Functional MRI of visual responses in the awake, behaving marmoset

Abstract

The visual brain is composed of interconnected subcortical and cortical structures that receive and process image information originating in the retina. The visual system of nonhuman primates, in particular macaques, has been studied in great detail in order to elucidate principles of human sensation and perception. The common marmoset (Callithrix jacchus) is a small New World monkey of growing interest as a primate model for neuroscience. Marmosets have advantages over macaques because of their small size, lissencephalic cortex, and growing potential for viral and genetic manipulations. Previous anatomical studies and electrophysiological recordings in anesthetized marmosets have shown that this species’ cortical visual hierarchy closely resembles that of other primates, including humans. Until now, however, there have been no attempts to systematically study visual responses throughout the marmoset brain using fMRI. Here we show that awake marmosets readily learn to carry out a simple visual task inside the bore of an MRI scanner during functional mapping experiments. Functional scanning at 500μm in-plane resolution in a 30 cm horizontal bore at 7T revealed robust positive blood oxygenation level-dependent (BOLD) fMRI responses to visual stimuli throughout visual cortex and associated subcortical areas. Nonvisual sensory areas showed negative contrasts to visual stimuli compared to the fixation dot only baseline. Structured images of objects and faces led to stronger responses than scrambled control images at stages beyond early visual cortex. Our study establishes functional MRI mapping of visual responses in awake, behaving marmosets is straightforward and valuable for assessing the functional organization of the primate brain at high resolution.

Graphical Abstract

1. Introduction

Vision is the dominant sensory modality of primates, providing spatially accurate information about the environment from a distance. From an ethological point of view, vision bestows organisms with the ability to identify foods, predators, mating candidates, and numerous other objects critical to survival. The high acuity of primate vision also allows for fine motor actions under visual guidance. Mechanisms of vision have long been studied in both human and non-human primates, and these studies have demonstrated that monkeys and humans share many neural specializations for visual perception (Orban et al., 2004; Rosa and Tweedale, 2005; Tootell et al., 2003). Old World macaques have long been the most widely used animal model of human vision, traditionally via anatomical and electrophysiological studies (Felleman and Van Essen, 1991; Hubel and Wiesel, 1968; Van Essen et al., 1992), and more recently in functional MRI experiments (Vanduffel et al., 2014).

Following the advent of new molecular and genetic technologies developed in mice, primate research would benefit greatly from the application of viral or transgenic manipulation of neural systems. As progress in the macaque has been slow, much attention has been turned to the common marmoset (Callthrix jacchus) as a genetically tractable primate model. Marmosets are New World monkeys that diverged quite recently (<40 million years ago) from the phylogenetic lineage leading to humans (Kishi et al., 2014; Mansfield, 2003; Okano et al., 2012; Okano and Mitra, 2014; Solomon and Rosa, 2014). Marmosets are comparable in size to rats, and the combination of small body and large brain offers unique opportunities for high spatiotemporal resolution MRI imaging in a small-bore, high field animal scanner (Hikishima et al., 2013; Newman et al., 2009; Silva et al., 2011). Unlike macaques or humans, the cerebral cortex of the marmoset is lissencephalic, lacking the complex folding of sulci and gyri. Yet in spite of its superficial resemblance to the rat’s brain, the functional organization of the marmoset brain is that of a primate, with specializations in the eye and brain that closely resemble those found in macaques and humans (Cheong et al., 2013; Lui et al., 2013; McDonald et al., 2014; Mitchell and Leopold, 2015; Rosa and Tweedale, 2005; Yu and Rosa, 2014).

In a recent study, we demonstrated the spatial organization of face-selective visual responses in the marmoset visual system (Hung et al., 2015). Here we systematically describe the methodology and results of mapping basic visual fMRI responses throughout brain of the behaving marmoset. For studying visual organization, fMRI offers an important complement to single unit recording. In the awake animal, visual fMRI involves cooperation of the subject, who must reliably direct its gaze to stimuli of interest. We demonstrate here that training and collecting high-resolution fMRI data from awake marmoset subjects in a small-bore scanner is straightforward. Using a block design, we mapped responses throughout the brain to images of natural objects, as well as scrambled control patterns. We found robust positive BOLD responses to visual stimuli in the lateral geniculate nucleus (LGN), pulvinar, superior colliculus and throughout the cerebral cortex, including occipital areas V1, V2, V3 and V4, temporal lobe areas TEO, TE3, and frontal lobe area 8aV. Several nonvisual areas, including primary areas of other sensory modalities, showed suppression of BOLD activity for visual stimuli compared to the fixation baseline. A subset of ventral stream visual areas, including areas V3, V4, TEO, and TE3, responded more strongly to natural, structured images than to the corresponding scrambled control images. These results show that fMRI is an effective tool for mapping cortical and subcortical visual responses in this species. Thus high-resolution fMRI, when combined with specific molecular and genetic interventions in the marmoset, offers new avenues to probe the functional organization of the primate brain.

2. Material and Methods

2.1 Subjects

Two healthy adult male common marmosets (Callithrix jacchus), marmoset E and marmoset B, each of them six years old and weighing 350–450 g, were used in the study. Marmosets were housed in cages with same-sex pairs in a room with a 12 h light/dark cycle. Their food and water intake was regulated, receiving food and at least 15 mL of unfiltered water once daily from 2 to 4 pm (Lunn, 1989). This regulated feeding schedule induced the animals to be more engaged during behavioral training and fMRI experiment sessions. All procedures were approved by The Animal Care and Use Committee of the National Institute of Neurological Disorders and Stroke/National Institute on Deafness and Other Communication Disorders.

2.2 Behavioral training and testing

Over the course of 3 weeks, animals were progressively acclimated to head and body restraint in the sphinx position as detailed in our previous work (Silva et al., 2011). Individualized custom-built helmets restrained head motion, while a semicylindrical cradle restricted whole body motion (as shown in Fig. 1). For the purpose of measuring responses to visual stimuli, we trained the animals to maintain their gazes on the stimuli. In this work, the animals were intentionally not trained to fixate on a small dot but free viewing the stimuli, though is has been proven feasible in other recent marmoset work (Mitchell et al., 2014). Over several weeks, the subjects learned to complete a visual task paradigm in which they were required to maintain their gaze within a 5°-radius window, corresponding to the size of the visual stimuli. Eye position was monitored using a video-based eye tracking system: Eyelink II (SR Research, Mississauga, ON, Canada). The center and gain of the eye signals were calibrated by training the animals to make saccades to fixation dots placed at five different positions on the screen. During the initial stage of behavioral visual training, the marmosets completed trials consisting of the presentation of one visual stimulus (from the stimulus categories specified below) for the duration of 1.5 seconds on an LCD screen (Tech Video System Co., Ltd, Suzhou, China). Correct gaze behavior was reinforced with a very small drop (~0.02 mL) of sugary liquid reward at the end of each trial. During the second stage of behavioral visual training, subjects learned to maintain their gaze within the defined tolerance range of 5° radius circle during a block design paradigm, where a sequence of images within a category or only the fixation dot were shown, consisting of one image every 0.5 seconds for 16 seconds.

Figure 1.

Visual fMRI setup for awake marmosets. A. Animals were restrained by a custom built helmet and jacket non-invasively in sphinx position in 7T horizontal scanner. They were trained to hold their gaze in the center portion of an LCD screen for a 16 s blocks while stimuli were presented. The subject’s gaze was tracked by a video-based eye tracking system with an MR-compatible camera. MonkeyLogic, an open source Matlab toolbox, was used to monitor the animal’s performance and control dispensing of the liquid reward through a plastic tube. An 8-channel receive-only array was built on the inner surface of the helmet (bottom left) to allow MRI to be obtained with high signal-to-noise ratio (SNR). B. Three representative coronal, axial, and sagittal views of the brain of marmoset E show high SNR in most cortical areas. The color scale indicates the SNR at each measured voxel.

Experimental testing using this block design paradigm consisted of 16 s blocks of stimulus presentation interleaved with 20 s inter-blocks interval of quiet rest in which there were no behavioral requirements (Fig. 2A). Stimulation blocks began with the appearance of the fixation point and a check of the animal’s gaze position. Failure to acquire and maintain gaze within the fixation window for the first 1.5 s immediately terminated the block. Within a block, drops of rewards were given every 1.5 seconds if gaze was maintained.

Figure 2.

Visual fMRI block design experiment paradigm and stimuli. A. Randomized stimuli within a category were presented every 500ms without gap during a 16 s block while EPI images were acquired every 2 s. A long, 20 s inter-block-interval was used during which time there were no particular behavioral requirements. Eye signals (blue trace) were recorded throughout the block and rewards were delivered if the subject maintained the gaze within the tolerance range (red-dashed lines). B. The stimuli set included structured categories (conspecific faces, conspecific body parts, and objects) and scrambled categories (spatial and phase scrambles of the face stimuli) as well as a fixation dot only condition. Within each block, twenty examples from a particular category were presented in randomized order.

2.3 Visual Stimuli

Six different categories of visual stimuli were used: conspecific faces, conspecific body parts, objects, spatial-scrambled faces, phase-scrambled faces, and a fixation dot only condition. A small cyan color fixation dot of 0.25° radius is presented along with the stimuli throughout the blocks. In case of the fixation dot only condition, only the 0.25° radius dot is presented. The images of marmoset faces, marmoset body parts, and objects, were from pictures taken from the colony. The contrast of the images was enhanced by histogram equalization and the luminance was calibrated to that of a randomly selected face image. To compare the visual activation of intact, structured, objects versus low-level visual features, we generated two kinds of scrambled stimuli. The spatial-scrambled stimuli were created by tiling the face images into 15 × 25 little squares and shuffling their positions. The phase-scrambled stimuli were generated by permuting the phase information of the spectrum of the original images. Twenty examples within a category were randomly selected and shown in each block (some are shown in Fig. 2B). The stimuli subtended a 5° visual angle.

2.4 Embedded eight-channel surface receiver coil

For each marmoset, an 8-element receive-only coil array was built using a flexible material, CuFlon (Polyflon Inc., Norwalk, CT, USA), with a 0.25 mm thick layer of copper deposited on a single side at 2 oz/ft². This thin and flat material allowed the embedding of each coil element on the inner surface of a 3D printed helmet. This helmet served to restrain head motion and thus was customized for each animal based on the animal’s 3D head profile. Additionally, a layer of 3-mm thick polyurethane foam covered the coils and helmet providing extra comfort to the animal during the experiments. The inner diameter of each circular loop was 12.5 mm and the copper width was 1.5 mm. The coil elements were arranged into two frontal rows of 3 elements overlapped in the anterior-posterior direction to cover the frontal and parietal cortices, and one back row of two elements overlapped in the left-right direction to cover the visual cortex, as shown in Fig. 1A. The coil circuitry consisted of a matching network and a PIN diode controlled blocking circuit for active detuning (Papoti et al., 2013), providing isolation better than 30 dB between the transmit RF coil and receive array. Decoupling between non-neighbors elements in the receive array was achieved by connecting each element to an 8-channel low input impedance preamplifier box (Dodd et al., 2009; Nascimento et al., 2008), resulting in isolation better than 20 dB due to preamp decoupling.

In vivo SNR maps (Kellman and McVeigh, 2005) were obtained from conscious awake marmosets using the 8-channel embedded coil and are displayed in Fig. 1B. The acquisition parameters for the SNR maps calculation were: FLASH sequence, TR/TE = 700/5.5 ms, FOV = 3.84 × 3.84 cm², slice thickness = 2 mm, matrix 256 × 256, 3 averages (scan time 8m 57 s).

2.5 fMRI procedures

All scans were performed in a horizontal 7T/30cm MRI spectrometer (Bruker-Biospin Corp., Billerica, USA). Two individualized helmets, each containing an embedded eight-channel surface coil array, as described in previous section were used to restrict head movement. Using a gradient-recalled EPI sequence with 18 axial slices (TE/TR= 26/2000ms; FOV/slice thickness= 32×32/1mm³; matrix= 64×64), blood oxygen level dependent (BOLD) functional images were acquired (512 volumes per run). Coplanar RARE T₂-weighted anatomical images (TE_effective/TR= 64/4000ms; FOV/slice thickness= 32×32/1mm³; matrix= 128×128) were collected each session for image registration.

Visual stimuli were presented in a block design paradigm as in the training sessions described above. During fMRI acquisition, eye position was monitored using the video-based eye tracking system iViewX (SensoMotoric Instruments, Teltow, Germany), with an MR-compatible camera and infra-red light source (MRC Systems, Heidelberg, Germany). Three to six 17-minute runs of EPI sequence were conducted in a session. Each run consisted of up to 28 blocks, depending on the animals’ performance.

2.6 Creating anatomical regions of interests (ROIs) from an MRI template brain

High-resolution, 0.166mm isotropic, T₁-weighted volumes were acquired in five marmosets in vivo (Bock et al., 2011). These volumes were skull-stripped, registered, and then averaged. The resulted brain volume served as a standard template on which functional data can be presented. To delineate the cortical and subcortical ROIs on this high-resolution MRI volume, we registered ROIs defined by Paxinos et al. in The Marmoset Brain in Stereotaxic Coordinates (Paxinos et al., 2011) onto the template.

Cortical ROI segmentation based on Paxinos’ atlas is publicly available online (http://www.3dbar.org:8080/getCafInfoPage?cafDatasetName=mbisc_11). This histology-based volume is from the left-hemisphere only; therefore, we mirrored the volume to create a right-hemisphere copy to provide bilateral coverage. To register this histological volume to the MRI template brain, we first masked the whole left and right cortex using ITK-snap (Yushkevich et al., 2006). Using nonlinear transformation algorithm 3dQWarp of AFNI (Cox and Hyde, 1997), we registered the left and right histological volume, containing annotations of cortical ROIs, onto the cortical masks of the MRI template respectively. We further segmented the primary visual cortex (V1) into two portions: the portion on the lateral surface, which corresponds to ~10° of the central vision, and the portion embedded in the calcarine sulcus, which corresponds to more peripheral visual fields. This segmentation help us differentially summarize the response to our visual stimuli, which subtended a 5° visual angle, in the central and peripheral representation of the primary visual cortex.

To specify subcortical regions of interest, we manually marked potentially relevant subcortical regions of interest (i.e. lateral geniculate nucleus (LGN), superior colliculus, pulvinar, caudate, putamen, hippocampus, amygdala, hippocampus, and dentate gyrus) using coronal sections ranging from inter-aural coordinate −0.70mm to +14.00mm in the Paxinos’ atlas. The individual atlas images were compiled into an image stack using ImageJ (Schneider et al., 2012). These subcortical ROIs were carefully labeled manually using ITKsnap. Nine cortical ROIs: anterior intraparietal area of cortex (AIP), middle temporal (MT), temporal area TF, occipital part (TFO), rostral auditory cortex (AuR), piriform cortex (Pir), amygdalopiriform transition area (APir), parietal area PG, medial part (PGM), area 23c of cortex (A23c), and temporal area TH (TH), distributed across the cortex, were also labeled to facilitate the registration of the manually-drawn ROIs to the MRI template brain. Two stages of registration were performed. First, a user-guided, coarse transformation was determined manually using tkregister2, a function of freesurfer (Dale et al., 1999). Second, we found an affine transformation using FLIRT, an FSL function (Jenkinson et al., 2002), by registering the 9 manually-drawn cortical ROIs to the corresponding Paxinos’ cortical ROIs, which were already registered to the MRI template brain. The subcortical, manually labeled ROIs were then transformed into the MRI template space using the affine transformation.

2.7 fMRI Data Analysis

Twelve and 9 MRI sessions were collected from marmoset E and marmoset B, respectively. Using AFNI, motion-correction and cross-session alignment were performed (Cox and Hyde, 1997). Blocks in which the subject’s gaze was outside of a 5°-radius window for ≥ 20% of the duration were excluded. Furthermore, volumes with motion larger than 0.02 mm for marmoset E and 0.1 mm for marmoset B were censored. With this threshold, we censored 5.1% and 2.0% of all the collected volumes in Marmoset E and Marmoset B, respectively. The coefficient of variance (CV) was calculated for each voxel in each session. Voxels with high CV values, indicating low signal-to-noise ratios, were censored. In marmoset E the threshold for this CV censoring was 0.04, whereas in marmoset B it was 0.06. On average, 62.79% ± 4.93% (mean ± standard deviation across sessions) voxels were censored for Marmoset E and 45.89% ± 5.93% for Marmoset B. Most of the censored voxels are within air or tissue outside the brain. Focusing on the brain, only 20.04% ± 5.56% voxels were censored for Marmoset E and 5.93% ± 3.89% for Marmoset B. These voxels were mainly located at the most posterior or ventral cortical areas where we had a lower Signal-to-Noise ratio due to susceptibility issues. The functional images from both animals were spatially smoothed with a Gaussian kernel of 1mm full width half height and transformed onto the T₁-weighted MRI template space for further data analysis and presentation.

Brain volumes, collected between 4 and 18 seconds after the 16-s block onset, were analyzed for visual responses. Statistical tests contrasting two sets of conditions during this period were calculated using two-sample t-tests with correction for multiple comparisons (false discovery rate q = 0.05) (Benjamini and Hochberg, 1995). The map of visual responsiveness was calculated by contrasting all five categories of stimuli versus the fixation point condition. The map of natural, structured selectivity was generated by contrasting responses to three intact, structured categories (conspecific faces, body parts, and objects) versus the two scrambled control image sets.

The time courses of average BOLD signal changes within an anatomical ROI were calculated by averaging the responses across voxels within the ROI defined on the template volume. A 3D brain surface model was generated from the T₁-weighted template brain using Caret (Van Essen, 2012). We then projected the functional contrast maps onto this surface using an algorithm from Caret, which samples a weighted average of the functional data from nearby voxels on each vertex (Van Essen, 2012). Contrast maps in volumetric and 3D-surface views were rendered using custom codes written in Matlab (The Mathworks, Inc., Natick, USA).

2.8 Structure selectivity index (SSI)

Structure selectivity index (SSI), which indicates the preference of natural, structured images over the scrambled controls, is defined as:

$$\mathbf{SSI}=\frac{R_i-R_s}{\left|R_i\right|+\left|R_s\right|}$$

, where R_i is the maximal BOLD responses among intact, structured images including faces, body parts, and objects relative to the fixation point condition and R_s is the maximal response among the two scrambled conditions relative to the fixation point condition. The index ranged from −1 to 1, with larger indices indicating a stronger preference for structured stimuli. The SSI is calculated for each voxel. For anatomical ROIs, the SSI of the ROI is the averaged SSIs of all voxels within the ROI.

3. Results

3.1 Development of visual fMRI setup for awake marmosets

The visual fMRI setup for awake marmosets is depicted in Figure 1A. For testing the marmoset inside the scanner, we adapted an experimental approach that had previously been used in macaques (Tsao et al., 2003), in our case with the marmoset lying belly down and restrained within a jacket fixed firmly to a cradle inside the bore. Unlike most macaque studies (but see (Srihasam et al., 2010)), the marmosets in this study did not receive any implants to restrain head motion. Instead, the animals were acclimated to have their heads firmly held using the customized 3D-printed helmets. Thus all data was collected noninvasively (Silva et al., 2011). As it is important in visual experiments to accurately track the eye position, we monitored the animals’ gaze in real-time using a video-based eye-tracking system. For this, we employed an MR-compatible high-magnification camera positioned 13 cm from the monkey inside the bore, with the eye illuminated by an infrared light source. The animals were trained to maintain their gaze on the stimuli presented on a display just outside the end of the bore. For the present study, our objective was not to attain strict fixation from the animals, though a previous study has shown that precise measure of eye position is possible in marmosets (Mitchell et al., 2014). The eye-tracking signals were relayed to a central computer running MonkeyLogic software, a MATLAB toolbox (Asaad et al., 2013), which controlled the visual display and dispensed liquid rewards if the animals’ gaze was maintained within the defined tolerance range of 5° radius circle.

Individualized helmets were 3D-printed to fit each animal’s head profiles. Eight-channel receive-only coils were made with flexible material, CuFlon, and were embedded within the helmet. The signal-to-noise ratio (SNR) maps obtained using the 8-channel receive-only embedded array are shown in Fig. 1B. The axial, sagittal and coronal images show whole brain coverage with high sensitivity throughout the entire cortical surface, except at the very frontal cortex, which falls outside of the restraint helmet. The placement of the coil elements inside the helmets minimized the distance from the coil to the animal’s head, consequently increasing the SNR near the cortex surface. The combination of overlapping nearest neighbor elements and decoupling due to low input impedance preamplifiers provided good coil-to-coil isolation, allowing the acquisition of images with acceleration factor of up to 2 without significantly compromising the SNR.

We measured whole-brain fMRI responses to assorted categories of visual stimuli presented in a block design paradigm (Fig. 2A). Each visual stimulation block lasted for 16 seconds; random example stimuli within a single category were presented and switched every half-second. A longer, 20-second, inter-block-interval was placed in between stimulation blocks to allow the hemodynamic responses to return to baseline. We acquired BOLD fMRI responses with echo-planar imaging (EPI) at a spatial resolution of 0.5×0.5×1.0mm (1.0mm in dorsal-ventral axis) and a temporal resolution of 2 s. Very small drops (~0.02 mL) of sugary liquid rewards were delivered every 1.5 s if the subjects maintained the gaze within the stimuli on the screen

The visual stimuli consisted of six categories (Fig. 2B). Three of the categories were structured natural photographic images of conspecific faces, conspecific body parts, and man-made objects familiar to the animals. Two other control categories of spatial and phase scrambled images were created based on the face stimuli, with the intention of preserving the low-level visual features. A baseline fixation condition consisted of only a fixation dot was presented throughout the block, with the animals rewarded as in the other visual stimulation conditions if they maintained their gaze within the tolerance window.

3.2 Visual responses in widespread cortical visual areas

To examine the overall level of visual responsiveness within the marmoset brain, we contrasted the BOLD signals of all five visual stimuli conditions: faces, body parts, objects, and two scramble stimuli, against the fixation dot only condition (Fig. 3). We observed stronger BOLD responses to the visual stimuli compared to the fixation dot only condition in multiple cortical and subcortical structures, as illustrated in representative axial, parasagittal, and coronal views (Fig. 3A–D). The most salient feature is the continuous swath of activation extending from occipital to temporal lobe. The data are shown using a conjoint color and transparency scale to map the t-value and at the same time allow for visualization of the underlying anatomy. The t-value of the two-sample t-test corresponding to p = 0.05 after multiple comparisons correction for all voxels within the brain was marked by the dashed line in the color bar and was represented by a very dim transparent color on the map.

Figure 3.

Cortical BOLD responses to visual stimuli. A. 3D model of the template marmoset brain from the right lateral view (top) and front view (bottom). The three lines indicate the position of the representative axial, parasagittal, and coronal slices on the right. B–D. Three representative slices showing the contrast map of all visual stimuli versus fixation baseline condition in marmoset E. The dashed line on the color bar indicates p < 0.05 after multiple comparisons correction for all voxels within the brain. The green dashed line indicates the lateral geniculate nucleus. E. The same contrast of all stimuli versus fixation, plotted on the 3D model of the right and left hemispheres in both animals. Clear positive contrast was observed in the occipitotemporal visual pathway. The auditory and somatosensory cortex showed areas of negative contrast, indicating lower activity for visual stimuli compared to the fixation baseline. F. The average time courses of the BOLD signals in selected anatomical region of interests (area V1c (central visual field representation of V1), V4, 8aV, and A1). The red line is the average time courses of all blocks from the five different visual stimuli conditions; the green line is the average time courses of the fixation dot only condition. The error bar indicates one standard error across blocks. The transparent blue/gray rectangle indicates the 16 s stimulation period.

To better visualize the widespread cortical visual activation in both animals, 3D surface representations of both the left and the right hemispheres were constructed (Fig. 3E). Since marmosets possess a lissencephalic cortex, most of the brain’s surface was visible without inflating and potentially distorting the brain. Widespread cortical activations in occipital, temporal, and frontal areas were consistently observed in both animals. Our EPI images suffered from signal drop-offs near the occipital pole and ventral temporal areas because of susceptibility issues. This led to lost information in the occipital pole and the more ventral portion of the temporal areas on the activation maps. Since the central vision representation of V1 is at the most anterior boundary, abutting area V2, on the lateral surface, where we had good signal-to-noise ratio, we were able to adequately assess functional responses in V1. The cortical visual responses spanned over most of the ventral stream extrastriate visual areas, including V2, V3, V4, TEO, and TE3. We also observed responses in frontal areas including area 8aV, some or all of which likely corresponds to the frontal eye fields (FEF) in macaque (Blum et al., 1982; Burman et al., 2006; Reser et al., 2013).

To quantify how visual responses vary in different cortical areas and subcortical structures, we defined areal boundaries in the cortex (Fig. 3E) based on a transformation of the Paxinos’ marmoset atlas (Paxinos et al., 2011) onto the MRI template space (see Methods for details). This revealed marked activation along many areas of the occipitotemporal pathway and made it possible to quantify the average responses in specific areas of interest. The time courses of BOLD responses in representative areas of the visual cortex, auditory cortex, and prefrontal cortex are shown in Fig. 3F. For example, in the central representation of V1 (V1c) and area V4 in the visual cortex, we observed clear responses to all visual stimulus conditions, but minimal or no responses to the fixation baseline condition. In the prefrontal cortex, activity was stronger during visual stimulation, but was also apparent during the fixation baseline condition, possibly reflecting components of visual attention shared across conditions. Unexpectedly, we also noted small but robust BOLD responses in the primary auditory cortex (A1) that were lower during the visual stimulus conditions than during baseline. This effect, which might be interpreted as suppression by a visual stimulus relative to the fixation baseline, was prominent outside the visual cortex in both animals (Fig. 3E), primarily affecting primary and secondary auditory and somatosensory areas. This effect may also be related to cross-modal deactivation that has been reported to occur when the subjects allocate attention to visual stimuli (Bense et al., 2001; Mozolic et al., 2008).

3.3 Visual responses in subcortical visual structures

In all three representative slices of Fig. 3B–D, subcortical activation is clearly visible. To better characterize and quantify the subcortical visual responses, we annotated several relevant subcortical structures based the Paxinos’ atlas (see Methods for details). In Fig. 4A, we highlight three subcortical structures: LGN in red, pulvinar in yellow, and superior colliculus (SC) in cyan, which are all critical for visual processing. We found robust visually evoked BOLD responses in each of these structures, as shown in several representative slices in Fig. 4B–F. The corresponding time courses of the BOLD signals averaged across all voxels within the anatomical ROIs show clear time-locked hemodynamic responses that peaked 4 s after stimulus onset and lasted for the duration of the blocks (Fig. 4G). The visual stimulation condition was higher than the fixation baseline, although the latter condition led to increased responses as well, perhaps due to the background illumination of the monitor, the fixation dot itself, or a change in the animal’s behavioral state with the initiation of each block. Closer examination reveals that LGN activation was strongest at the posterior portion, corresponding to its central visual field representation (White et al., 1998). The visual stimulus-selective responses in SC located within the rostral portion of the superficial layer, which is also consistent with visually evoked responses from the central visual field (Kaas and Huerta, 1998).

Figure 4.

Subcortical BOLD activation to visual stimuli. A. Lateral geniculate nucleus (LGN), pulvinar, and superior colliculus (SC) colored with red, yellow, and cyan respectively, were superimposed on the 3D model of the template marmoset brain. The top panel shows the right lateral view and the bottom panel shows the frontal view of the brain. The lines indicates the positions of the slices in panel B–F. B. The contrast map represents all visual stimulus conditions versus the fixation baseline in marmoset E. The color map convention is the same as in Fig 3. This axial slice shows bilateral activation in LGN with a line indicating the slice in D. C. Coronal slice at +2.0mm inter-aural shows activation in both superior colliculus and pulvinar. Two lines show the position of the slices in E and F. D–F. The parasagittal views of the activations in LGN, superior colliculus, and pulvinar. The anatomical ROIs are outlined in colors(red - LGN, yellow - pulvinar, cyan - SC). G. Average time courses of the BOLD signals for visual stimuli (red) versus fixation (green) in the LGN, pulvinar, and SC.

3.4 Selectivity for natural, structured stimuli in higher visual areas

We next assessed the selectivity of natural, structured visual stimuli by contrasting the intact photographic images versus spatially and phase-scrambled versions of the face stimuli. The contrast maps (Fig. 5B–D) demonstrated a pattern of selectivity for structure, particularly in a subset of rostral extrastriate visual regions. In assessing the BOLD time courses for each stimulus condition in several visual areas (Fig. 5E), it was apparent that coherent visual structure was more important for higher visual areas (V3, V4, TEO, and TE3) than in the earliest retinotopic areas (V1, V2). Importantly, in the highest visual areas, scrambled stimuli are approximately as ineffective as the fixation baseline condition for driving visual responses. We further quantified the selective responses to natural, structured images over the scrambled controls using the structure selectivity index (SSI) (see Methods) commonly adopted in visual neurophysiology experiments. We found that the SSI generally increases along the posterior-to-anterior axis along the occipitotemporal cortex in both subjects. We calculated the average SSIs for different anatomical ROIs along the ventral visual stream, which presumably relates to visual object shape processing. This revealed an increase in structure selectivity across this visual hierarchy, with the lowest SSIs (close to zero in marmoset B indicating a no preference of structured versus scrambled images) in V1 and V2, intermediate SSIs in V3 and V4, and the highest SSIs in TEO and TE3 (Fig. 5F).

Figure 5.

Selective responses to structured versus scrambled images in higher cortical visual areas. A–D. The contrast map of structured versus scrambled stimuli shows activation of rostral extrastriate visual areas. The color convention is the same as in Fig. 3. E. The BOLD signal time courses of visual areas along the ventral stream (V1c, V2, V3, V4, TEO, and TE3) for each individual stimuli category. The color code of the curves are assigned at right panel, where face =red, body = orange, object = yellow, spatial scrambles = dark blue, phase scrambles = light blue, fixation point alone = green. The responses to the structured and the scrambled images separated in the more rostral higher order visual areas F. Structure selectivity index (SSI) reveals a gradient of increasing structure selectivity along the occipitotemporal pathway. Average SSIs of the anatomical ROIs along the occipitotemporal visual pathway: area V1, V2, V3, V4, TEO, and TE3, are calculated in both subjects. It reveals a trend of higher SSIs in higher visual areas along the ventral pathway.

4. Discussion

4.1 Visual fMRI experiment in awake, behaving marmosets

In this work, we further established the feasibility of conducting visual fMRI experiments in awake, behaving marmosets. The procedures were non-invasive and thus carried out in animals without the need for head implants. It has been previously speculated that marmosets are difficult to train and will not cooperate in tightly set psychophysics experiments such as the ones that are routinely performed by macaques. Recently, researchers have reported success in training the marmosets to perform visual (Mitchell et al., 2014) and auditory (Osmanski et al., 2013; Remington et al., 2012) discrimination tasks. Our present results provide another example that head-restrained marmosets can engage in visual tasks just as well as Old World monkeys within an fMRI setup.

The anatomy, physiology, and function of marmoset vision is generally similar to that of other primates, including humans, as summarized in recent review articles (Mitchell and Leopold, 2015; Solomon and M. Rosa, 2014). Previous electrophysiology studies using anesthetized marmosets have provided detailed knowledge about the organization of the marmoset’s visual system (Cheong et al., 2013; Lui et al., 2013; McDonald et al., 2014; Rosa and Tweedale, 2005; Yu and Rosa, 2014). Our study shows that fMRI experiments in the behaving animal can be used to complement electrophysiological studies. This approach has certain advantages, such as the capacity to assess activity simultaneously throughout the brain and the elimination of potential confounds introduced by anesthetic agents (Liu et al., 2013).

Recently, our lab has established experimental conditions for conducting fMRI experiments in awake marmosets during resting state (Belcher et al., 2013) as well as somatosensory stimulation (Liu et al., 2013) experimental paradigms. This was achieved by using a customized 3D-printed helmet, which provided a tight fit on each animal’s head. The helmet not only constrains the animal’s head, but also serves as a medium for embedding receive-only RF coils near the animal’s brain to increase the signal-to-noise ratio. One drawback is that compared to traditional head-post implantation, the use of helmets as a restraint device allows some small head motion occasionally. These motion artifacts, nevertheless, could either be corrected for, or censored, in post-processing routines as is standard in human studies. By successfully training the animals to actively maintain their gaze on the screen using positive reinforcement learning to obtain rewards, we extended our fMRI measurement capabilities to the visual system.

One major advantage of our setup is that the animals are placed in the MRI bore in the sphinx position, which is a posture that marmosets assume naturally, specially when they are resting. Thus the animals have a natural body position, and they are looking out the MRI bore, so that axial images of the brain are acquired in the X-Z Cartesian plane. We have previously shown an overall increase in T2* in the brain of marmosets when animals are placed in the sphinx position versus the conventional human supine position due to the parallel orientation of major white matter fibers in the optical radiation relative to the direction of the main magnetic field B₀ (Sati et al., 2012). Thus the sphinx position has also added benefits of higher image quality and less distortion and signal dropouts due to susceptibility, especially at high field strengths. However, one region that does suffer from susceptibility effects is the occipital pole of the brain, as the edge of the pole is perpendicular to the direction of the main magnetic field B₀. The helmet construction enhances this susceptibility effect to some degree, due to the edge created in the back of the head. This increased susceptibility reduces data quality in the most posterior portion of V1. To solve this problem, we would either need to modify the animals’ posture, and/or to improve the helmet design, and this is a goal for future work. Alternatively, we could combine the use of intravenous blood volume contrast agents, such as monocrystalline iron oxide nanoparticles (MION), with a pulse sequence with shorter echo time TE, to get better EPI image quality in this area.

4.2 Robust visual responses in cortical and subcortical visual areas

We observed visual responses in subcortical areas, striate cortex, and multiple extrastriate visual areas. We found robust activations in three main subcortical areas: the superior colliculus (SC), the lateral geniculate nucleus (LGN), and the pulvinar. The SC is a hub of sensory and motor processing in the dorsal midbrain, which can detect salient visual input and generate eye movements (Boehnke and Munoz, 2008). The SC receives direct retinal projections and wide ranges of cortical projections from visual areas to superficial layers and from visuomotor areas to deeper layers (Collins et al., 2005; Wurtz and Albano, 1980). The LGN receives primary outputs from retina and feedback signals from V1 (Sherman and Guillery, 2002; Weller and Kaas, 1989). The pulvinar is a part of the thalamus that is particularly well developed in primates. The pulvinar is a higher-order relay center that coordinates activity between cortical areas through cortico-pulvinar-cortical connections (Kaas and Lyon, 2007; Sherman and Guillery, 2002). The cortical visual activation spans a big portion of the occipitotemporal cortex, including V1, V2, V3 (also named ventrolateral posterior, VLP (Rosa and Tweedale, 2005)), V4 (ventrolateral anterior, VLA), V4t, TEO, and TE. Moreover, we found visual responses in the frontal lobe, mainly in area 8aV and area 45; some of these areas are presumably the homolog to the frontal eye fields (FEF) in macaques (Blum et al., 1982; Burman et al., 2006; Reser et al., 2013). On the other hand, the extrastriate visual areas of the dorsal stream within the parietal lobe did not show significant responses to the stimuli. This may due to the use of static images, which are not the most suitable stimuli to activate the dorsal visual pathway in the parietal cortex (Kravitz et al., 2011).

To our surprise, we observed a slightly negative contrast in some cortical areas outside the visual cortex when contrasting the responses to all visual stimuli versus the fixation condition baseline (Fig. 3E). These negative contrast regions were located mainly within the primary and secondary auditory and somatosensory areas. A similar finding has been previously reported in a human fMRI study, where visual stimuli ‘deactivate’ the auditory cortex, and the auditory stimuli ‘deactivate’ the visual cortex (Laurienti et al., 2002). Moreover, it was also reported that selective attention to one sensory modality can ‘deactivate’ other modalities (Bense et al., 2001; Mozolic et al., 2008). In our case, all five different visual stimuli blocks provide stronger visual input, and likely demand more visual processing resources and selective attention when compared to the fixation baseline condition. This explained why we observed the negative contrast in the other sensory modalities.

4.3 Structure selectivity gradient along occipitotemporal cortex

When contrasting the visual responses to natural structured images versus the ones to scrambled controls, we demonstrated a progressive increase in selectivity for structured natural images in higher visual areas along the posterior-anterior axis of the occipitotemporal cortex. This contrast of structured versus scrambled images has been used in human fMRI studies to identify a region anterior and lateral to V4, namely the lateral occipital complex (LOC) (Grill-Spector et al., 1999; Kourtzi and Kanwisher, 2001; Malach et al., 1995). A follow up study showed that the structure selectivity actually presented a posterior-anterior gradient from early retinotopic areas to the LOC (Lerner et al., 2001). Neurophysiology studies in macaque also have revealed an increased preference for complex and natural stimuli across the visual hierarchy (Kobatake and Tanaka, 1994; Logothetis and Sheinberg, 1996; Tanaka, 1996). Here, we report a comparable gradient of image structure selectivity from earlier visual areas (V1, V2) to intermediate extrastriate visual areas (V3, V4) and finally to the higher visual areas in the inferotemporal cortex in the ventral pathway of the marmoset. We previously observed such a gradient of structure selectivity in high gamma-range local field potential power using electrocorticography (ECoG) (Hung et al., 2015). In that work, we showed that not only were anterior channels more responsive to structured stimuli, but that posterior channels were more responsive to scrambled images (see Fig. 4A in (Hung et al., 2015)). Our fMRI results here agree with the ECoG data and provide a more comprehensive view of this gradient of increasing structure selectivity because of the powerful feature of whole-brain imaging. Overall, this finding is compatible with our knowledge of the macaque and human visual brain, and supports the view that the functional organization of the ventral visual stream is similar in marmosets and Old World primates (Mitchell and Leopold, 2015; Solomon and Rosa, 2014).

4.4 Delineation of Cortical Visual Area Boundaries

We determined the cortical areal boundaries based on the morphed Paxinos’ atlas on the template brain. As a coarse approach, this type of segmentation could be applied to all parts of the cortical areas (visual, somatosensory, auditory, prefrontal etc.), though it serves only as an approximation. As there is variation in the structure of individual brains, as well as in the position of areal boundaries, other methods would be required to achieve greater precision. While the issue of individual variation in this species is not well characterized, a previous MRI study mapping specific areas based on cortical myelin content showed notably little variation in the size of areas across several animals (see Table 1 of (Bock et al., 2011) for details). The future application of this method to individual animals will provide a more accurate picture of areal extents, as will retinotopic mapping. The latter process is routinely practiced in humans and macaques but, as it relies on extended periods of fixation at which marmosets do not excel, has not yet been proven in this species. In our previous study (Hung et al., 2015), we achieved a coarse approximation of areal boundaries by presenting flashing checkerboards along the cardinal meridians while the animals was require to look at a very small movie shown in the center of the screen. This approach allowed for an approximation of the anterior boundary of V1 as well as the identification of foveally responding voxels. Whether it is possible to achieve a more complete retinotopic mapping in the awake marmoset by extending fixation or possibly updating stimulus presentation based on shifts in gaze position remains a question for future research.

4.5 Potentials of awake marmoset fMRI for basic and translational research

The marmoset is increasingly recognized as a promising primate model for biomedical research, particularly given its demonstrated potential for genetic manipulations(Kishi et al., 2014; Okano and Mitra, 2014). The creation of transgenic marmosets with germline transmission (Sasaki et al., 2009) makes it possible to generate transgenic animal lines for both basic and translational research. Both in basic and translational research, anatomical and functional MRI methods will be important tools to access the neurophysiological status in animal models of neurological and neuropsychiatric disorders (Hikishima et al., 2015; Jagessar et al., 2014). Our findings of robust visual BOLD responses in awake, behaving marmosets demonstrate that, just as in humans, it is straightforward to measure aspects of functional brain organization using fMRI. This tool, when combined with the increasing repertoire of genetic and molecular manipulations available to marmosets, points to a powerful new approach for studying the functional principles of the healthy and diseased brain.