Diffusion MRI of the Hippocampus

Diffusion MRI as a probe for hippocampal microstructure

Diffusion magnetic resonance imaging (dMRI) is a technique that sensitizes the MRI signal to the diffusion of water (Brown, 1828) which permits the noninvasive analysis of microscale properties. In the mammalian brain, these microscale structures (generally referred to as microstructure) encompass axons, dendrites, and cell bodies (including the glia). While histology provides a ground-truth reference of microstructure, the promise of dMRI lies in its noninvasive nature and the ability to capture microstructure in 3D. Many methods have been developed that aim to extract measures of microstructure (such as neurite density) via explicit models or to quantitatively summarize the diffusion process without making assumptions about the underlying microstructure (see Diffusion and MRI Background, Diffusion MRI models). Regardless of the method used, most work conducted with dMRI has focused on characterization and validation of white matter. However, with recent advancements in hardware and software, there has been great interest in leveraging the dMRI signal for characterization of gray matter as well (Palombo et al., 2020; Jelescu et al., 2022).

The hippocampus is a structure in the medial temporal lobe that is largely but not exclusively composed of the gray matter. It plays key roles in cognitive processes such as episodic memory formation, consolidation, recollection, and spatial navigation (Voss et al., 2017). The importance of hippocampal structural integrity in these functions is highlighted by the impact hippocampal damage or degeneration can have in many neurological disorders (Small et al., 2011). For example, the well-characterized cell loss of the subiculum and cornu ammonis (CA) 1 hippocampal subfields in Alzheimer's disease has been reported to be linked to cognitive deficits (Small et al., 2011). The ability to probe hippocampal microstructure noninvasively can provide valuable insights into its typical neural structure supporting cognitive functions and deterioration in diseased states.

In this review we aim to highlight recent advances in research that have applied dMRI in the hippocampus of both human and nonhuman (primarily rodent) populations. We start by offering an overview of dMRI, including standard modeling techniques. We then discuss the complexity of hippocampal microstructure, providing a background for the kinds of tissue properties that can potentially be captured with dMRI. We then review the current literature, highlighting a wide array of dMRI methods that have been used to study hippocampal microstructure in the neurotypical brain and in neurological diseases. Subsequently, we discuss the relationships between different dMRI measures and histology. We end by summarizing open questions and by suggesting future directions in this research domain.

Diffusion MRI models

Diffusion MRI models are typically separated into two approaches: signal representations (aka model-free) and biophysical models (aka model based; see Fig. 2 for an overview) where model selection is limited by the dMRI acquisition (i.e., choice of diffusion time and b-value). Signal representations are largely based on the statistical properties of water diffusion and make no a priori assumptions about the underlying microstructure. While these signal representations are generalizable and well-defined mathematically, they tend to lack microstructural specificity (Jelescu et al., 2017).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Visual depictions of a variety of dMRI models. The left column shows signal-based approaches (model-free), including DTI (Basser et al., 1994), DKI (Jensen et al., 2005), and the Q-ball method (Tuch, 2004). DTI involves calculation of the diffusion tensor (D). The eigendecomposition of D allows for the calculation of fractional anisotropy and mean diffusivity. DKI quantifies the kurtosis or deviation from Gaussian diffusion, which can be more sensitive to microstructure barriers that restrict the movement of water. The Q-ball method uses the Funk–Radon transform (a spherical integration technique) to obtain the orientation distribution of diffusion that ideally corresponds to the orientation of the underlying microstructure (top left inset under Q-ball section). The right column shows biophysical (model-based) approaches, including NODDI (Zhang et al., 2012), SANDI (Palombo et al., 2020), NEXI (Jelescu et al., 2022), and CSD (Tournier et al., 2007). NODDI assumes there are three compartments, where f_* corresponds to the volume fraction of a compartment, d_* corresponds to the diffusivity of a compartment, and E_* is the diffusion signal arising from a compartment. There is a spherical free water compartment (thought to largely correspond to water in CSF), a tensor extracellular compartment, and a “stick” intracellular compartment. The orientation of the extracellular tensor and intracellular “stick” are linked together via a Watson distribution, which captures the orientation distribution with a mean orientation (μ) and dispersion (κ) parameter. The SANDI model adds an additional spherical compartment to capture intra-soma diffusion, with an additional soma radius parameter. The NEXI model drops the spherical soma compartment to measure the exchange time of water more accurately between the intracellular and extracellular space. CSD assumes that the diffusion signal is the convolution of the fiber (axon bundle) orientations with a fiber response function, which is the signal that arises from a single coherent axon bundle (for example, from the corpus callosum).

The most common signal representation is diffusion tensor imaging (DTI), which characterizes the molecular displacement of water as a 3D Gaussian distribution (Basser et al., 1994). In DTI the diffusion process is represented as a 3 × 3 matrix (i.e., a tensor), where its eigendecomposition provides useful metrics that describe diffusion. These include the diffusivity along three orthogonal directions (eigenvalues) and the primary orientation of diffusion (first eigenvector). From this eigendecomposition, metrics such as mean diffusivity (MD—mean of the eigenvalues) and fractional anisotropy (FA—variance of the eigenvalues) can be derived. These have been used extensively to interrogate microstructure. MD characterizes the average measured diffusion along three orthogonal directions, which would be expected to decrease in regions where there is more restriction (i.e., in white matter) and increased in regions where there is less restriction (i.e., cerebrospinal fluid). FA is thought to quantify the degree with which diffusion is preferentially oriented along a particular direction. Put differently, it quantifies how “pointed” the diffusion tensor is. For example, in the cylinder example in Figure 1, diffusion would be preferentially oriented along the cylinder's main axis, and as a consequence FA would be high. However, the complexity of neural tissue typically results in diffusion that cannot be well described by a single tensor. A common example is given in situations in which two fiber populations are crossing orthogonally to each other, where the measured FA is low while the true underlying anisotropic movement of water is high. Thus, a limitation of DTI is that it conflates orientation dispersion (e.g., crossing axons) and anisotropy.

Diffusion kurtosis imaging (DKI) also serves as a method for signal representation but aims to estimate the diffusion kurtosis that represents a departure from Gaussian diffusion. The diffusion kurtosis is effectively a measure of heterogeneity of the diffusion environment, suggesting that it has the potential to be more sensitive than DTI to microstructure (Jensen et al., 2005; Jensen and Helpern, 2010). However, since neither DTI nor DKI make explicit distinctions between microstructural environments, their interpretations in terms of specific microstructure are not straightforward (Jensen and Helpern, 2010).

Contrasting with signal representation methods, biophysical models assume a particular geometry of the underlying microstructure, with the goal of extracting the most meaningful features from the dMRI signal. Mathematical expressions of the diffusion signal are then derived based on the assumed geometry, with the typical assumption that water cannot exchange between these geometrical components (typically referred to as compartments; Jelescu et al., 2020). For example, diffusion in an intra-axonal compartment may be well represented by the geometry, size, and diffusion properties of a cylinder (Jelescu et al., 2020). While biophysical models can improve microstructure specificity, they can be ill-posed if the assumed geometry is not accurate. Most biophysical models were built for white matter and generally assume that there are two or three compartments in which diffusion is Gaussian. This overarching compartment framework is the so-called standard model of diffusion (Novikov et al., 2018). Neurites and potentially glia processes are represented as impermeable cylinders with zero radius (i.e., “sticks”). Diffusion in the extra-neurite space is described by a diffusion tensor, while a third optional compartment is the cerebrospinal fluid (CSF), which is modeled as a sphere representing completely free diffusion (i.e., no barriers or restrictions). The “sticks” are assumed to be distributed according to an arbitrary orientation distribution function (ODF).

The standard model provides a general framework through which many other biophysical models can be understood. Neurite orientation dispersion and density imaging (NODDI) is one such biophysical model that assumes three microstructural compartments for water diffusion: intracellular (sticks), extracellular (tensor), and cerebrospinal fluid (sphere; Fig. 2; Zhang et al., 2012). NODDI captures the primary orientation of diffusion and microstructure dispersion (i.e., the ODF) through a Watson distribution. Commonly derived parameters from NODDI include the neurite density index (similar to the intracellular signal fraction from other biophysical models) and the orientation dispersion index (ODI). The potential gain in specificity of such parameters is clear compared with the signal representations described above; however, validation remains a challenge (Jelescu et al., 2020). Previous research has found good agreement between NDI and measures of fiber content obtained via electron microscopy (Sepehrband et al., 2015) while another study found no significant association between MRI-derived myelin water fraction and NDI (Qian et al., 2020). As well, conflicting results have been found regarding how well ODI can capture histologically derived fiber dispersion (Grussu et al., 2017; Schilling et al., 2018). Understanding the direct microstructural correlates of parameters derived from biophysical models is a bourgeoning area of research.

While modeling gray matter tissue has many parallels to that of white matter, there are generally two distinctions that are made. Firstly, since myelin content is limited in the gray matter, there is the potential for water to exchange across the membrane of the neurites, which implies that the diffusion from each compartment is no longer additive or independent. Secondly, there can be signal contributions from the soma that are not represented in the standard model described above. Tackling the first issue, the neurite exchange imaging (NEXI) model builds upon the standard model by attempting to measure the exchange time between compartments (Jelescu et al., 2022). In contrast, the soma and neurite density imaging (SANDI) model adds a spherical compartment to account for the presence of the soma, providing measures of soma fraction and radius (Palombo et al., 2020).

To improve characterization of microstructural orientations, constrained spherical deconvolution (CSD) has been proposed to capture orientations through the fiber orientation distribution function (fODF), an alternative to the dODF described above (Tournier et al., 2007). The diffusion signal is assumed to be the convolution of the fODF with the fiber response function, which is the diffusion signal from a single coherent white matter bundle (Tournier et al., 2007). The shape of the fODF is generally sharper than the dODF and can be used to derive parameters with increased biologically saliency like apparent fiber density (AFD).

The above-described signal and biophysical models have been designed for single diffusion encoding (SDE) sequences, which probe diffusion along one direction (Fig. 1). However, there has been interest in going beyond SDE with double (planar) or triple (spherical) diffusion encoding, which can help disentangle features of various microstructure compartments (like orientation dispersion) within a voxel (Afzali et al., 2022).

Radial hippocampal microstructure

Radially oriented hippocampal microstructure (Fig. 3A,C) is defined as aligning with the orientation that crosses the cell layers or put differently, that is oriented orthogonal to the main hippocampal curvature. Generally, radial microstructure stays within the subfields. Within the hippocampus, the dominant radial structure is composed of the pyramidal neurons (Fig. 3B,C), which include somas in the stratum pyramidale layer, axonal projections, basal dendrites, and large apical dendrites that project toward the stratum moleculare. Notably, the microstructure of the stratum pyramidale has distinct transitions across the subfields, including changes to the density and shape of the somas (Fig. 3C; Braak, 1974; Amaral, 1978; Ding and Van Hoesen, 2015; Williams et al., 2023). The wide pyramidal layer of the subiculum contains large triangular and globular shaped somas, which tend to be more sparsely packed than CA1 (Rosene and Van Hoesen, 1987; Ishizuka et al., 2001; Williams et al., 2023). The large stratum pyramidale layer of CA1 includes somas that are largely triangular, relatively small, and interspersed with thin axons (Braak, 1974; Mouritzen Dam, 1979; Stephan, 1983; Ishizuka et al., 1995; Pyapali et al., 1998; Duvernoy, 2005; Ding and Van Hoesen, 2015; Williams et al., 2023). The narrow stratum pyramidale layer of CA2 contains large and oval shaped somas that are densely packed (Amaral et al., 1984; Duvernoy, 2005; Ding and Van Hoesen, 2015; Williams et al., 2023). The curved CA3 stratum pyramidale layer contains somas with a similar shape and slightly less packing density than CA2 (Braak, 1974; Insausti and Amaral, 2004; Ding and Van Hoesen, 2015; Williams et al., 2023). However, a unique feature of CA3 is the presence of unmyelinated mossy fibers which project from the DG (Fig. 3C; Ishizuka et al., 1990; Duvernoy, 2005). CA4 is oriented within the concavity of the DG, with large, very sparse, and oval shaped somas intertwined with partly myelinated mossy fibers (Mouritzen Dam, 1979; Duvernoy, 2005; Williams et al., 2023). Finally, the densely packed somas of the granular neurons in the DG are characteristically small and round (Blackstad, 1956; Williams et al., 2023).

The axons of the pyramidal cells project through the stratum oriens into the alveus, where they become a part of the fimbria and subsequently of the fornix, forming a large hippocampal white matter efferent (Daitz and Powell, 1954; Wyss et al., 1980; Duvernoy, 2005; Saunders and Aggleton, 2007). The basal dendrites of the pyramidal neurons make short local connections typically bending into the stratum oriens and alveus (Gloor et al., 1993; Duvernoy, 2005). The large apical dendrites project toward the stratum moleculare (Fig. 3B). The stratum moleculare layer contains few neurons and interneurons, and it contains the arborizations of pyramidal neuron apical dendrites (Fig. 3B; Ramon y Cajal, 1968; Duvernoy, 2005). Projections from the entorhinal cortex, which form the perforant path, enter the subiculum in a radial fashion before taking on a more complex curved orientation (Fig. 3C; Dolorfo and Amaral, 1998; Duvernoy, 2005; Zeineh et al., 2017). Finally, the majority of hippocampal interneurons appear to be radially oriented, although their components can adopt more complex orientations (Nieuwenhuys et al., 2008).

Tangential hippocampal microstructure

Tangentially oriented hippocampal microstructure (Fig. 3A,C) follows the curvature of the hippocampus; that is, it projects across the subfields. The perforant path is one of the major inputs to the hippocampus. Once perforating the subiculum (as mentioned in the section on radial microstructure), the perforant path has a tortuous trajectory, allowing it to project to the granule cells of the DG and the apical dendrites of the CA fields (Dolorfo and Amaral, 1998; Duvernoy, 2005; Zeineh et al., 2017). Importantly, a large part of this trajectory is thought to adopt a tangential orientation (Zeineh et al., 2017). Tangential projections from the dentate granule cells reach the pyramid cells of CA3 as either mossy fibers in the stratum radiatum or as the endfolial pathway in the stratum oriens (McLardy, 1962; Zeineh et al., 2017; Fig. 3C). A large branch of the axons from the CA3 pyramidal cells project to the alveus. At the level of the stratum oriens, coarse collaterals are derived from the axons that project to the stratum lacunosum or the stratum oriens in CA1 (Ishizuka et al., 1990; Nieuwenhuys et al., 2008). These collaterals are collectively known as the Schaffer collaterals, which largely project tangentially to the hippocampal surface. In turn, the CA1 pyramidal cells have projections to the apical dendrites of the subiculum (Knowles and Schwartzkroin, 1981; Nieuwenhuys et al., 2008). Finally, the subiculum and its constituents project back to the entorhinal cortex. This closes what is described above as a largely tangentially oriented pathway that is known as the trisynaptic circuit in the broader hippocampal literature.

Anterior–posterior and other hippocampal microstructure

While we have primarily focused on the radial and tangential hippocampal microstructure, there are aspects that are oriented along the anterior–posterior axis or that do not have a preferred orientation. A major anterior–posterior bundle is the fimbria, which sits atop of CA3, CA4, and the DG (Fig. 3C; Daitz and Powell, 1954; Zeineh et al., 2017). The thin fibers of the largely tangential but slightly anterior–posterior alveus converge to constitute the fimbria (Nieuwenhuys et al., 2008; Zeineh et al., 2017). At the posterosuperior tail of the hippocampus, the fibers of the fimbria enter the fornix, one of the major fiber systems that connect the hippocampus to other cortical and subcortical areas (Daitz and Powell, 1954). It is pertinent to note that while the information-carrying microstructure (i.e., the neurites and fiber pathways) are of great interest, other important microscale structures such as oligodendrocytes, astrocytes, microglia, and more can contribute to the diffusion MRI signal. Most of these properties tend to not have a preferred orientation and may have components that extend in many directions. The abundance of these structures greatly increases the microscale complexity of the hippocampus, which is discussed in the Concluding Remarks and Future Directions.

DTI metrics of FA and MD are sensitive but not specific to the microstructure of the hippocampal laminae

A large volume of studies have modeled the diffusion MRI signal within the hippocampus using DTI (Basser et al., 1994; see Diffusion and MRI Background, Diffusion MRI models). DTI-derived fractional anisotropy and mean diffusivity (FA and MD) have been two popular metrics to quantify diffusion at the voxel level. Both FA and MD have shown good contrast across the hippocampal laminae ex vivo, suggesting that both metrics may be sensitive to its underlying microstructural composition (Shepherd et al., 2006, 2007; Coras et al., 2014; Stolp et al., 2018; Fig. 4). FA was shown to provide noticeable differentiation of the alveus, stratum pyramidale, stratum lacunosum moleculare, and the DG molecular layer ex vivo (Shepherd et al., 2007; Coras et al., 2014; Figs. 3B, 4A,D). Correspondingly, MD was found to have an opposite contrast to FA, with MD being lower in regions dominated by axon architectures (fimbria and alveus) and higher in layers of neural soma and neuropil (stratum granulosum, stratum pyramidale, hilus, stratum moleculare, and stratum radiatum; Shepherd et al., 2006, 2007). In contrast to the above studies, it was found that MD could differentiate two cell-dense (CA1 pyramidale and DG granular layer) from two cell-sparse but projection-heavy (CA1 stratum radiatum and DG molecular layer) laminae, while there was no significant differences in FA (Stolp et al., 2018). The lack of sensitivity of FA to these laminae is likely a result of the high orientational heterogeneity of the hippocampal microstructure (Yoo et al., 2021). Interestingly, it has been shown that the DTI model most accurately captures the diffusion signal in areas of lightly myelinated axons, such as in the stratum radiatum moleculare (Jespersen et al., 2010). At high ex vivo resolutions (50–200 µm; Table 1), FA and MD generally appear to be useful in contrasting the hippocampal laminae, from highly myelinated axonal regions to the less restrictive neuropil, to the perikarya.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Illustration of hippocampal dMRI results based on some of the categories defined in the right column of Table 1 (i.e., Analysis types). For a brief summary of the presented results, see Summary of Key Findings. A, Adapted from Coras et al. (2014) with permission. The left panel represents a healthy control and the right panel represents a sclerotic hippocampus. Gray-scaled images represent FA maps. Insets represent histological slices stained for myelin basic protein. B, Adapted from Beaujoin et al. (2018) with permission. Top left depicts a T2w image with the zoom shown in the bottom left figure. I (alveus, str. orients), II (str. pyramidale), and III (str. lucidum radiatum lacunosum moleculare) correspond to the three laminar layers of the hippocampus. Top right represents the intracellular volume fraction map obtained with the NODDI model. C, Adapted from Khan et al. (2015) with permission. Coronal sections of a rhesus macaque hippocampus. Top left is the primary eigenvector map (V1) derived from DTI and scaled by FA. Top right is a 200 µm thick tissue obtained at 10× magnification using 3D confocal microscopy. Bottom left is a zoomed in depiction of the DTI color map depicted in the top left. Bottom right depicts one slice of the 3D confocal microscopy acquired at 63× magnification and digitized at 0.25 µm resolution. Arrows depict the 3D structure tensor analysis using kernel sizes of 0.5 and 3.0 µm to define the local region used for computing spatial derivatives. D, Adapted from Shepherd et al. (2007) with permission. Colored image is the FA-scaled primary eigenvector (V1) with transverse (blue), vertical (green), or through-plane (red) orientations. E, Adapted from Coras et al. (2014) with permission. Left column represents a healthy control and the right column represents a sclerotic hippocampus. Colored image represents DTI tractography. F, Adapted from Wu et al. (2016) with permission. Left figure is laminar tractography in CA1. Right figure is a similar slice with AAV tracer acquired using serial two-photon microscopy. G, Adapted from Wu et al. (2016) with permission. Left column shows stained coronal sections of MAP2 (dendrites and cell bodies) while the middle column shows similar slices of NF (axons) across the laminae in CA1 and the DG. The far-right column represents the fiber orientation distribution (fODF) maps derived from dMRI. H, Adapted from Ly et al. (2020) with permission. Top left represents a coronal slice of tractography from a human hippocampus. Bottom left represents combined immunohistochemistry seen on the right side of the same coronal slice. Right column represents immunohistochemical staining of DAPI (general cellularity), FOX3 (neurons), and GFAP (glial cells).

An improved understanding of hippocampal microstructure as seen through the lens of FA and MD has come through studying diseased states, in which known microstructural changes occur. In a rat model of epileptogenesis, FA was found to increase over time in the hilus and DG regions, suggesting that it may be useful in capturing the axonal plasticity of the mossy fibers after epileptic injury (Laitinen et al., 2010; Sierra et al., 2015). Similarly, a major microstructural hallmark of mesial temporal lobe epilepsy (TLE) is hippocampal sclerosis, defined by a loss of pyramidal neurons, reactive gliosis, and morphological changes of the DG. Probing sclerotic hippocampi ex vivo, it was found that FA was not sensitive to microstructural changes such as cell death and expanding extracellular space (Coras et al., 2014). MD was increased in the stratum pyramidale of sclerotic hippocampi versus controls, suggesting that it is likely sensitive to changes to the pyramidal somas (Shepherd et al., 2006, 2007; Coras et al., 2014; Stolp et al., 2018). In humans with TLE, it was found that MD elevations were regionally heterogeneous within the hippocampus, and this heterogeneity varied across TLE patients; these types of findings suggest that the measure may be capturing region-specific microstructural changes (Treit et al., 2019). In a group of Alzheimer's disease (AD) patients, MD was found to reflect aspects of the neurodegenerative process, with significant changes in the anterior head of the hippocampus, potentially reflecting perforant path disruption (Fellgiebel and Yakushev, 2011). In another study at ultra-high ex vivo resolutions, however, no significant changes in FA or MD were found within the layers of the DG between AD and control tissue (Shih et al., 2023). While nonsignificant, a decreasing trend in MD was found for AD tissue, which seemingly contradicts previous research in which known hippocampal microstructural degradation increased MD (Coras et al., 2014; Stolp et al., 2018; Treit et al., 2019; Shih et al., 2023). The authors proposed that DTI measures are likely sensitive to many different barriers that hinder water movement, including axonal swelling, accumulation of intracellular NFT, amyloid plaque deposition, and microvascular changes (Shih et al., 2023). In summary, the DTI-based scalar metrics appear to have only limited value in capturing hippocampal circuitry and laminar microstructure. The lack of specificity to any particular cellular organization and inability to represent complex microstructure configurations limits their use in drawing conclusions about any specific microstructural property (Beaujoin et al., 2018). The general limitations of DTI, including those mentioned here, have been extensively acknowledged (Mori and Zhang, 2006).

Biophysical modeling is a promising avenue for improving characterization of hippocampal microstructure

Due to the known shortcomings of DTI, there have been several diffusion-based studies employing increasingly complex signal representations like DKI (Jensen et al., 2005) and biophysical models to improve tissue specificity. A popular dMRI biophysical model is the NODDI (Fig. 2) model, which provides metrics of orientation dispersion and intracellular volume (Zhang et al., 2012). At high ex vivo resolutions, the intracellular volume fraction was shown to provide excellent contrast (i.e., sensitivity) to the hippocampal laminae (Beaujoin et al., 2018; Fig. 4B). Furthermore, heterogeneity in the intracellular volume fraction was seen along the anterior–posterior axis, supporting the idea of hippocampal long-axis differentiation (Beaujoin et al., 2018). Another study using NODDI and surface-based hippocampal modeling in vivo found the intracellular volume fraction was significantly correlated with a proxy measure for myelin content (Karat et al., 2023). Furthermore, the subfields were found to be highly separable based on measures of orientation dispersion and intracellular volume fraction, suggesting that these biophysical measures may be sensitive to subfield variability in cyto- and myeloarchitecture (Karat et al., 2023). Further biophysical models were developed to describe diffusion in gray matter (Jespersen et al., 2010). By applying these models in rats, it was found that the cylinder density parameter to capture neurites was very low in the stratum radiatum, which contains nonmyelinated axons and dendrites (Jespersen et al., 2010). Due to the short exchange time of water between dendrites and glia to the extracellular space, it is likely that the cylinder density parameter reflected the sparsity of myelinated axons in this area. It was found that these two models describe the diffusion signal better than DTI (Jespersen et al., 2010). Similarly, a more detailed biophysical model was built for diffusion in gray matter, with the goal of characterizing astrocytes and microglia (Garcia-Hernandez et al., 2022). It was found that the parameters prescribed for the microglia and astrocytes were sensitive only to their corresponding activation (Garcia-Hernandez et al., 2022). However, the microglia stick fraction was sensitive to myelin, which is corroborated by Jespersen et al. (2010). Thus, biophysical modeling is a promising avenue for providing more complete descriptions of hippocampal microstructure, including laminar differentiation, myelinated fibers such as the perforant path, and the abundant glial cells.

A handful of techniques other than DTI and biophysical models have been applied in the hippocampus. AFD derived from CSD (Tournier et al., 2007; Fig. 2) was shown to provide better delineation than DTI of the CA1 pyramidale layer and DG granular layer, which are cell-dense regions, versus the CA1 stratum radiatum and DG molecular layer, which are cell-sparse but projection-heavy layers (Stolp et al., 2018).

Acquisitions with ultra-short diffusion times have shown significant sensitivity to hippocampal microstructure injury

Using oscillating-gradients has allowed for dMRI acquisitions of ultra-short diffusion times, altering the sensitivity of the diffusion signal to local tissue microenvironments. Examining the change in apparent diffusion coefficients (ADCs) across varying diffusion times using oscillating gradient spin-echo (OGSE) sequences has shown unique laminar-specific contrasts in the mouse hippocampus (Aggarwal et al., 2013). Furthermore, in response to hippocampal injury, a significant decrease in the change in ADC across diffusion times was found to be specific to the pyramidal and granule cell layers in OGSE-based work (Aggarwal et al., 2013). Similarly, an OGSE and standard PGSE sequence has been employed to model the diffusion kurtosis in injured and control mouse brains (Wu et al., 2018). The change in measured kurtosis (ΔMk) between the two sequences was readily visible in the hippocampus, which may reflect a difference in the scales of the microstructure that is restricting water diffusion (Wu et al., 2018). For example, the 10 µm radius of the mouse pyramidal neurons is commensurate with the mean displacement of water molecules at the diffusion times of the OGSE sequence but shorter than the scales probed by the PGSE sequence (Wu et al., 2018).

Microstructural orientation dispersion is likely abundant in the hippocampus

Moving from standard linear tensor encoding (LTE) to spherical tensor encoding (STE) has allowed for the characterization of microscopic fraction anisotropy (uFA), which can help disentangle anisotropy within microscopic compartments from variability in size and orientation dispersion. Using STE and LTE in humans, uFA has been found to be much higher than FA, suggesting there is substantial orientation dispersion in hippocampal gray matter in which crossing fibers are abundant (Yoo et al., 2021; Fig. 3C).

The angular richness of the dMRI signal can provide distinct hippocampal microstructural orientations

Along with dMRI quantitative modeling, there has been a plethora of research and models that have capitalized on the angular richness of the dMRI signal. Distinct microstructural orientations have been found within the hippocampus by examining the primary eigenvector from DTI (V1; Fig. 2). At high ex vivo resolutions, the stratum oriens of CA1 and CA3, along with the alveus, have a tangential V1 orientation that is likely a result of axonal projections from the pyramidal neurons bending into the alveus (Shepherd et al., 2007; Figs. 3B,C, 4D). Radial V1 orientations were found within the stratum pyramidale and radiatum, which is thought to represent the large apical dendrites from the pyramidal cells (Shepherd et al., 2006, 2007; Figs. 3B,C, 4D). Similarly, the molecular layer of the DG showed considerable radial orientations, likely resulting from the projections of granule cell dendrites (Shepherd et al., 2007; Fig. 4D). Orientations in the stratum lacunosum moleculare are less distinct than the other regions, which may be due to large amounts of orientational heterogeneity (Shepherd et al., 2007; Figs. 3B,C, 4D). Coinciding with the previous result, within CA1 a radial V1 orientation was found within the stratum pyramidale, while a tangential orientation was found within the alveus (Khan et al., 2015; Fig. 4C). By inducing epilepsy in rats, it was shown that after 10 d, the orientation of V1 was altered, suggesting that it is sensitive to underlying hippocampal microstructural changes (Sierra et al., 2015). Using the mean orientation of the Watson distribution provided by NODDI (which is analogous to V1), it was found that diffusion measured in vivo tended to be oriented along the radial, tangential, or anterior–posterior axes, potentially capturing the known stereotyped organization of the hippocampal microstructure (Karat et al., 2023; Fig. 3). However, due to the complex arrangement of hippocampal microstructure, the full diffusion tensor ellipsoid has been shown to be largely spherical, limiting the use of V1 in capturing distinct microstructural orientations, in particular in regions of high orientational heterogeneity (Jespersen et al., 2010; Laitinen et al., 2010). Thus, a handful of studies have gone beyond DTI and used models that can provide accurate orientations even in such regions. The biophysical model developed in Jespersen et al. (2010) appears to capture the more complex arrangement of hippocampal microstructure. Two clear crossing fiber populations thought to represent the radial pyramidal neurites and Schaffer collaterals were found at the level of CA3 using the Q-ball method (Tuch, 2004; Beaujoin et al., 2018; Fig. 2). Similarly, clear shifts in microstructural orientation across the laminae were found using CSD (Wu and Zhang, 2016; Stolp et al., 2018; Figs. 2, 4G). Overall, these studies demonstrate the utility of dMRI orientation estimates for specification and delineation of hippocampal microstructural properties.

By leveraging the dMRI orientations, it is possible to estimate continuous structural connections across extended space. This process is often referred to as tractography and involves propagation of connections (called streamlines) in consecutive steps by following peaks in the estimated orientations. It is a particularly useful technique when attempting to capture intrahippocampal circuitry noninvasively. Clear streamline orientation shifts were found between laminae in humans and mice by using both DTI and CSD to perform tractography (Wu and Zhang, 2016; Ly et al., 2020; Fig. 4H). Similarly, orientations from DTI, CSD, and Q-ball tractography appeared to show approximate trajectories that may reflect the perforant path, Schaffer collaterals, mossy fibers, and other components of the trisynaptic and monosynaptic intrinsic hippocampal pathways (Shepherd et al., 2006; Augustinack et al., 2010; Zeineh et al., 2012; Modo et al., 2015; Beaujoin et al., 2018; Ly et al., 2020; Fig. 3C). Tractography has also been applied to the hippocampus in disease states. Streamline differences were found when comparing a sclerotic and nonsclerotic hippocampus in TLE, suggesting a potential shift in the underlying microstructural orientations and larger scale intrahippocampal circuitry (Coras et al., 2014; Fig. 4E). A similar comparison of streamlines between an AD and control hippocampus showed distinct regions of disconnectivity and orientation shifts (Shih et al., 2023). It should be noted that some previous studies have used anisotropic voxel resolution for performing tractography, which can be limiting given the potential for preferential averaging of orientations along a given axis (Jones and Leemans, 2010). This can potentially bias measured connectivity between two subfields. Regardless, tractography appears to be a promising technique to elucidate the intricate 3D structure of hippocampal microcircuitry (Fig. 3C).

DTI metrics correlate with many histological properties of the hippocampus

While the previously mentioned methods may be useful to model hippocampal microstructure in healthy and diseased states, the relationships of resulting measures to the underlying ground-truth microstructure still remain poorly understood at present. As such, a wide array of histological and optical imaging methods have been employed in combination with dMRI so as to better understand the sensitivity and specificity of the latter to hippocampal microstructure. Beginning with DTI, a positive correlation between histological staining of myelin and FA was found, with both FA and myelin content being high in the alveus and fimbria and low in the neuropil and pyramidal cell layers (Laitinen et al., 2010; Coras et al., 2014; Fig. 4A). However, FA has also been found to be elevated in regions of high astrocyte density (Stolp et al., 2018). In a rat model of epilepsy, FA was shown to be increased over time in the hilus and DG regions, while histology revealed a loss of pyramidal cells, gliosis, and activated astrocytes, making it unclear what microstructural changes the FA measure might be sensitive to Sierra et al. (2015). Similarly, MD was shown to have high spatial correspondence with histological markers of neuronal cell density and gliosis in temporal lobe epilepsy (Treit et al., 2019). Within the granular and pyramidal cell layers, MD was found to be correlated with general cell staining and glia (Stolp et al., 2018). Furthermore, V1 orientation was found to change across laminae when the microstructural composition changed, as shown by a general cell stain, a neuron cell stain, and a glial cell stain (Ly et al., 2020; Fig. 4H). Within a small slice of CA1 and the alveus, it was shown that V1 matches the orientations uncovered with light microscopy stained for Dil (nonspecific membrane labeling), suggesting that ultra-high-resolution DTI may be able to capture the general orientation of the ensemble microstructure (Khan et al., 2015; Fig. 4C). Conversely, using a more specific Golgi stain to label hippocampal dendritic arborizations, it was found at lower resolutions that the DTI ellipsoids were generally spherical and not able to capture the true orientation complexity of the dendrites (Jespersen et al., 2010). Overall, DTI-derived measures tend to correlate with many tissue properties but converging evidence for specificity in these mappings remains limited.

Metrics of dMRI other than DTI correlate with more specific aspects of hippocampal microstructure

Moving beyond DTI, it was found that within the stratum radiatum, there was a distinct radial orientation of the fODF calculated using CSD (Fig. 2), which was reflected in histological staining of axons, astrocytes, and blood vessels (Stolp et al., 2018). Counterintuitively, AFD was negatively correlated with the staining of axons (Stolp et al., 2018). The fODF was further shown to have regions of noticeable correspondence with axon and dendrite staining, suggesting that it can capture the basic organization of the hippocampal axonal and dendritic networks (Wu and Zhang, 2016; Fig. 4G). Within the stratum radiatum of CA1, the large radial fODF lobes were parallel to the densely populated dendrites as shown on MAP2 stains (Wu and Zhang, 2016; Fig. 4G). In the stratum lacunosum moleculare of CA1, there were large tangential fODF lobes which were parallel to axons believed to be the Schaffer collaterals (Wu and Zhang, 2016; Fig. 4G). Diffusion MRI tractography within CA1 was found to closely correspond to the spatial projection pattern uncovered by tracer injection (Wu and Zhang, 2016; Fig. 4F). Using OGSE, it was found that differences in diffusivity across a range of diffusion times corresponded to apparent cell loss (H&E staining) and neurodegeneration and reactive gliosis (Fluoro-Jade C staining) within the pyramidal and granular cell layers (Aggarwal et al., 2013). Finally, a detailed biophysical model built for characterizing glia in gray matter was shown to be sensitive and specific to astrocyte and microglia alterations (Garcia-Hernandez et al., 2022). By inducing neuroinflammation without neurodegeneration, a fast microglia response was found, with increased cell body size, dispersion, and retracted cellular processes that were mirrored by the microglia parameters of the biophysical model (Garcia-Hernandez et al., 2022). As well, 24 h after inducing neuroinflammation, the astrocytes grew in volume, a result that was also reflected in the astrocyte parameter of the model (Garcia-Hernandez et al., 2022).

Outstanding questions

Here we highlight some outstanding and interesting questions that arose as a result of the synthesis of the reviewed studies.

1. Can biophysical models specifically developed for gray matter (which incorporate water exchange and/or a soma compartment) improve specificity to hippocampal microstructure?

2. To what degree is the hippocampal microstructure oriented radially and tangentially? Can dMRI metrics be related to specific hippocampal microstructure based on previous knowledge of their orientations? Finally, to what extent do hippocampal microstructural orientations vary across individuals and can distinct changes in microstructural orientations be discerned in diseased states?

3. As MRI hardware improves and stronger gradients become increasingly available, shorter diffusion length scales can be probed. Can diffusion time be leveraged to capture particular aspects of hippocampal microstructure that exist at various spatial scales?

4. Is tract-tracing a viable method for validating the sensitivity of diffusion tractography to hippocampal circuitry?

5. To what degree are the conclusions drawn from hippocampal dMRI affected by resolution? At what resolution does extrahippocampal myelinated microstructure dominate the dMRI signal?

6. Are there consistent hippocampal long-axis (anterior–posterior) differences in the dMRI signal across different models?

7. Does microscopic fraction anisotropy (uFA) provide contrast to the laminae and subfields?

8. Does hippocampal uFA have a dominant histological correlation?

9. Particular regions of the hippocampus were found to have unique cellular compositions. For example, the subiculum and CA1 were shown to have lower amounts of glial cells. Can a spatially adaptive hippocampal dMRI model improve microstructural specificity by taking advantage of known microstructural compositions?

10. What are the relationships between microstructural properties revealed with higher-order dMRI modeling and specific cognitive processes known to be supported by the hippocampus? How are these relationships altered by disease states that affect hippocampal integrity?