Over the past decade, investigators have begun to implement high spatial resolution functional magnetic resonance imaging (hr-fMRI) of human MTL, in which anatomical landmarks correlating with MTL cytoarchitectonics are used to guide segmentation of the MTL into component hippocampal subfields and surrounding cortical structures. Guided by MTL atlases (Duvernoy, 2005; Insausti and Amaral, 2004) and standards developed in structural MRI studies of the MTL (Insausti et al., 1998; Pruessner et al., 2002; Pruessner et al., 2000), investigators can now routinely segment human hippocampus into the subiculum, CA₁, and a combined subregion consisting of the dentate gyrus, CA₂, and CA₃ (DG/CA_2/3; these subfields are difficult to unambiguously segment at current functional MR resolutions). Moreover, the anterior extent of the parahippocampal gyrus can be divided into entorhinal cortex (ERC) and perirhinal cortex (PRC), and differentiated from parahippocampal cortex (PHC) in the posterior portion of the gyrus (Figure 1a).

The use of hr-fMRI to explore human MTL subfield function was first described by two research teams almost a decade ago. Small et al. (2000a, 2000b) examined resting-state activity using conventional T2*-weighted gradient-echo images at 1.5T (5 slices, 0.9 × 0.9 × 5 mm resolution), and Zeineh et al. (2000) examined MTL novelty encoding effects using a T2*-weighted gradient-echo echo-planar imaging (EPI) sequence at 3T (16 slices, 1.6 × 1.6 × 3 mm resolution). In both studies, acquisition was perpendicular to the long axis of the hippocampus, yielding high resolution in the coronal plane of the hippocampus and surrounding cortex, and thus affording increased substructure visualization. During data analysis, participant-specific anatomically defined regions of interest (ROIs) were demarcated, using either the high-resolution functional images themselves (Small et al., 2000a, 2000b) or a separate set of co-registered structural images with even higher spatial resolution (0.4 × 0.4 × 3 mm; Zeineh et al., 2000). Average signal intensity was extracted from each subfield from each participant, and then submitted to group-level (across-participant) statistical analysis. Moreover, to better visualize subfield boundaries and the topography of functional activation in the MTL, Zeineh and colleagues developed a method for flattening the gray matter MTL ribbon into a two-dimensional representation (see Figure 1b for a recent demonstration of the flat map technique), adapted from flat map methods routinely used to study visual cortex (e.g., Van Essen et al., 1998).

Since these initial reports, most subsequent hr-fMRI studies of human MTL have been conducted at 3T, and have collected both high-resolution functional and structural volumes, with the latter having superior resolution that allows for demarcation of anatomical ROIs. To ensure that these higher-resolution ROIs are properly aligned with the corresponding regions in the functional data, researchers typically use one of several standardized within-participant linear co-registration tools. Moreover, to preserve the high resolution of the functional images, most hr-fMRI studies forgo the practice of spatial smoothing, or else apply only minimal smoothing within the boundaries of each MTL substructure in an effort to boost the signal-to-noise ratio (SNR) for voxel-level analyses.

Two of the earliest hr-fMRI studies of the MTL directly examined hippocampal encoding and retrieval, and yielded striking findings suggesting that a functional distinction may exist between input and output structures of the tri-synaptic hippocampal circuit, with input structures (DG/CA_2/3) being selectively active during successful encoding, and output structures (CA₁ and subiculum) being differentially active during successful subsequent retrieval (Eldridge et al., 2005; Zeineh et al., 2003). In the first of these studies, Zeineh et al. (2003) scanned participants as they performed a face-name association task across a series of alternating encoding and recall blocks. During associative encoding, activity levels in DG/CA_2/3, but not in subiculum, closely tracked participants' learning curve, whereas during face-cued recall for the associated name, activity in subiculum, but not in DG/CA_2/3, tracked performance (Figure 2a). In the second of these studies, Eldridge et al. (2005) used an event-related paradigm to link activation changes to memory success on an item-by-item basis. In this study, participants initially were scanned as they intentionally encoded a series of object pairs, and again the next day as their memories were tested. Through a subsequent memory analysis (for reviews, see Blumenfeld and Ranganath, 2007; Paller and Wagner, 2002; Spaniol et al., 2009; Uncapher and Wagner, 2009), retrieval performance was used to back-sort encoding trials according to whether each item was associated with vivid recollection, familiarity, or forgetting. The activation profile of DG/CA_2/3 suggested selective involvement in the formation of new memories: encoding activity was correlated with subsequent memory success, and retrieval activity was greater for correctly rejected foils (i.e., novel items) than for successfully recognized pairs. By contrast, CA₁ encoding activity was not clearly related to subsequent memory performance, whereas retrieval-related activation was greater for recollected than forgotten items. In subiculum, encoding activity demonstrated a complex relationship to later memory outcome, being greater for later familiar than later forgotten or recollected items; at retrieval, subicular activity clearly tracked recollection success, being greater for recollected than for familiar or forgotten items.

Three recent hr-fMRI studies have revealed related intra-hippocampal functional dissociations during episodic encoding and retrieval. In an experiment examining the degree to which MTL subfields are recruited during the study, delay, and probe phases of a delayed-match-to-sample (DMS) face memory task, Olsen et al. (2009) observed a subregion × task phase interaction, wherein DG/CA_2/3 was differentially active during the study and delay phases of the task, during which encoding operations predominate, whereas CA₁ and subiculum were differentially active during the probe (retrieval) phase. Furthermore, in a pair of experiments examining the encoding (Carr et al., in press) and retrieval (Viskontas et al., 2009) of durable episodic memories—that is, memories retaining their episodic quality over a one-week delay—a significant subsequent memory effect was selectively observed in DG/CA_2/3 at encoding, whereas during retrieval, recollection was selectively associated with activation in the subiculum.

Zeineh et al. (2003) argued that the differential encoding/retrieval responsivity of DG/CA_2/3 and subiculum, respectively, may challenge leading computational models of hippocampal function, which vest encoding and retrieval mechanisms within the same hippocampal subfields (O'Reilly and Rudy, 2001; Rolls, 1996). Other hr-fMRI data, however, are more broadly consistent with extant models, as encoding effects have been observed not just in DG/CA_2/3, but in all subfields of the hippocampal circuit. For example, Suthana et al. (2009) found significant performance-related CA₁ activity as participants learned to navigate a spatial environment independent of starting point and direction. Computational theories posit that during learning, CA₁ mediates the binding of CA₃ patterns with ERC representations, allowing CA₁ to serve as a “translator” during subsequent retrieval processes due to back-projections to ERC (e.g., O'Reilly and McClelland, 1994; Rolls and Kesner, 2006). Moreover, examining incidental encoding of faces and scenes, Preston et al. (2010) observed a significant novelty encoding effect (novel > repeated stimuli) in the subiculum, and further showed that subicular encoding activity correlated with subsequent memory performance (Figure 2b). These findings complement related hr-fMRI observations of novelty encoding effects in human subiculum (Bakker et al., 2008; Zeineh et al., 2000), and argue against a restricted role of the subiculum in retrieval. Based on animal data, O'Mara (2006) proposed that the subiculum may be important in forming object-space associations by combining direct input from MTL cortical regions (including PRC) with input from the CA fields.

Consistent with these models, recent electrophysiological data from rodent hippocampus suggest that DG plays the clearest role in pattern separation, whereas both pattern separation and completion are performed to varying degrees by CA₃ and CA₁ (Guzowski et al., 2004; Lee et al., 2004; Leutgeb et al., 2007; Leutgeb et al., 2004; Vazdarjanova and Guzowski, 2004). Specifically, in response to a parametric manipulation of pattern overlap, DG has been shown to exhibit greater differentiation when presented high pattern overlap than does CA₃ (Leutgeb et al., 2007). Studies examining differences between CA₃ and CA₁ further suggest that CA₃ responds in a non-linear manner to pattern overlap, with low overlap triggering establishment of a novel hippocampal pattern (i.e., pattern separation) and higher overlap triggering establishment of a previously encoded pattern (i.e., pattern completion), whereas CA₁ responds more linearly (Guzowski et al., 2004; Lee et al., 2004; Leutgeb et al., 2004; Vazdarjanova and Guzowski, 2004). The conflicting tendency of CA₃ towards pattern separation or pattern completion, depending on the degree of overlap between past and present experience, further highlights the difficulty of classifying hippocampal subfield function according to singular constructs, such as encoding and retrieval.

Using hr-fMRI to determine whether differential biases toward pattern separation and completion are seen in human hippocampal subfields, Bakker et al. (2008) presented participants with a stream of visual objects that included objects presented for the first time (novel), identical repeats of previously presented objects (repeats), and novel objects that were similar to (e.g., were other exemplars of) previously presented objects (lures) (Figure 3a). Of interest was the degree to which each subfield displayed repetition suppression—that is, decreased activity relative to novel objects—for repeats and lures. Results revealed that DG/CA_2/3 activation was greater during both novel and lure trials relative to repeat trials, whereas activation in CA₁, and to a lesser extent subiculum, was greater during novel trials relative to both repeat and lure trials. Bakker et al. interpreted these findings to mean that DG/CA_2/3 is the principal hippocampal subfield biased toward pattern separation (because lures were putatively processed like novel events), whereas CA₁ and subiculum were biased toward pattern completion (lures were putatively processed like repeated events).

Bakker et al.'s data are the first to suggest functional heterogeneity within human hippocampus with respect to pattern separation and completion (for a potentially related hr-fMRI finding, see Johnson et al., 2008). Nevertheless, to better elucidate the relationship between input pattern overlap and hippocampal subfield responses, future studies will require a more systematic and quantitative manipulation of stimulus similarity. Such studies could then evaluate hypotheses regarding non-linear (DG/CA_2/3) and linear (CA₁) subfield responses to pattern overlap, thus affording more precise comparisons to the rodent literature. While intriguing, it is also important to note that Bakker et al.'s (2008) results diverge from those described in a related hr-fMRI investigation (Kirwan and Stark, 2007), where far less conclusive findings of hippocampal subfield functional heterogeneity were reported despite using the same stimuli and similar stimulus conditions as in the Bakker et al. study. A potentially critical difference between these two experiments was the participants' retrieval orientation, with Kirwan and Stark requiring participants to make explicit judgments about the novel/repeat/lure status of stimuli, whereas mnemonic processing was incidental to the task in Bakker et al. Accordingly, task demands may influence the manner in which stimuli are processed by hippocampal subfields. Indeed, other recent data similarly suggest that automatic processes carried out by the hippocampus may be difficult to evaluate if participants adopt mnemonic strategies that alter hippocampal processing via top-down influences (Kumaran and Maguire, 2009; see also, Dudukovic and Wagner, 2007; Duncan et al., 2009). Because the interplay between goal states and hippocampal computation is currently poorly understood, future studies are needed to determine how choice of task influences hippocampal activity, the results from which will likely inform subsequent efforts to evaluate hippocampal subfield involvement in pattern separation and pattern completion.

More broadly, Bakker et al.'s investigation of pattern separation and completion processes in human hippocampal subfields is illustrative of how hr-fMRI can serve as a valuable tool in enabling comparisons between human and animal findings. Two recent hr-fMRI studies have similarly evaluated specific predictions about MTL subregion function advanced in the animal literature, with the first examining hippocampal subfield involvement in spatial navigation (Suthana et al., 2009). In this study, the authors used hr-fMRI to evaluate MTL subfield activity as participants learned to virtually navigate a small town starting from either multiple start-points (allocentric navigation) or the same start-point (egocentric navigation). Results revealed greater activity in CA₁ for multiple than single start-point navigation (Figure 1b), as well as a correlation between CA₁ activity and behavioral performance in the multiple start-point condition. Such findings support data from rodents suggesting a key role for the hippocampus in spatial navigation (for a review, see, e.g., Bird and Burgess, 2008), and for CA₁ in particular in processing spatial relationships among landmarks (Goodrich-Hunsaker et al., 2008) and integrating novel spatial information with previously-learned information (Vinogradova, 2001).

An exciting new direction in understanding hippocampal and MTL cortical contributions to memory is the application of multi-voxel pattern analyses (MVPA) to hr-fMRI data. Whereas univariate analyses focus on individual voxel activity in isolation or averaged activity across all voxels in an ROI, MVPA capitalizes on the rich information represented in spatially distributed activity patterns (Haxby et al., 2001; Haynes and Rees, 2006; Norman et al., 2006). Thus, the MVPA approach may be more sensitive than standard univariate approaches because MVPA can detect condition-specific changes in activation topography within a region, even when the mean activity in the region is indistinguishable across conditions (e.g., Harrison and Tong, 2009; Serences et al., 2009). However, if one were to perform MVPA on standard-resolution fMRI data from the MTL, classification performance might be hindered by the small number of voxels within each subregion; furthermore, the granularity of standard-resolution data may be too coarse to robustly detect information coding in the MTL. High-resolution MTL scans, capable of capturing more fine-grained activation patterns, may provide richer datasets for MVPA.

Two recent hr-fMRI studies have used an MVPA approach to shed light on categorical selectivity within the MTL cortex (Diana et al., 2008) and on the role of the hippocampus and surrounding cortices in spatial navigation (Hassabis et al., 2009). To characterize representational selectivity in MTL cortical regions, and more specifically to assess whether PHC exhibits selectivity for spatial information, Diana et al. (2008) examined MTL activity patterns as participants viewed blocks of stimuli from five distinct categories (scenes, faces, objects, toys, and abstract shapes). Strikingly, MVPA revealed that patterns of activity within PHC contain sufficient information to decode which of the five categories a given participant was currently viewing, including above-chance classification between the non-scene categories (Figure 4). When these data were submitted to univariate analyses, however, only those contrasts involving scenes reached significance in PHC (see Preston et al., 2010, for related univariate hr-fMRI data revealing differential PHC responsiveness to scene vs. face novelty and subsequent memory). In contrast to PHC, category information could not be reliably extracted from distributed activity patterns within PRC or the hippocampus, nor did univariate analyses reveal significant effects in these regions (see Preston et al., 2010, for hr-fMRI data demonstrating face and scene novelty and subsequent memory effects in PRC). Diana et al.'s findings are generally consistent with theories of MTL function positing that PHC can represent both spatial and non-spatial information (Diana et al., 2007; see also, Buffalo et al., 2006; Litman et al., 2009). While the behavioral significance of these distributed patterns of PHC activation have yet to be documented, Diana et al.'s (2008) data suggest that MVPA may be a fruitful method for characterizing MTL regional selectivity profiles.

In addition to its putative role in binding item and contextual details into an integrated conjunctive memory, the hippocampus is also posited to play a key role in spatial navigation (e.g., Bird and Burgess, 2008; Burgess et al., 2002). Motivated by findings of hippocampal place cells in rodents that show selectivity for specific environmental locations, Hassabis et al. (2009) used MVPA techniques to assess whether it is possible to predict an individual's location in a virtual reality environment based on distributed patterns of MTL activity. High-resolution fMRI data were collected as participants navigated two unique rooms, each consisting of four target positions. Rather than evaluate activity patterns in pre-defined MTL subfields, Hassabis et al. used a searchlight approach (Kriegeskorte et al., 2006) to evaluate spherical cliques of voxels within a bounding box surrounding the MTL. Searchlight MVPA revealed activation clusters within the posterior hippocampus that supported above-chance classification of an individual's location within a room, and activation clusters within PHC that supported differentiation between the two rooms. Univariate analyses, on the other hand, failed to reveal activity differences associated with specific locations or rooms. Hassabis et al. suggest that the ability of the hippocampus to discriminate individual locations within a room may form the basis of an allocentric cognitive map representing the room's layout (in agreement with Suthana et al., 2009), whereas PHC may extract contextual information from each room. The authors posit that their ability to predict a participant's location within a room using MVPA techniques may reflect hippocampal neuronal ensembles with predictable topographical functional organization—if neural codes were random and uniformly distributed across hippocampal voxels, classification of location would have been impossible. (It bears noting that hr-fMRI data submitted to flat-mapping analysis may not be well-suited for searchlight MVPA because local patterns of activity among neighboring voxels may be disrupted as functional data are projected from 3D to 2D).

Unfortunately, the physiological factors that give rise to signal dropout also result in signal displacement in the phase-encoding direction (Ojemann et al., 1997; Olman et al., 2009)—in the case of most high-resolution studies acquiring coronal images perpendicular to the long axis of the hippocampus, in the superior-inferior direction or vice-versa. The magnitude and form of displacements again depend on the particular functional sequence adopted, with EPI protocols being particularly prone to displacement. Although not always implemented in hr-fMRI studies, such displacement can be attenuated by collecting B0 field maps and subsequently correcting for the measured displacement. Several analysis software packages now include standardized tools for unwarping EPI images (e.g., FSL: http://www.fmrib.ox.ac.uk/fsl/fugue/feat_fieldmap.html; SPM: http://www.fil.ion.ucl.ac.uk/spm/software/spm8/), enabling future studies to more easily implement such corrections and benefit from increased precision in functional localization. Finally, signal displacement can be reduced by increasing image acquisition speed or reducing field-of-view in the phase-encoding direction, though at a cost of reduced SNR (Olman et al., 2009).

With respect to ROI analyses, a non-trivial factor in evaluating subfield function is the time required to manually trace multiple ROIs in both hemispheres. More importantly, issues pertaining to observer bias must also be considered, with the same observer ideally tracing all ROIs in a given study to maintain consistency across participants. Recent advances in automated segmentation of MTL subfields (e.g., Van Leemput et al., 2009) may reduce the need for manual demarcation of subfields, and thus offer an associated reduction in both the time requirements and observer biases related to analyzing hr-fMRI datasets.

With respect to testing theories of hippocampal subfield function arising from computational models and animal studies, a major limitation of current hr-fMRI methodology is the difficulty in anatomically discriminating DG from CA_2/3. Despite the gains in resolution afforded by high-resolution imaging, segmentation of DG from CA_2/3 is difficult due to the relative absence of anatomical landmarks, particularly in anterior portions of the hippocampus. Recent advances in structural imaging techniques allow for segmentation of DG from CA_2/3 in a small number of slices (Ekstrom et al., 2009b). However, given that DG comprises a very small number of structural voxels, it is likely that, at the resolution of the functional images (often acquired at approximately three times lower resolution than that of structural data), many functional voxels will contain data from both DG and CA_2/3. To this end, further gains in functional resolution may be required for these regions to be reliably differentiated.

Further increases in functional resolution likely will facilitate other efforts to examine whether human MTL demonstrates functional heterogeneity within its subfields similar to that observed in the animal literature, such as the presence of parallel streams supporting the processing of spatial information (postrhinal, medial ERC, proximal CA1, distal subiculum) and non-spatial information (PRC, lateral ERC, distal CA1, proximal subiculum) (for a review, see Knierim et al., 2006). Given the rapid advances in parallel imaging technology, as well as the increasing availability of 7T scanners, there is reason to be optimistic that hr-fMRI at 3T and higher fields will ultimately achieve the necessary enhanced functional resolution required to more fully integrate the human and animal literatures on MTL subfield function. It bears noting, however, that with gains in sensitivity afforded by higher field strength come increased dropout and distortion in anterior regions of the MTL (e.g., Krasnow et al., 2003; Kruger et al., 2001). Moreover, the effective resolution of hr-fMRI data ultimately will be constrained by the point-spread function of the BOLD response (Logothetis and Wandell, 2004).

A further challenge for relating hr-fMRI MTL subfield findings and recording data in animals is that, unlike direct single or multi-unit recordings from the animal brain, the BOLD signal is thought to primarily reflect input and local neuronal processing from many thousands of neurons, rather than spiking/output activity (Angenstein et al., 2009; Logothetis et al., 2001; Logothetis and Wandell, 2004). Thus, hr-fMRI of the MTL may be too coarse to detect certain sparse coding properties of MTL neurons (e.g., grid cell responses in medial ERC), being better suited to comparisons with animal studies examining local field potentials (LFPs) or other measures of population activity. Furthermore, because fMRI is an indirect measure of neuronal function in which the BOLD signal peaks several seconds after neural response onset, the temporal resolution afforded is modest. As a result, hypotheses requiring knowledge of millisecond timing differences between regions, such as those regarding the order in which information flows through cortical and hippocampal circuitry, cannot be feasibly evaluated using fMRI alone. It is partly for this reason that a number of laboratories have begun to combine the high spatial resolution of structural and functional imaging with the high spatial and temporal resolution of direct intracranial recordings in human patients undergoing pre-surgical mapping for pharmacologically resistant temporal lobe epilepsy. While intracranial electroencephalography in patients is not widely available and the placement of electrode contacts is solely based on clinical necessity, extant data indicate that patients typically have multiple contacts throughout the MTL that can be effectively localized to specific subfields using high-resolution imaging (Ekstrom et al., 2008). In addition to addressing MTL function, efforts to integrate hr-fMRI, single-unit, and LFP data from patients can advance understanding of the relationship between MTL BOLD activity and direct neural responses (Ekstrom et al., 2009a).

Finally, a trade-off exists in collecting high-resolution images of the MTL in that only a limited number of high-resolution slices can be prescribed without exceeding a reasonable per-volume acquisition time (e.g., repetition times of 2–4 s). The tradeoff between volume coverage, spatial resolution, and temporal resolution typically results in slice prescriptions that exclude many brain areas outside the MTL and its immediate surroundings. This restricted coverage prevents progress in understanding the functional connectivity between the MTL and regions outside the slice prescription. Replication of hr-fMRI experiments at standard resolution offer one means of assessing how extra-MTL regions, such as structures in prefrontal and parietal cortex, interact with the MTL.