Resting-state fMRI data of awake dogs (Canis familiaris) via group-level independent component analysis reveal multiple, spatially distributed resting-state networks

Resting-state networks are spatially distributed, functionally connected brain regions. Studying these networks gives us information about the large-scale functional organization of the brain and alternations in these networks are considered to play a role in a wide range of neurological conditions and aging. To describe resting-state networks in dogs, we measured 22 awake, unrestrained individuals of both sexes and carried out group-level spatial independent component analysis to explore whole-brain connectivity patterns. In this exploratory study, using resting-state functional magnetic resonance imaging (rs-fMRI), we found several such networks: a network involving prefrontal, anterior cingulate, posterior cingulate and hippocampal regions; sensorimotor (SMN), auditory (AUD), frontal (FRO), cerebellar (CER) and striatal networks. The network containing posterior cingulate regions, similarly to Primates, but unlike previous studies in dogs, showed antero-posterior connectedness with involvement of hippocampal and lateral temporal regions. The results give insight into the resting-state networks of awake animals from a taxon beyond rodents through a non-invasive method.

image acquisition. Functional MRI acquisitions were performed on a Siemens Prisma 3 T scanner (Siemens Healthcare, Erlangen, Germany) using a Gradient Echo Echo Planar Imaging (GRE-EPI) sequence with TR = 2640 ms including a 500-ms delay at the end of each volume, and TE = 30 ms. We decided to introduce the delay for welfare reasons. During piloting we realized that our dogs seem to find continuous scanning less comfortable than protocols with short gaps between subsequent scans. The 500 ms delay was selected empirically during piloting to include the shortest gaps that the dogs are still comfortable with. The protocol had an in-plane Field-of-View 128 mm × 128 mm, using 2 mm in-plane resolution and 2 mm slice thickness, measuring 31 slices with an inter-slice gap of 0.5 mm, using an excitation flip angle of 86°. Phase-encoding direction was set to left-right. A single loop coil (d = 11 cm) was used for signal detection, fixed onto the head of the dog and to the table of the scanner. In each run, 139 volumes were acquired, with the first 5 of them being discarded before processing, resulting in a total functional scanning time of 367 seconds/run. A T1-weighted anatomical scan was carried out separately as part of another study on each awake dog for spatial registration on a 3 T Philips Ingenia scanner (Philips Medical Systems, Best, The Netherlands), using a 3D Turbo Field Echo (TFE) sequence, with TR = 9.85 ms, TE = 4.6 ms, and an isotropic resolution of 1 mm. image analysis. FMRI preprocessing included affine realignment (6 parameters, least square aproach) and reslicing of the images of the individual runs in SPM12 (http://www.fil.ion.ucl.ac.uk/spm/), followed by manual coregistration of the mean image to the individuals' own structural T1 image in Amira 6.0 (Thermo Fisher Scientific). The individual structural images were normalized and transformed (linear, non-rigid transformation) to a stereotaxic breed-averaged, T2 weighted template brain 15 with Amira. The resliced images were then coregistered and normalized to this transformed mean functional image via SPM's standard nonlinear warping function with 16 iterations and smoothed with an FWHM of 4 mm.
We applied band-pass filtering with a 0.01 and 0.1 Hz cutoff and linear detrending in CONN 16 . A single run had been censored due to exceeding motion threshold, after 127 scans (out of 134). As mean (scan-to-scan) motion was 0.035 ± 0.016 mm (mean ± SD), censoring of fMRI volumes during runs due to excessive motion, as applied by e.g. 17 was deemed inappropriate, because only 3 volumes (each in separate runs) out of the 5975 collected images during the study (0.05%) would have been affected by such a treshold. We decided against censoring via removing volumes midrun from our functional datasets, because censoring in combination with frequency filtering can introduce additional artefacts 18 and excessive motion was very infrequent (scan-to-scan displacement larger than 0.2 mm occurred in only 5% of the scans). We applied published, averaged, segmented white matter (WM) and cerebrospinal fluid (CSF) dog MRI brain masks 15 in our analysis as nuisance regressors to filter out non-neural signal fluctuations. These standard masks were applied to our normalized images. Figure 2 shows a sample of the coregistration of the grey matter mask to a functional image. www.nature.com/scientificreports www.nature.com/scientificreports/ independent component analysis. Independent Component Analysis (ICA) is a model-free, data driven method. Spatial ICA seeks components that are systematically non-overlapping, temporally coherent and maximally independent in space, without constraining the shape of the temporal response. We used the CONN software 16 to carry out group level spatial ICA. CONN is a Matlab-based cross-platform software for the computation, display, and analysis of functional connectivity in fMRI. ICA identifies a number of networks of highly functionally-connected areas. CONN's implementation uses Calhoun's group-level ICA approach 19 , with variance normalization pre-condititioning, subject concatenation of BOLD signal data along temporal dimension, group-level dimensionality reduction (to the target number of dimensions/components), fastICA for estimation of independent spatial components, and GICA1 backprojection for individual subject-level spatial map estimation. We included the WM mask, CSF mask and the realignment parameters to estimate physiologic noise (e.g. cardiac and respiratory cycles) and included them as nuisance regressors to filter out non-neural signal fluctuations. We applied a whole-brain mask to restrict the analysis space. Before the analysis, we also inspected QA plots regarding registration of the normalised functional images and the outline of the grey matter ROI. Anatomical labelling was carried out based on relevant anatomical brain atlases 15,[20][21][22] . During evaluation of the components, we relied on the guidelines published in Griffanti et al. 23 . We inspected the components with multiple thresholds, with different planes, while looking at both positive and negative clusters. During evaluating the components, we took into account the location of the susceptibility artefact and took a conservative approach to avoid classifying components containing clusters predominantly from the affected area as neural signal. As currently there is little information regarding the location and extent of susceptibility artefact in case of different dog breeds' functional brain scans, we created a guideline after going through and summarizing our raw images, looking for the regions which were affected by distortion/signal loss in the individual runs. Figure 3 shows the guideline we utilized during our evaluation, to flag clusters which are likely to be the result of susceptibility. While increasing model order increases the functional neuroanatomical precision, it reduces the repeatability of the ICA decomposition 24 . As model orders of 10-20 were reported as most suitable to detect large functional network clusters 24 , we run gICA with model orders of 10, 15 and 20.
Assessing reproducibility of independent components. To evaluate the reproducibility of the resulting ICA components, we calculated the Dice similarity coefficient. The Dice similarity coefficient (DSC) is used to evaluate reliability or reproducibility of MRI volumes, providing a reproducibility validation metric via calculating a spatial overlap index. The value of a DSC ranges from 0, indicating no spatial overlap between two sets of binary segmentation results, to 1, indicating complete overlap 25 . In case of evaluating consistency of resting-state networks from ICA, a DSC value ≥0.3 is considered as good spatial overlap consistency in studies explicitly designed to investigate reproducibility 26 .
To get an estimate of spatial overlap consistency of our data, we split it into two parts (first and second run of the subjects), then we run two additional, separate subICAs and calculated the spatial overlap between these sets and the original ICA. First, to match the subICA components to the originals, we compared them to the original ICA components, and labelled them based on the followings: the ICA component having the highest DSC value for the original network component X was labelled X. If a component had the highest DSC value for multiple networks, the label with the highest DSC value was selected. If the highest DSC value for a network was below 0.3, it was labelled with a new letter.

Results
To evaluate the quality of the collected data regarding head motion, we calculated mean motion from the realigment parameters. In our sample mean motion was 0.035 mm ± 0.016 mm (mean ± SD, n = 22), which is comparable to the reported mean motion values in human samples 27 . This confirms the effectiveness of our training procedure and shows the feasibility of carrying out resting-state measurements with awake unrestrained dogs.   www.nature.com/scientificreports www.nature.com/scientificreports/ threshold of p-FDR corrected <0.001 and a cluster threshold of p-FDR corrected <0.005. Component A (Fig. 4) covered parts of the rostral composit gyrus, rostral regions of the cingulate gyrus and the straight gyrus, the subcallosal area and diagonal gyrus; the left premotor area; medial and bilateral caudal regions of the cingulate gyrus and splenial gyrus, bilateral regions of the hippocampus and parahippocampal gyrus and the caudal composit gyrus. Component B (Fig. 5) included anterior brain regions in the prefrontal area, namely the genual gyrus, prorean gyrus, straight gyrus and dorsal anterior cingulate cortex. This component was located dorsally www.nature.com/scientificreports www.nature.com/scientificreports/ from the frontal hubs of Component A. Component C (Fig. 6) was a right lateralized frontal network, including secondary somatosensory cortices, such as the right rostral composit gyrus, the right caudate nucleus and the right rostral suprasylvian gyrus. While this network was less extensive on the left side, it showed clear indications of bilaterality, covering parts of the frontal lobe, namely the left rostral composit gyrus and the prorean gyrus. We found two components which were in part symmetrical, one of them left while the other right lateralized. Component D (Fig. 7) was left lateralized, including the left amygdala, left caudate nucleus, striatum, cerebellum, bilateral insular cortex, thalamus, fronto-parietal regions, the medial prefrontal cortex, and the mid cingulate gyrus. Component E (Fig. 8) covered the right amygdala, the striatum, the cerebellum, the right insular cortex, parts of the left visual corex, the septal nuclei, the left sylvian and ectosylvian gyrus. Component F (Fig. 9) covered  www.nature.com/scientificreports www.nature.com/scientificreports/ regions at the junction of the frontal, temporal and parietal lobes, including frontal regions of the sensorimotoric cortex, the bilateral rostral ectosylvian gyrus and rostral suprasylvian gyrus. Component G (Fig. 10) covered the cerebellum and the mesencephalon (corpora quadrigemina). This component is analogue to cerebellar networks previously reported in human 28 and non-human animal studies 6,8 . Component H (Fig. 11) consisted of the  www.nature.com/scientificreports www.nature.com/scientificreports/ mid cingulate cortex, an integrative part of the external limbic circle, related to affective processes and memory. Component I (Fig. 12) covered the bilateral auditory cortices, namely the rostral and caudal part of the sylvian gyrus; middle, rostral and caudal parts of the ectosylvian gyrus; and the middle, rostral and caudal regions of the suprasylvian gyrus. Component J (Fig. 13) included the primer and associative sensory cortical areas, namely the bilateral marginal gyrus and the ectomarginal gyrus. Component K (Fig. 14) included the mid cingulate gyrus, frontal gyrus, genual gyrus, the pre-and postcruciate gyri, the splenial gyrus, the posterior cingulate gyrus and  www.nature.com/scientificreports www.nature.com/scientificreports/ vessels. It covered the sinus cavernosus and the Willis' circle, large veins and arteries located at the ventral part of the brain. Component O (Fig. 18) showed characteristics consistent with a motion artefact (a ring around the edge of the brain). We classified component O as motion artefact because it contained a ring around the edge of the brain (which was more visible when both the positive and negative clusters are displayed as they are complementary) and the clusters did not follow known anatomical boundaries. In contrast, component M did not show these characteristics, it also contained bilateral frontal clusters which were not ring-like (slices 4-8) and www.nature.com/scientificreports www.nature.com/scientificreports/ component M showed a pattern which was overlapping with brain regions known to process stimuli from the visual modality [29][30][31] .
Compared to the 15 component gICA, the 10 item gICA contained only one miscancellous/noise component, while multiple components showed signs of mixed signal sources merged together, rendering this modell less  Assessing reproducibility of independent components. Labelling subset components. In subset 1 (Table 2), no component was labelled as O, and in subset 2 (Table 3)  Reproducibility across runs. After this, we calculated the DSC value between the two, labelled subICA components ( Table 4)  www.nature.com/scientificreports www.nature.com/scientificreports/ except B (0.15) and F (0.16) which were also overlapping with some nonmatching components. The low DSC value of B is most likely a result of that in subset 2, two separate components showed a DSC value higher than 0.3. As our study was not designed to carry out this analysis (hence the low number of datapoints), this high consistency in spatial overlap between separate ICAs shows the robustness and reproducibility of the reported components. www.nature.com/scientificreports www.nature.com/scientificreports/

Discussion
In this exploratory study, our main aim was to test whether applying the currently available methods, which had been successfully used to explore resting-state networks in other species, yield interpretable results with our setup (proof-of-concept) and describe the spatial characteristics of these networks. To achieve this goal, we decided to use a data-driven method, which is not relying on a priori hypothesis regarding the supposed function of certain brain regions. We found multiple, spatially distributed RSNs in dogs, the evolutionarily most distant taxa from humans scanned without anesthesia so far. To evaluate spatial overlap consistency, we calculated the Dice coefficients of the components from separate ICAs, which corroborated the robustness of the results. The localisation of these network correspond to the gross functional anatomical regions of the dog brain as described in the main neuroanatomical textbooks 32-35 (e.g. primary visual field, auditory cortex, limbic circuit, sensorimotor region) and www.nature.com/scientificreports www.nature.com/scientificreports/ to the resting-state networks previously described in other species [7][8][9][10]12 , suggesting that our setup and analysis pipeline is suitable to detect resting-state networks in awake dogs.
Motion is still present in the data after standard pre-processing. The ICA method is able to deal with these structured noise effects via separating these in form of an additional (noise) ICA component 36 . The presence of such a component signals that the algorithm was successful at detecting and grouping the motion related fluctuations together and not a sign of problems with data quality. ICA is often used in human studies to remove the effects of residual motion as part of automated denoising pipelines 37 . www.nature.com/scientificreports www.nature.com/scientificreports/ We described two visual network components (L & M), a finding similar to that of macaques 9 and humans 38 , with one network corresponding to primary visual areas (medial visual cortical areas) and another encompassing visual association areas. Interestingly, Component M also shows a similar location to the pDMN component reported in the previous dog study 11 . In our study, we found two lateralized networks (Component D & E) including the striatum, amygdala, nucleus caudatus, cerebellum, insular cortex and thalamus, and a third, separate component covering the mid-cingulate gyrus (Component H). These regions are all nodes within the human saliency network 39 , and they may also play a similar role in dogs, but to determine the stability, repeatability and functional characteristics of this component(s), further studies are needed. Based on our findings, a necessary next step is to investigate the functional connectivity specifically between these putative saliency nodes on another dataset. We also found a bilateral frontal component, a finding similar to the results from praire voles 10 , marmosets 8 and macaques 9 , but not reported outside of the superorder Euarchontoglires (primates and rodents) taxon so far.
Our results indicate that awake, unrestrained dogs' possess a network (Component A) showing antero-posterior connectedness, containing areas from both the prefrontal cortex and the anterior cingulate cortex, with additional involvement of hippocampal regions. These regions correspond to regions indicated in the default mode network (DMN) described previously in humans 40 and in animals 8,10 . However, the network we report here contains more regions than the traditional human DMN. For example, the composite gyrus, the hippocampi, splenial gyrus, and premotor area(although see 40 on recent reports on the potential involvement of some of these additional regions in the human DMN). Therefore, it is problematic to label this component as DMN with only resting state ICA. In humans, specific regions of the DMN (MPFC, PCC and IPL) deactivate in response to a task. In order to definitively label the network we report here as the DMN, more research will need to show that analogous regions in the dog brain deactivate during a task. This dog network (Component A) does not include the parietal cortex. In humans, these parietal regions (angular gyrus, temporoparietal junction) are thought to be involved in complex cognitive processes such as accessing conceptual representations about events or items and theory of mind 40 , cognitive functions whose extent and level in dogs are debated 41 . Similarly to our findings, a putative DMN containing both the posterior cingulate cortex and frontal cortical areas have been desribed in rats 7 , voles 10 , ferrets 12 , marmosets 8 and macaques 9 so far. Our results do not support previous results in dogs 11 , which suggested dissociation of anterior and posterior regions of a DMN-like network, however our study diverged in several aspects which could account for the different findings. We utilized a larger sample size (4 vs. 22), longer data acquistion runs, different temporal filtering, lower dimensionality in the ICA. It is important to note that seed-based analysis (a model-driven approach) requires strong a priori hypothesis, a slight difference in the location of the spatial seed can have a significant impact on the spatial characteristics of the resulting RSNs 36 . Currently, we lack extensive fMRI studies regarding the functional properties of different dog brain regions, high resolution anatomical/cytological maps or molecular evidence to say which brain regions could be considered truly homologous e.g. how should be the dog cingulate gyrus divided into anterior, mid and posterior regions.  Table 1. Spatial topography of ICA models at model orders of 15 and 10, listing the most characteristic brain regions. Capital letters refer to components from gICA 15 , while the numbers refer to components from the gICA 10 . Similarity is based on the involved brain regions and visual characteristics. www.nature.com/scientificreports www.nature.com/scientificreports/ Dogs have some advantage regarding the scanning setup when compared to other model animals in fMRI studies. After special training, dogs are suitable to be scanned without sedation, which is beneficial as anaesthesia has a large impact on resting state functional connectivity 42 , and unlike other non-human animal species scanned awake, there is no need of mechanical restraints either, which also have the risk of influencing rs-network connectivity 43 .
In case of the relatively long TR which we used in our study, physiological artifacts can alias into the band of interest. For human adults, a normal resting heart rate is between 60 and 100 beats per minute (bpm). Reports of resting heart rate in dogs seem to show a large variation depending on the context in which they were measured. Veterinary visits or research procedures in case of laboratory dogs are stressful situations to dogs, and heart rates  www.nature.com/scientificreports www.nature.com/scientificreports/ measured under these conditions reflects this. While the grand average of HR measured at veterinary visits 44 and under laboratory conditions 45 was around 124 beats/min, in another study, where dogs participated in a behavioural experiment with minimal physical activity, the grand average of HR was only around 80 beats/min 46 and in a polysomnography study, where dogs were resting with their owners (even less physical activity), the grand average HR was only around 70 beats/min (Bálint et al. 47 ). Respiration rate is indeed higher than in adults, but it is similar to the respiration rate of children (approx. 30/min) 48 . ICA is suited to deal with such physiological noise (e.g. respiration, pulsation), resulting in components with clusters mainly located in the white matter, cerebrospinal fluid and blood vessels (particularly arteries) 23 . Hence, we believe that regressing out signals from these regions via masking takes this into account, as these fluctuations related to physiological noise are also present in the white matter and cerebrospinal fluid space.
Our study is the first dog fMRI study utilizing atlas-based brain segmentation, mainly due to a lack of published brain masks delineating the limits of grey and white matter regions (while there is a binary atlas available, currently no probabilistic atlas exists). The creation, comparison and validation of such masks through different segmentation algorithms for dog brains is outside the scope of our current resting-state functional MRI study. Although we collected T1 weighted structural images from our subjects for the purpose of coregistration, the contrast of these images was not suitable to carry out reliable, individual segmentation based on intensity.   www.nature.com/scientificreports www.nature.com/scientificreports/ While our subjects did not look at a fixation cross during scanning, a study with adults, where the authors applied seed-based analysis showed no significant difference between eyes open and a fixation cross condition regarding intranetwork connectivity strengths 16 and in another study on human children, where ICA results of a movie watching condition and a no stimuli resting condition were compared, revealed no significant differences in within-network connectivity between rest and movie watching 49 . In addition, resting state networks have been also successfully extracted from task-based co-activation patterns 50 . Based on these studies, it is unlikely that the presence of a static human during scanning would have significantly altered the spatial extent of the described resting-state networks.
The dog subjects are willingly motionless during scanning for an extended period of time (as performing a conscious action inhibiting their movements), a situation more closely resembling the conditions of human rs-fMRI measurements. While one may argue that the dogs are executing a rewarded "task" and not really resting, it is important to keep in mind that humans also consciously perform the same "hold still" task during fMRI measurments and most often do so in exchange to some predetermined reward (in form of financial compensation). Additionaly, while resting state networks are usually analysed in data collected during rest, they are also present when performing cognitive tasks 51 .
Based on the current study, we cannot provide information whether the described networks should be considered task positive or task negative (a dichotomy that has been recently challenged, see e.g. 52 ) or regarding their functional characteristics (e.g. how is dogs' visual network operating compared to visual networks of other species). The identification of the networks, like in other animal studies was based on the anatomical properties of the network, without investigating whether they show deactivation during the performance of specific tasks (as in anesthetized prairie voles 10 , rats 7 and awake marmosets 8 ). The description of these dog resting-state networks provide information for future studies to set up more specific a priori hypotheses to test the functional characteristics and interactions of the networks e.g. via also combining it with behavioural data.