Magnetic resonance techniques such as dMRI and resting state FC were first introduced in the late 1990s to investigate the structural and functional architecture of the central nervous system in human subjects (see Pierpaoli et al.
1996; Biswal
2012). By providing information that can be collected non-invasively in a digital format, these approaches have great utility for clinical and research purposes. Using these data to construct maps of connections, whether as tractography or functional connectivity, provides considerably opportunities for studying brain architecture and its functional role. In the animal model, especially in the macaque, detailed knowledge of cortical connectivity and functional organization of motor control has been collected based on neural tracer and electrophysiological experiments, which enables some validation to be made of these MR-based approaches and, thus, assessment of how accurate they may be in representing genuine anatomical and functional organisation. A growing number of studies have compared data on the cortical connectivity of the macaque brain with dMRI data (e.g., Markov et al.
2014; Thomas et al.
2014; van den Heuvel et al.
2015; Azadbakht et al.
2015; Donahue et al.
2016), arriving at different conclusions on the accuracy of dMRI in tracing neural connections. Other studies have used rs-FC to highlight functional correlations between cortical areas of the macaque brain which may indicate some similarity in the functional architecture of the macaque and the human brain (Hutchison et al.
2011,
2015). In the present study, we test the efficacy and accuracy of different dMRI and rs-FC techniques in identifying two large-scale cortical motor control networks (i.e., the lateral grasping network and the exploratory oculomotor network) in the living macaque brain, for which there is detailed, solid knowledge in terms of involved areas and interconnections based on neural tracing and electrophysiological studies. In contrast with other studies, the cortical sectors used as seeds have been defined in every monkey in order to include the core of each cortical area of the networks under study.
This investigation showed that dMRI and rs-FC produce different results in terms of estimating connections. Tractography was able to detect connections with higher specificity than resting state techniques, but less sensitivity. Importantly, both deterministic and probabilistic methods provided false positives and negatives and were not always in accordance as to the connections identified. Tractography has limited utility for studying short-range connections within directly adjacent cortex, although it can be used to show U-shaped fibres between adjacent gyri (Catani et al.
2012,
2017; Guevara et al.
2017). Long range fibre bundles are more reliably reproduced with tractography (e.g., Rojkova et al.
2016; Warrington et al.
2020). As such, it is noteworthy that neural tracing studies show that connections of a given cortical area typically involve, qualitatively and quantitatively, mostly adjacent cortex and neighbouring areas and that long-distance connections, although important from a functional point of view, generally represent a minor component of the total labelled cells. For example, after neural tracer injections in F5p, about 60–70% of the labelled neurons are located within the primary motor/premotor cortex, whereas only about 4–5% of the labelled cells are located in IPL areas AIP and PFG, which are sources of visual information crucial for selecting and controlling object-oriented hand actions (Gerbella et al.
2011). Thus, comparing long-distance, point-to-point cortical connectivity is somewhat challenging when comparing the accuracy of dMRI in identifying connections between different sectors. For this reason, we instead compared the presence of connections between techniques, however, this may be an area for future study. On the other hand, resting state techniques were better able to identify local over long-range functional connectivity, but their interpretation is less clear. As discussed below, the present data, in agreement with some previous studies (Thomas et al.
2014; Reveley et al.
2015), indicate some expedients that may be helpful in improving the quality of these neuroimaging techniques in representing the underlying anatomy.
Comparing dMRI with neural tracing
In the present study, different sectors of the LGNet and EONet were defined on the cortical mantle, and for tractography analysis were slightly extended into the immediately contiguous white matter. Tractography was primarily effective in identifying fibres connecting parietal and frontal cortex, in particular those projecting from AIP to SII, F5a and F5c, although projections to prefrontal sectors could not be identified. Connections between the posterior bank of the arcuate sulcus, ventral premotor regions (F5) and rostral inferior parietal regions (PFG, PF, AIP) likely constitute the macaque homologue of the ventral branch of the superior longitudinal fasciculus (SLF III), making up the core component of the LGNet (Croxson
2005; Schmahmann et al.
2007; Mars et al.
2011; Warrington et al.
2020). Despite marked differences in gyral and areal organization, similarities in the organization of parieto-frontal connections in the macaque and the human brain have been reported (e.g., Thiebaut de Schotten et al.
2012). In the human brain, the SLF III connects the rostral IPL (supramarginal gyrus; BA40) with the ventral portion of the precentral gyrus (BA6) and the inferior frontal gyrus (BA44, 45) (Thiebaut de Schotten et al.
2011b).
On the other hand, both structural and functional connectivity techniques could not identify connections running directly between the anterior bank of the arcuate sulcus (FEF, 45B) and caudal inferior parietal regions (LIP). These fibres constitute part of the middle branch of the superior longitudinal fasciculus (SLF II), the core parieto-frontal connection of the exploratory oculomotor network (Sani et al.
2019). These fibres run in parallel but more medial to the SLF III in an anterior–posterior direction, with a clear distinction between each tract bundle. We speculate that tractography algorithms may be unable to bend and track dorsally to reach the prearcuate cortex, where the oculomotor areas FEF, 45B, and caudal 46v are located. Hence, to reproduce fibres of this tract in the macaque, regions-of-interest within deeper white matter beneath the arcuate sulcus may be required.
We compared representative trajectories of labelled axons extending from AIP with spherical deconvolution modelling used to generate tractography, to assess whether the fODF within a single voxel was reflective of likely axon trajectories. We first showed AIP-frontal projections running dorsally above SII, and tractography was able to reliably reproduce the projections that also extended into SII. We also showed that AIP-F5 connections shown with tracers could also be reproduced using tractography. However, further along this tract, within the frontal lobe, streamlines projecting through and past F5 (mainly F5a) were not connected with prefrontal sectors (12r and 46v). A previous study comparing postmortem dMRI tractography with histological analysis showed that the presence of uniform and dense sheets of fibres running below and parallel to layer VI poses challenges for diffusion tracking into sulcal regions, as well as into gyral crowns (Reveley et al.
2015). These dense U-shaped connections surround the arcuate sulcus, as well as within the parietal lobe and may hence influence the efficacy of tractography in projecting through this area (Schmahmann and Pandya
2006; Catani et al.
2012,
2017).
Previous studies, which have compared dMRI with neural tracer data in the macaque brain (Markov et al.
2014; Thomas et al.
2014; van den Heuvel et al.
2015; Azadbakht et al.
2015; Donahue et al.
2016) also show that confining regions of interest (ROIs) to gray matter strongly reduces the sensitivity of dMRI. For this reason, it is necessary to extend ROIs into the underlying white matter, as we did, however this may reduce the specificity of results. Comparisons of different dMRI analyses show that by changing parameters it is possible to alter sensitivity (measured as true positives) and specificity (measured as false positives) of results. We compared different tractography approaches, but in fact showed that there was fairly good correspondence between techniques, with only a slightly increased risk of increasing false positives when using probabilistic approaches. It has been hypothesized (van den Heuvel et al.
2015) that false positive results obtained with dMRI could also be explained by the lack of efficacy of neural tracers in identifying some connections (false negatives). We were unable to evaluate this in the present study, as we focused on identifying well-documented connections from neural tracer studies, although it may hence be important to also compare tractography techniques with other invasive approaches such as polarised light imaging (Axer et al.
2011).
Our results indicate that comparing genuine fibre trajectories, identified with neural tracing, with non-invasive approaches may help to highlight areas for which tractography lack accuracy (for example projections from AIP to prefrontal sectors). This information is relevant in improving the accuracy of tractography output (Smith et al.
2012; Jbabdi et al.
2015). For example, in human studies, this approach has improved tracking of Meyer’s loop of the optic radiation as well as the acoustic radiation, both of which are challenging to reproduce with most tractography algorithms (Chamberland et al.
2017; Maffei et al.
2018). There have been a number of recent studies showing inherent biases in tractography output, which vary depending on the acquisition, preprocessing and dissection approach used (Dyrby et al.
2011; Maier-Hein et al.
2017; Jeurissen et al.
2019). While simulated phantoms are commonly used to appraise the reliability of models and algorithms in human studies (e.g., Poupon et al.
2008; Neher et al.
2014), the availability of tracing data and the growing field of comparative MRI may provide a meaningful opportunity to identify anatomical and orientational priors to improve tractography (Rheault et al.
2019).
Comparing resting state MRI with neural tracing
We showed that rs-FC was partially effective in identifying the various different nodes of both the LGNet and the EONet. Indeed, the rs-FC matrix in Fig.
5 shows several examples of coherent fluctuations of BOLD signals of ROIs, which are anatomically connected and are involved in the LGNet or the EONet. For example, various F5 subdivisions appeared to be functionally correlated with rostral IPL areas, area SII and the insula, and area 45B appeared to be functionally correlated with the temporal cortex.
However, the connectional matrix also shows that all the various ROIs tend to be strongly functionally correlated with adjacent cortical areas of the same or the other network. Furthermore, the maps in Fig.
5 show that the various ROIs were typically at the centre of a relatively large, fairly homogeneously extending, functionally correlated region. Thus, the FEF, for example, was functionally correlated with adjacent oculomotor prefrontal areas, but also with ventral premotor cortex and even with the primary motor area F1. Accordingly, our data provide evidence for a clear tendency of rs-FC to show false positive, short distance “connections”.
The matrix also showed relatively poor long-distance rs-FC for many ROIs. For example, area LIP did not correlate with either of the prefrontal oculomotor areas, or the temporal cortex. Accordingly, long-distance rs-FC may be affected by false negatives. This could partially be attributed to inter-individual differences in areal localization that can affect the identification, at the group level, of subtle differences in rs-FC. This was also confirmed by the results of the UC Davis individual level rs-FC analysis, which showed long-distance rs-FC that was absent at the group level (e.g., LIP-FEF and AIP-F5c). These connections may have been better identified through the use of individual definition of the intrasulcal areas.
The previously described limitations appear to be a common problem of rs-FC in macaques, even when different approaches are used (Mars et al.
2011; Neubert et al.
2014; Hutchison et al.
2015; Sharma et al.
2019). Using a different rs-FC analysis, not requiring a priori seed definition, Hutchinson and colleagues (
2011) defined eleven different networks across the entire macaque cortex. In most cases, these networks included mostly neighbouring regions, rather than anatomically connected zones, as well as very few distant regions, except for homologous contralateral areas. Sharma and colleagues (
2019) describe different patterns of rs-FC of F5 subdivisions using a contrast agent for enhancing signals in awake macaque monkeys. Their results appear to be very similar to those we observed. Indeed, although the observed patterns appear to vary according to the location of the seeds, all tended to involve a large cortical region around the seed, including the FEF and neighbouring prefrontal oculomotor areas.
Some similarities with our results were also observed in other studies (Mars et al.
2011; Neubert et al.
2014) in which, for example, prearcuate oculomotor areas showed rs-FC with F5 and F1 (false positive) and did not show rs-FC connectivity with inferotemporal areas (false negative). It therefore seems from the present and other studies that rs-FC shows relatively similar and reproducible patterns even when different approaches are used. Furthermore, most studies (e.g., Babapoor-Farrokhran et al.
2013; Neubert et al.
2014; Hutchison et al.
2015; Sharma et al.
2019) show specificity of these patterns, even when adjacent zones are compared. Abrupt variations in rs-FC patterns among adjacent zones have been used to define areal parcellation in the human brain (Cohen et al.
2008; Xu et al.
2019). However, this FC-based parcellation approach should be adopted with caution. Indeed, as already pointed out in several studies and in line with our data, coherent fluctuations of BOLD signal do not necessarily reflect direct cortical connectivity or common functional properties. Furthermore, the possible contribution of indirect, polysynaptic connectivity, or of common subcortical input may be variable across cases and cannot be assessed easily with this method. It is also challenging to interpret how there can be a lack of functional connectivity between areas that are relatively strongly anatomically connected. Finally, another aspect that still remains to be clarified is why rs-FC and tracer patterns appear to be in better concordance for somatomotor areas than for other regions such as, e.g., prefrontal cortex (Van Essen et al.
2019). One possible explanation could be that variability across monkeys in functional connectivity appears to be lower for the primary sensory and motor areas than the high-order association regions which make up the majority of long-distance connections (Xu et al.
2019).
One paradigmatic example, based on our data, of the possible difficulties in explaining rs-FC data in light of current interpretations is represented by the correlation between the FEF and F1, observed here and in other studies discussed above. It is well established that these two areas lack direct anatomical connections and do not appear to share common connections with other cortical areas, but also have markedly different input from the thalamus.
Limitations
Non-human primate neuroimaging is a growing field, although its quality is not yet at the stage of human neuroimaging, as there are unique challenges to be faced when acquiring this data. High field strengths are commonly used, much higher than those regularly used for human studies (around 3T), with custom built surface coils which can result in B1 homogeneity and varying coil coverage which can cause alterations in image intensity. This also leads to distortions and dephasing due to susceptibility. As such, pipelines to process these data have to be carefully optimised, and until the release of the recent PRIME-DE resource (Milham et al.
2018,
2020), there has been no established benchmark from which to establish data quality. There were some discrepancies between the connections identified here and those described in previous studies (Warrington et al.
2020; Schmahmann et al.
2007; Sani et al.
2019), which may indicate that data quality or data processing may not have been optimised to visualise these connections. Future studies may use the sectors provided here to study different datasets, such as those acquired postmortem, to evaluate whether the connections can be identified. In terms of rs-FC analysis, it is important to note that it can be affected by state, and both cohorts of monkeys were anaesthetised (Xu et al.
2019). It also remains to establish whether rs-FC is more similar to anatomical tracing and tractography results within the awake state.