Structural connectivity-based segmentation of the human entorhinal cortex

The medial (MEC) and lateral entorhinal cortex (LEC), widely studied in rodents, are well defined and characterized. In humans, however, the exact locations of their homologues remain uncertain. Previous functional magnetic resonance imaging (fMRI) studies have subdivided the human EC into posterior-medial (pmEC) and anterior-lateral (alEC) parts, but uncertainty remains about the choice of imaging modality and seed regions, in particular in light of a substantial revision of the classical model of EC connectivity based on novel insights from rodent anatomy. Here, we used structural, not functional imaging, namely diffusion tensor imaging (DTI) and probabilistic tractography to segment the human EC based on differential connectivity to other brain regions known to project selectively to MEC or LEC. We defined MEC as more strongly connected to presubiculum and retrosplenial cortex (RSC), and LEC as more strongly connected to distal CA1 and proximal subiculum (dCA1pSub) and orbitofrontal cortex (OFC). Although our DTI segmentation had a larger medial-lateral component than in previous fMRI studies, our results show that the human MEC and LEC homologues have a border oriented both towards the posterior-anterior and medial-lateral axes, supporting the differentiation between pmEC and alEC.

In anatomical and functional studies of the human brain, magnetic resonance imaging (MRI) has become an invaluable tool. Functional MRI (fMRI) studies have shown that the properties of the rodent and non-human primate EC also apply to the human EC (Doeller, Barry, & Burgess, 2010;Reagh & Yassa, 2014;Schultz et al., 2012). Based on the subdivision of the rodent EC into MEC and LEC, studies have tried to localize their respective homologue regions in humans. Previous fMRI studies tested connectivity 'fingerprints' of EC subregions to other parts of the brain. Studies in rodents and non-human primates have demonstrated a largely similar organization of EC connectivity across species (Canto et al., 2008), thus predicting distinct fMRI connectivity fingerprints for the two subregions in humans as well. The resulting delineations of putative human homologue regions of the rodent MEC and LEC were labeled posteromedial EC (pmEC) and anterolateral EC (alEC), based on the outcome of two independent fMRI studies that tested local and global connectivity, respectively (Maass et al., 2015;Navarro Schröder et al., 2015). However, it remains unclear whether the results were affected by the nature of the imaging modality (fMRI) or the choice of seed brain regions used to identify the subregions.
In addition to the neuroimaging modality, the second reason for a re-evaluation has gained additional importance since the assumption about EC connectivity on which parts of the previous fMRI studies (Maass et al., 2015;Navarro Schröder et al., 2015) were based has been recently revised. For years, the existence of two parallel cortical connectivity streams through the EC has been the accepted model (Nilssen, Doan, Nigro, Ohara, & Witter, 2019;Ranganath & Ritchey, 2012;Witter, Doan, Jacobsen, Nilssen, & Ohara, 2017). This comprises one pathway into the hippocampus via the parahippocampal/postrhinal cortex (PHC/POR) and MEC (the "where" pathway), and a parallel pathway via the perirhinal cortex (PRC) and LEC (the "what" pathway). However, recent evidence substantially challenged this view. Doan and colleagues found that POR in rats, which corresponds to the PHC in humans, does also project to LEC. These authors further argue that existing data in monkeys substantiate this notion (Doan, Lagartos-Donate, Nilssen, Ohara, & Witter, 2019). This is in line with new findings in humans indicating that the hippocampal-entorhinalneocortical connections are far more complex than a pure segregation into "where" and "what" pathways (C.-C. Huang, Rolls, Hsu, Feng, & Lin, 2021).
The objective of this study is therefore to identify the human homologues of the rodent MEC and LEC using DTI, incorporating the novel insights from rodent anatomy. To achieve this, we performed probabilistic tractography on high-quality DTI data acquired by the Human Connectome Project (Fan et al., 2016). We identify the EC subregions by analyzing the connectivity from regions of interest (ROIs) that project selectively to either of them and compare these to the results from previous fMRI studies.

Results
To visualize the connectivity paths between the EC and the regions hypothesized to be connected with its subregions, we ran probabilistic tractography between the regions. By seeding paths from all voxels in the EC, presubiculum, dCA1pSub, RSC and OFC ROIs, maps of the connectivity paths between the EC and the other ROIs were created. The resulting group averaged paths are shown in Figure 1. In all figures, blue color schemes are used for MEC-related regions, i.e. presubiculum and RSC, while red color schemes are used for LEC-related regions, i.e. dCA1pSub and OFC. The maps show that all the regions exhibit clear connectivity with the EC. Connections with dCA1pSub extend further anteriorly in the EC than the connections with the presubiculum, and the connections with presubiculum and RSC seem to take a similar route to the EC. The paths between OFC and EC, however, stand out from the others as they take a more lateral route, but the inferior part seems to pass close to dCA1pSub. Note that the colormap intensity in these maps does not represent the actual number of white matter tracts, but instead scales with the probability that the true path between the ROIs lies in that point. Corresponding connectivity paths for one example participant are shown in Figure 1-figure supplement 1.
Because we wanted to segment the EC into the MEC and LEC homologues based on the connectivity with other regions, a voxel-by-voxel measure of connectivity probability was needed. We therefore also ran the tractography only seeding from the EC ROIs. Then, for each voxel in the ROI, we counted how many of the seeded paths reached the other ROIs.
These connectivity counts were normalized to a probability, providing connectivity maps for the EC with the other four ROIs. The resulting smoothed and thresholded group averaged connectivity maps are shown in Figure 2. The sagittal slices show that the connectivity with presubiculum and RSC appears to be strongest in the posterior part of the EC, whereas the connectivity with dCA1pSub and OFC is strongest anteriorly in the EC. Further, the presubiculum connectivity does not show a clear medial-lateral gradient, but the connections with dCA1pSub, RSC and OFC are stronger laterally in the EC in the selected coronal slices.
Corresponding connectivity maps for one example participant are shown in Figure    For segmentation into the MEC and LEC homologues, the main hypothesis was that these regions could be identified based on connectivity with presubiculum vs. dCA1pSub, respectively. The actual segmentation was performed on a voxel-by-voxel level in the EC determining with which of the other two regions the connection probability was highest, using the connectivity maps described in the previous paragraph. For comparison, the MEC-LEC segmentation was also performed based on connectivity with RSC vs. OFC, respectively. This was first performed individually for all participants, and inter-participant segmentation variability maps for the presubiculum vs. dCA1pSub and RSC vs. OFC segmentation approaches are shown in Figure 3. For most participants, MEC is clearly located more posteriorly and LEC is located more anteriorly for both segmentation approaches, and in addition they are located more medially and laterally with respect to each other for the presubiculum vs. dCA1pSub approach. The RSC vs. OFC approach also shows this mediallateral trend of MEC and LEC across participants, although not as clear as for presubiculum vs. dCA1pSub. Corresponding MEC and LEC segmentations for one example participant are shown in Figure 3-figure supplement 1.
The same connectivity-based MEC-LEC segmentation was performed on a group level using the group averaged connectivity maps from Figure 2. As described above, the group segmentation was also performed using two different approachespresubiculum vs.
dCA1pSub, and RSC vs. OFCand the resulting segmentations are shown in Figure 4. We see that for the MEC and LEC predictions from presubiculum vs. dCA1pSub, there is a clear medial-lateral (ML) and posterior-anterior (PA)-oriented border between the subregions. For RSC vs. OFC, however, the PA-oriented border is most prominent, but it is also slightly MLoriented, most visible in the left EC. Because the results from the two approaches were slightly different, we also tried to interchange the order of the ROIs, and MEC and LEC segmentations from using presubiculum vs. OFC and RSC vs. dCA1pSub can be seen in where we averaged the connectivity maps for presubiculum and RSC, and the maps for dCA1pSub and OFC ( Figure 5A and B). Figure 5C shows the resulting MEC and LEC homologues from this combined segmentation approach. With this approach, as with separate combinations of seed regions, we find both a PA-and ML-oriented (although most visible in the left hemisphere) border between MEC and LEC. These final MEC and LEC masks are also available in the Supplementary files.
In a next step, since the borders of the segmentations from different approaches showed slightly different orientations along the posterior-anterior (PA) and medial-lateral (ML) axes, we wanted to quantify this directional difference by calculating the "degree" of PAand MLorientation of the borders. This was defined as a percentage from 0 to 100%, dependent on the angle between the MEC-LEC center of gravity vector and a pure PA or ML vector. Table   1 shows the resulting degrees of PA-vs. ML-oriented borders for the different segmentation 8 approaches including the fMRI segmentations from previous studies (Maass et al., 2015;Navarro Schröder et al., 2015). The center of gravity vectors are also plotted in a common   reference frame in Table 1-figure supplement 1. All DTI segmentation approaches have a border with a PA-orientation of around 50-60%, and a varying degree of ML-orientation from 6% for RSC vs. OFC up to 67% for presubiculum vs. dCA1pSub. The borders between the segmentations from fMRI have a high PA-orientation of around 92%, and a lower degree of ML-orientation than all of the DTI approaches. Interestingly, when comparing the different combinations of DTI approaches, using dCA1pSub as the defining region for LEC yields a higher degree of ML-orientation than using OFC. Similarly, using RSC as the defining region for MEC yields a slightly higher degree of PA-orientation of the border than using presubiculum, but this is less prominent. Finally, we wanted to compare the resulting sizes of the MEC and LEC homologues from all the different segmentation approaches, and these are shown in Table 2. For all DTI approaches, the MEC is larger than LEC, while fMRI on the other hand yields a larger LEC than MEC. The subregions are most equally sized when using the RSC vs. dCA1pSub approach.

Discussion
In this study, we used DTI and probabilistic tractography in 35 healthy adults to segment the human EC into homologues of what in other mammals have been functionally and cytoarchitectonically defined as MEC and LEC. We based the segmentation on EC connectivity with four brain regions known to selectively project to either of the EC subregions in multiple species. Different combinations of these four regions all showed both a posterior-anterior (PA) and a medial-lateral (ML)-oriented border between the human homologues of MEC and LEC. This orientation of the thus defined border is similar to that  (Maass et al., 2015;Navarro Schröder et al., 2015). Note however that our DTI results show a larger degree of ML-orientation, and a correspondingly lower degree of PAorientation of the border between the subregions compared to the previous fMRI results.
The results from our study substantiate the pmEC and alEC subdivision of the human EC suggested in previous fMRI studies (Maass et al., 2015;Navarro Schröder et al., 2015).
Although some earlier fMRI studies on mnemonic processing in the EC found a dissociation primarily along the medial-lateral axis (Reagh & Yassa, 2014;Schultz et al., 2012), it is important to realize that even the orientation of the cytoarchitectonically defined border between MEC and LEC in rodents does not align along a pure medial-to-lateral axis. Rather, the MEC in rodents is actually located in the posterior-medial EC, and the LEC is located in the anterior-lateral EC (van Strien et al., 2009). Also, in macaque monkeys, tracing studies show differential connectivity in caudal vs. rostral portions (Witter & Amaral, 2021). A pure medial-lateral subdivision of human EC is thus not to be expected. Nevertheless, the somewhat different orientations of the border between the human homologues of MEC vs.
LEC subdivisions found using DTI vs. fMRI studies raises the question of which of the two imaging modalities should be preferred to define the position and orientation of this border.
There are several possible explanations as to why our DTI study showed slightly different segmentation results than the fMRI studies. First, DTI and fMRI are two different imaging modalities with inherently different mechanisms of connectivity. While DTI exploits the diffusion of water molecules in order to trace the structural paths of connectivity between brain regions (Mori et al., 1999;Mori & Zhang, 2006;Powell et al., 2004;Zeineh et al., 2012), fMRI identifies functional connectivity by correlating blood-oxygen-level-dependent Interestingly, using different seed regions to identify MEC and LEC resulted in varying degrees of PA-and ML-orientation of the border between them. It is unclear whether this is inherently linked to the DTI method, or due to an actual connectivity difference between the regions. Using presubiculum and dCA1pSub as the seed regions, which are situated medially and laterally with respect to each other, respectively, resulted in a border with higher degree of ML-than PA-orientation. On the other hand, using RSC and OFC, which are situated posteriorly and anteriorly in the brain, respectively, resulted in a border with higher degree of PA-than ML-orientation. Although it is not unnatural to assume that the brain is organized such that connected regions are situated more closely to each other, this could also be an effect of using probabilistic tractography, where the apparent connectivity probability depends on e.g. the length of the path and the size of the ROIs (Behrens, Johansen-Berg, Jbabdi, Rushworth, & Woolrich, 2007). In other species, including rodents and monkeys, the presubiculum and RSC show inputs to the EC with a similar spatial distribution (Witter & Amaral, 2021), aligning with our maps of connectivity paths with these two seed regions. However, comparing the different MEC and LEC segmentations from the different seed region combinations shows that while interchanging presubiculum and RSC yields only slightly different orientation of the border along the PA and ML axes, the difference when interchanging dCA1pSub and OFC is more substantial. In other species, dCA1pSub are known to project to both rostral and dorsolateral parts of EC, whereas posterolateral OFC mainly projects dorsolaterally in the EC (Kondo & Witter, 2014;Saleem et al., 2008;Witter & Amaral, 1991;Witter & Amaral, 2021). Whether these regions in humans project to different parts of the homologue of LEC, or whether our results are affected by using DTI and probabilistic tractography, should be further investigated by also comparing EC functional connectivity to these areas using fMRI. Note also that the location of projections from dCA1pSub along the medial-lateral axis of the EC depends on where the seed is placed along the posterior-anterior axis of the dCA1pSub (Witter & Amaral, 2021), which emphasizes the importance of anatomically accurate seed ROIs.
In order to determine and compare the connectivities between EC and the other ROIs, we normalized the connectivity maps by dividing them by the maximum probability of each map.
This could introduce a bias in the results. By doing this, we intrinsically assume that the maximum connectivity strength to each of the other ROIs are equal, and the segmentation process does not take into account that the MEC connections might be stronger than the LEC connections, or vice versa. However, little is known about the strength of connectivities at this level of detail, particularly since it is not straightforward to examine or even define connectivity strength. Connectivity strength surely depends on axonal density, but other factors like synaptic density and efficacy are other important variables. Nevertheless, even if we were to know that some of the connections are stronger than the others, probabilistic tractography provides a relative instead of an absolute measure of connectivity and is also dependent on path lengths, ROI sizes and the number of possible path directions in a voxel.
Normalizing the connectivity maps based on different connectivity strengths would therefore be a highly complex task. Therefore, we did not impose any further assumptions about connectivity strengths in our analyses.
Our study has some limitations. To define our ROIs, we chose to use regions from automatic cortical segmentation protocols. This could have influenced the anatomical precision of our analysis. Manual segmentation would be labor-intensive and requires high skills in neuroanatomy, possibly limiting the number of participants that could be included in the study. However, we manually adjusted some of the automatically segmented ROIs, and also intersected the registered ROIs from MNI space with the participants' individual automatic segmentations in order to increase the anatomical accuracy. Another limitation is that there are inherent challenges to the EPI sequence used for diffusion imaging. This results in a generally low signal-to-noise ratio in the EC and the whole medial temporal lobe. In addition, these regions appear geometrically distorted in the EPI images, and although this has been corrected for, it is not possible to recover all of the lost signal. Imperfect correction can also affect the accuracy of the ROIs. Because of the probabilistic nature of the tractography technique it is unlikely that noise will introduce false connections, but it can leave some connections undetected. At last, a relatively low number of participants were included in our study, which might have influenced the statistical power of the results.
In conclusion, our DTI results support the definition of pmEC and alEC as the human homologues of MEC and LEC. Also inspired by novel insights coming from rodent anatomy, we present a segmentation based on a combination of differential presubiculum/RSC and dCA1pSub/OFC structural connectivity which indicates a border between the two subdivisions of EC with an orientation that is angled both towards the posterior-anterior axis, as well as to the medial-lateral axis. The fact that there are some differences in the orientation of the border based on DTI and fMRI data in addition to the seed regions used, indicates the need for investigation in a larger number of participants across both modalities.

Preprocessing
The MRI data were minimally preprocessed by the Human Connectome Project as described in (Fan et al., 2014). In brief, this preprocessing pipeline included gradient nonlinearity correction, motion correction, Eddy current correction and b-vector correction.

Registration
Both structural and diffusion images were brain extracted using the brain mask from running the FreeSurfer functions recon-all and dt-recon on the participant's structural and diffusion images, respectively (Fischl et al., 2002;Fischl et al., 2004), before refining the result using FSL's BET (http://fsl.fmrib.ox.ac.uk/fsl/) (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012;Smith, 2002). For the diffusion images, brain extraction and registration were performed on the participant's average b = 1000 image. The individual brain-extracted structural and diffusion images were registered to each other, as well as to the MNI152-09b standard brain template (Fonov, Evans, McKinstry, Almli, & Collins, 2009), using symmetric non-linear registration in the Advanced Neuroimaging Toolbox (ANTs) based on mutual information (Avants et al., 2011).

Regions of interest
Regions of interest (ROIs) including the EC, presubiculum, CA1 and subiculum were extracted from the automated cortical and subcortical parcellation obtained from running FreeSurfer's recon-all and segmentHA_T1 functions on the MNI152-09b template (Fischl et al., 2002;Fischl et al., 2004;Iglesias et al., 2015). The EC ROI was further refined by masking it by a probabilistic EC ROI, thresholded at 0.25 from the Jülich-Brain Cytoarchitectonic Atlas (Amunts, Mohlberg, Bludau, & Zilles, 2020). Since the resulting EC ROI extended too far posteriorly towards the parahippocampal cortex and laterally beyond the collateral sulcus, we also performed a manual adjustment. We created ROIs of distal CA1/proximal subiculum by splitting each of the two hippocampal structures in half along its proximodistal axis. Of all voxels encompassing CA1, the half located distally was included, and of all the voxels encompassing subiculum, the half located proximally was included: these two halves thus make up what we here define and refer to as 'distal CA1/proximal subiculum' (dCA1pSub). To create RSC and OFC ROIs, respectively, the FreeSurfer parcellations named "isthmus cingulate" and "lateral orbitofrontal" were used as a starting

DTI analysis
All DTI analyses were performed in the participant's native diffusion space. Voxel-wise fiber orientation distribution functions (fODFs) were computed by running the FSL function bedpostx on the diffusion data, using the zeppelin deconvolution model, a Rician noise model, and burn-in period 3000 (Sotiropoulos et al., 2016). Probabilistic tractography between the EC and presubiculum, dCA1pSub, RSC and OFC ROIs was then performed by running FSL's probtrackx2 on the fODFs (Behrens et al., 2007;Behrens, Woolrich, et al., 2003). Tractography was performed both in ROI-by-ROI and voxel-by-ROI connectivity mode, with number of samples 250,000, minimal path length 5 mm, and a midline termination mask (Behrens, Johansen-Berg, et al., 2003;Ezra et al., 2015;Johansen-Berg et al., 2004;Máté et al., 2018;Saygin et al., 2011

MEC and LEC segmentation
The voxel-wise connectivity maps were normalized to [0,1] by dividing by the maximum probability for each hemisphere separately, and then thresholded by 0.01 to reduce false positive connections (Behrens, Johansen-Berg, et al., 2003;Saygin et al., 2011). This threshold was determined empirically by testing a range of thresholds and choosing the one that in most cases removed connections outside the grey matter, because due to remaining distortions in the DTI images some of the EC ROIs slightly extended into white matter and air voxels. Crucially, we then define the MEC as the region that is most strongly connected with the presubiculum and/or RSC, while the LEC is the region that is most strongly connected with dCA1pSub and/or OFC (Caballero-Bleda & Witter, 1993;Honda & Ishizuka, 2004;Hoover & Vertes, 2007;Jones & Witter, 2007;Kondo & Witter, 2014;Saleem et al., 2008;Witter & Amaral, 1991;Witter & Amaral, 2021;Wyss & Van Groen, 1992). For each participant, a hard segmentation was performed on the normalized and thresholded voxelwise connectivity maps using FSL's find_the_biggest (Behrens, Johansen-Berg, et al., 2003;Johansen-Berg et al., 2004), where the voxels that had a stronger connection probability with the presubiculum/RSC than with dCA1pSub/OFC were classified as MEC, and vice versa for LEC.

Group analysis
Group probability maps of the connectivity paths between the ROIs, as well as group where the connectivity maps for presubiculum + RSC and for dCA1pSub + OFC, respectively, were combined and averaged before segmentation.

Segmentation comparisons
To assess the different segmentation approaches and compare the resulting locations of MEC and LEC, we calculated the orientation of the MEC-LEC border along the posterioranterior (PA) and medial-lateral (ML) axes, respectively. This was performed by first calculating the centers of gravity of the differently defined MECs and LECs, and the vector between these centers of gravity. Next, the angle between this vector and a pure PA or ML vector was determined. The degree of PA-or ML-oriented border was then defined between 0 and 100% such that an angle of 0° to the PA or ML vector means that the border is 100% oriented along the PA or ML vector, respectively. Correspondingly, an angle of 90° would mean that the border is 0% oriented along the respective axis, i.e. it is orthogonal to that axis. In addition, the different segmentations were compared with respect to the sizes of the resulting MECs and LECs, and the size ratios between these were calculated. All these segmentation comparisons were also carried out on the two fMRI-based segmentations of pmEC and alEC available for download from earlier studies (Maass et al., 2015;Navarro Schröder et al., 2015).