We present novel hippocampal maps in a standardized folded and unfolded space for each of the datasets outlined above. This includes 30 distinct group-averaged maps which have been attentively preprocessed and curated. Within each methodology, some interpretation and summarization via dimensionality reduction is offered, and finally we compare all maps across methodologies in the “Feature combinations” section.
Histology
Histology is considered a neuroanatomical gold standard, and is the basis for most parcellations and descriptions of brain regions (Amunts et al., 2020; Brodmann, 1909; Eickhoff et al., 2018; Paquola et al., 2019). Here we examined data collected from BigBrain Merker staining for cell bodies (Amunts et al., 2013), 3D polarized light imaging (PLI) of neural processes (M. Axer et al., 2011), and the AHEAD dataset with different stains serving as markers of neurons, myelin, and subtypes of interneurons (Alkemade et al., 2022) (Figure 2A). Most features showed banding in the proximal-distal direction, in alignment with the subfield atlas shown in Figure 1.
Microstructural (or laminar) profiles are shown for five ROIs across the proximal-distal axis of the BigBrain Merker stain (Figure 2B). They show a tight unimodal distribution in the distal CA fields, and a more bimodal distribution in the subiculum as expected based on their known laminar architectures (Duvernoy et al., 2013). Profiles for all vertices were concatenated across all stains to make multimodal profiles, a common method for characterizing laminar structure (Schleicher et al., 1999). Next, multimodal covariance matrices between vertices were calculated (mMPC matrix) (Figure 2C). Diffusion map embedding, a non-linear dimensionality reduction technique (Coifman et al., 2005; Margulies et al., 2016; Vos de Wael et al., 2020), decomposed the mMPC matrix into primary components, or gradients, that highlighted the differences between vertices with respect to all modalities and depths. In the first gradient, a sharp boundary was seen between the subicular complex and proximal CA1 and the rest of the hippocampus. The second and third gradients in turn highlighted the CA2-3 regions and CA1 with parts of the subiculum, respectively. This is data-driven evidence that subfields across the proximal-distal extent of the hippocampus, rather than anterior-posterior or other patterns, account for structural variance in the hippocampus with respect to these stains. These data-driven decompositions, thereby, echo classical and recent neuroanatomy descriptions of hippocampal microstructure (Ding & Van Hoesen, 2015; Duvernoy et al., 2013; Olsen et al., 2019).
Structural MRI
MRI is a key tool for studying human neuroanatomy and structure-function relations due to its non-invasive nature and potential for biomarker discovery. 7 Tesla (7T) and ex-vivo 9.4T scanning are especially powerful, achieving greater resolution and contrast than typical 3T or 1.5T clinical scans (Duyn, 2012; Opheim et al., 2021). Here, we provide healthy normative maps for such scans (Figure 3A) including popular acquisitions: quantitative T1 relaxometry (qT1) and its non-quantitative ex-vivo inverse: R1, T2* and its inverse R2*, proton density, diffusion weighted imaging (DWI) estimates of fractional anisotropy (FA) and apparent diffusivity coefficient (ADC), and magnetic transfer ratio (MTR).
Multiple scans were available for averaging (n=4 left+right hippocampi at 9.4T and n=20 left+right hippocampi at 7T), enabling a calculation of consistency across samples via Pearson’s R (Figure 3B). DWI and qT1 maps were also calculated in a second validation dataset, consisting of 82 locally scanned healthy participants (including the subset from the MICA-MICs dataset) with a 3T scanner, which showed similar patterns (Figure S1). mMPCs were generated as above and were reduced using diffusion map embedding into primary gradients, which again highlighted differences across subfields. Only the third gradient showed anterior-posterior differences, largely within the CA1 subfield.
Functional MRI
Functional MRI during the resting state (rsfMRI) allows interrogation of intrinsic brain function via the analysis of spontaneous activity and its statistical dependencies, and has become a key technique in the mapping of functional-anatomical systems (Biswal et al., 1997; Buckner et al., 2008; Smith et al., 2009). Here, we examined several features of rsfMRI in 88 healthy participants scanned at 3T. Intrinsic timescale is a measure of the time it takes for the temporal autocorrelation to drop below a threshold (Golesorkhi et al., 2021; Wolff et al., 2022) (Figure 4A). On a functional level, this is thought to be driven in part by recurrent connections that maintain activity patterns on the order of seconds (Fallon et al., 2020). Regional homogeneity considers the similarity between adjacent vertices’ time series, which is thought to indicate the extent of horizontal (i.e., between cortical columns) excitatory connectivity (Zang et al., 2004)(Figure 4B). Finally, macroscale functional connectivity is by far the most popular rsfMRI feature, with many rich properties that have been explored with respect to white matter connections (Damoiseaux & Greicius, 2009; Greicius et al., 2009; Honey et al., 2009), network properties (Schmittmann et al., 2015; van den Heuvel & Sporns, 2013), organizational gradients (Bernhardt et al., 2022; Margulies et al., 2016; Paquola et al., 2019; Park et al., 2021), and many other summary metrics. For simplicity, we examined connectivity between all hippocampal vertices and neocortical parcels from the Schaeffer400 parcellation (Schaefer et al., 2018)(Figure 4C). The consistency of maps was examined as above, and all measures were significantly greater than zero. Repetition of these analyses in a smaller sample of 7T rsfMRI data showed consistent results (Figure S2).
As mentioned above, functional connectivity is a rich measure that can be summarized in many ways. Here, we identified gradients of intrinsic hippocampal connectivity variations (Figure 4D) using the aforementioned non-linear decomposition techniques. Consistent with previous work (Genon et al., 2021; Poppenk et al., 2013; Przeździk et al., 2019; Strange et al., 2014; Vogel et al., 2020; Vos de Wael et al., 2018), we found anterior-posterior differentiation in the first hippocampal gradient, together with proximal-distal banding with CA1 in particular differing from the other subfields. Neocortical counterparts of this gradient show that anterior and CA1 regions shared more connectivity with temporal pole, insula, and frontal regions whereas more posterior and non-CA1 subfields shared connectivity with more posterior parietal and visual areas, again consistent with previous findings (Vos de Wael et al., 2018). The second gradient also showed differentiation of CA1 from subiculum and CA2-3 in the more middle and posterior regions, with neocortical correspondences to medial prefrontal and posterior cingulate regions for CA1 and more visual areas for CA2-3 and posterior subiculum.
Intracranial EEG
Invasive recording methods such as iEEG provide a direct measure of neural activity at high temporal resolution, but typically have lower spatial coverage and are limited to neurological patient populations. In that sense, they can be considered as scattered spatial data, which can be interpolated or extrapolated for hippocampal mapping as described in Figure 1C, or following previous approaches (Frauscher et al., 2018). We employ common measures of the periodic component of iEEG data, as shown by power spectrum density and additionally further simplified to Delta, Theta, Alpha, Beta, and Gamma band powers from low to high frequencies, respectively. Power spectrum densities and band powers derived from hippocampal channels resembled those derived from all channels (Figure 5A). Extrapolating channel information across neighbouring vertices from a given hippocampus, a spatial pattern emerged in which both proximal-distal and anterior-posterior differences were seen (Figure 5B). Band power is a limited measure of the full power spectrum density though, and so in Figure 5C we performed gradient mapping of the full power spectrum density. This showed a primary anterior-posterior gradient driven by higher Theta and Alpha power in the posterior and higher Delta power in the anterior hippocampus. The second gradient showed increased Delta power in the anterior and posterior hippocampus, while the third gradient showed a slight increase in Delta and decrease in Theta in the subiculum. Results were consistent when using an open iEEG atlas (Frauscher et al., 2018) or locally collected data in patients (Paquola, Seidlitz, et al., 2020), showing largely conserved patterns in Figure S3.
Feature combinations
The biggest advantage of a common hippocampal mapping space is that it allows for direct spatial correlation between features from different scales and methods. In Figure 6A, we examined relationships between all maps shown above using Pearson’s R with an adapted spin test significance testing to control for spatial autocorrelation in the data. This revealed many greater-than-chance correlations, both within methodologies and between. Finally, we additionally compared morphological measures of thickness, gyrification, and curvature which are generated within the HippUnfold workflow (Figure S4). Previous work (J. DeKraker et al., 2020) showed that these features differed between MRI and histology, with the latter showing greater detail including more gyrification and lower thickness. After group-averaging, each of these features was significantly spatially correlated between histology and MRI.
We performed a dimensionality reduction across all features from all figures using diffusion map gradient embeddings. For visualization, we plotted components 1 and 2 with colour coding according to subfield and continuous anterior-posterior and proximal-distal gradients (Figure 6C). The proximal-distal and anterior-posterior axes of the hippocampus are closely aligned to gradients 1 and 2, respectively, with gradient 1 explaining approximately twice the variance (Figure 6B). This suggests that while these two axes emerge as natural summaries of many hippocampal feature maps, the proximal-distal direction is stronger.
Figure 6D provides a summary of which measures are most correlated with the anterior-posterior and subfield axes of the hippocampus. As expected, the strongest subfield relationships were observed in histological features such as Calbindin and Calretinin staining, or thickness measures at a histological level of precision. Many structural 9.4T and 7T features also showed strong subfield correlations, especially qT1 and qR1. This is encouraging given the increasing availability and adoption of quantitative T1 sequences (Bidhult et al., 2016; Haast et al., 2016; van der Weijden et al., 2021). The employed rsfMRI and iEEG features were still moderately correlated with subfield division, but iEEG and rsfMRI gradients in particular showed strong correlations with the anterior-posterior hippocampal axis. Some caution should be exercised here: iEEG data were sparsely sampled and so after extrapolation each band power map was very smooth, which could amplify correlation values (but not significance, since spin test permutations were used to control for spatial autocorrelation). Note also that laminar profiles were not used in this analysis, and histological measures in particular can benefit from the information added by such methods due to their high precision.
Usability experiment and documentation
HippoMaps as an open toolbox and online data warehouse paves the way for multiple new research avenues, examples of which are shown in Figure 7. We anticipate that as hippocampal mapping studies are performed in other research areas, authors can use the initial maps provided here as comparisons and will upload their own maps in the spirit of open and reproducible science, and also to boost the visibility of their work. To this end, we provide a set of Python tools, well documented example code to reproduce the maps shown here (labeled as tutorials), and guidelines for how other experimenters should upload their maps to this repository. We have and will continue to answer questions and create community resources via GitHub (https://github.com/HippAI/hippomaps or https://github.com/MICA-MNI/hippomaps), and all current maps are available on the Open Science Framework (https://osf.io/92p34/).
Figure 7A shows a generic use case of HippoMaps wherein a new finding is contextualized by comparison to other maps in HippoMaps, and data is in turn contributed to HippoMaps to extend its utility in future work. Figure 7B illustrates an example experiment with task-fMRI using the Mnemonic Similarity Task (MST) designed to probe pattern separation, a task thought to preferentially involve hippocampal subregions (Pishdadian et al., 2020; Stark et al., 2019). This can be seen most strongly in subiculum for the successful pattern separation trials, whereas trials with novel stimuli showed anterior-posterior differentiation. Comparing these maps directly to microcircuit features provides context for the demands of these two task conditions: pattern separation was strongly correlated to detailed maps of curvature, thickness, and neocortical connectivity, whereas novelty was moderately correlated to intrinsic timescale, beta band power, and gamma band power (Figure 7B, right). Further task-fMRI results from an object-pairing memory task, as well as replication data of the MST at 7T, are shown in Figure S5.
Figure 7C illustrates an example experiment comparing 35 temporal lobe epilepsy (TLE) patients to 81 healthy, age- and sex-matched controls scanned at 3T MRI. Reductions in hippocampal thickness and gyrification are seen, with the greatest changes in CA1 and CA4 subfields, which have previously been identified as vulnerable areas (Blümcke et al., 2012, 2013; Duvernoy et al., 2013; Steve et al., 2020). Comparing thickness reduction patterns to other maps shows moderate correlations with rsfMRI properties of intrinsic timescale, neocortical connectivity, and histological Bieloschowsky staining. Gyrification loss was strongly correlated with healthy gyrification in histology and 7T MRI, and iEEG delta band power.