directions during exposure to each source (handler, stranger, computer) and condition (reward and no-reward) for each dog for all volumes, including those subsequently censored. Scan-to-scan movement was computed as: sqrt($\mathrm{d}{x}_i^2+\mathrm{d}{y}_i^2+\mathrm{d}{z}_i^2$) where dx_i, dy_i, and dz_i are the changes in corresponding head position of the ith volume (e.g., dx_i = x_i−x_i−1). Scan-to-scan movements greater than 10 mm occurred occasionally if a dog moved out of the field of view. In these cases, the motion estimates were not reliable and so these values were capped at 10 mm to avoid biasing the average. We also computed the proportion of censored volumes during exposure to each source and condition. This allowed us to determine whether unbalanced motion parameters or censoring between reward and no-reward conditions might introduce strong bias into our BOLD findings.

EPI images were smoothed and normalized to %-signal change. Smoothing was applied using 3dmerge, with a 6 mm kernel at Full-Width Half-Maximum (FWHM). The resulting images were then input into the General Linear Model.

For each subject, a General Linear Model was estimated for each voxel using 3dDeconvolve. The task-related regressors in this model were: (1) familiar human reward signal, (2) familiar human no-reward signal, (3) unfamiliar human reward signal, (4) unfamiliar human no-reward signal, (5) computer reward signal, and (6) computer no-reward signal. Because our previous work measuring the hemodynamic response function (hrf) in dogs on this task revealed a peak response at 4–6 s after signal onset (Berns, Brooks & Spivak, 2012), the six task regressors were modeled as impulse functions. Events were convolved with a single gamma function approximating the hrf. Motion regressors generated by the motion correction were also included in the model to further control for motion effects. A constant and linear drift term was included for each run. Finally, to control for possibly confounding physiological factors (e.g., if the reward condition led to more rapid breathing, which might boost neural BOLD signal), a spherical ROI (3 mm radius) was drawn manually on each dog’s structural image in the posterior ventricle, just posterior to the splenium of the corpus callosum. Using AFNI’s 3dmaskave, average timecourses for these ROIs were then extracted after transforming to each dog’s structural space. This timecourse was entered into the general model as a nuisance variable for each subject. Because BOLD signal changes in CSF and white matter reflect physiological effects, not neural processing, this controlled for undue influence of physiological effects on the primary contrasts (Weissenbacher et al., 2009; Murphy, Birn & Bandettini, 2013).

ROIs and mixed-effects modeling

To measure the interaction of reward and no-reward signals with signal source, we used a mixed-effects ANOVA to compare mean caudate activation in the six conditions (2 signals × 3 sources) across all dogs. The four primary temperament factors (as computed from the CBARQ behavioral questionnaire and subsequent PCA discussed above) were also included in the statistical model. To ensure that we were, in fact, measuring caudate activity, we used anatomically defined ROIs. A left and right ROI was drawn on each dog’s structural image (Fig. 3). Then, these ROIs were used to extract average beta values for each dog in each condition from the first-level GLMs, after transforming to each dog’s structural space. The end result was 12 values for each dog (the 6 conditions for both left and right caudates).

Analysis was conducted using the Mixed Models procedure in SPSS v21 (IBM). A 2 × 3 ANOVA was formulated with fixed effects for hand signal (reward, no reward) and source (familiar, unfamiliar, computer) with dog as a random effect. Because no significant left/right differences were observed, side was not included as a fixed effect. We also conducted a second analysis that included the C-BARQ temperament factors as well their interactions with the other fixed effects. This allowed us to determine the relative caudate activation in each of the six conditions and examine how this might be modulated by temperament.

PPI

In addition, we used our data to compute a psychophysiological interaction (PPI) analysis at the individual and group levels. PPI highlights areas that increase functional coupling (defined as synchronous firing patterns) with a seed area in a certain condition, or in a certain condition in contrast to some other (Friston et al., 1997), and has been shown to produce reliable and robust measures of task-specific functional connectivity (Kim & Hortwitz, 2008). PPI allows one to determine which brain areas increase functional connectivity with an area of interest during processing of a specific task.

We computed our PPI in AFNI as described in http://afni.nimh.nih.gov/sscc/gangc/CD-CorrAna.html. First, functionally localized, bilateral caudate seeds were generated for each dog from the allrew–allnorew contrast (that is, across all three source conditions: familiar, unfamiliar, and computer). For each animal, the seeds were generated as spheres with a 3 mm radius centered on the voxel with the greatest differential activation in the left and right caudate respectively.

The activation time series for the left and right caudate seeds were extracted for each subject using 3dmaskave. The time series were then deconvolved using 3dDetrend using a basic gamma function. Next, the interaction of the condition times (reward, no reward, and neutral baseline) and the deconvolved caudate seed time series was computed to produce an interaction term. The interaction term was convolved using the AFNI waver command with a basic gamma function. The convolved interaction term and the initial extracted caudate seed time series were then entered as regressors into the initial GLM (along with the primary task regressors, drift terms, motion regressors), and physiological regressor).

Increased BOLD activation corresponding to the interaction term in the GLM should then be seen in areas that increased functional coupling (i.e., synchronous firing) with the caudate during reward versus no-reward conditions. Such areas may be understood to be differentially connected to the caudate during processing of the reward signal in this task. Note that, as a connectivity analysis, PPI is driven by patterns of increase and decrease in BOLD signal, and is not necessarily indicative of a main effect difference in BOLD strength in any particular region.

Whole-brain group analyses

To apply the transformations to a statistical contrast, the appropriate individual-level contrast (obtained from the GLM as described above) was extracted from the AFNI BRIK file and normalized to template space. Group normalization was conducted using the Advanced Normalization Tools (ANTs) software, as described in Berns, Brooks & Spivak (in press). Briefly, three spatial transformations were computed for each dog: (1) rigid-body mean EPI to high-resolution structural (6 dof); (2) affine structural to template (12 dof); and (3) diffeomorphic structural to template. Spatial transformations were then concatenated and applied to individual contrasts obtained from the above-described GLM model. This allowed the computation of group level statistics. For group-level statistics, a high-resolution MRI beagle brain atlas was used as the template (Datta et al., 2012). AFNI’s 3dttest+ + was then used to compute a t-test across dogs with the null hypothesis that each voxel had a mean value of zero. All twelve dogs were used in this group analysis. The group contrast we conducted was reward—no reward (across all source conditions), computed as the contrast [(Rew_fam + Rew_unf + Rew_com)]−[(noRew_fam + noRew_unf + noRew_com)]. The same approach was taken to compute the group PPI results. Instead of a contrast, the beta values for the interaction term (computed between stimulus presentation schedule and caudate time series) from each individual dog were entered into a second-level model using 3dttest+ +. For both the primary GLM and the PPI analysis, we then calculated the average smoothness of the residuals using 3dFWHMx and then used 3dClustsim to estimate the significance of different cluster sizes across the whole brain after correcting for familywise error (FWE).