a high-resolution computational atlas of the human hippocampus...

NeuroImage 44 (2009) 385–398

Contents lists available at ScienceDirect

NeuroImage

j ourna l homepage: www.e lsev ie r.com/ locate /yn img

A high-resolution computational atlas of the human hippocampus from postmortemmagnetic resonance imaging at 9.4 T

Paul A. Yushkevich a,⁎, Brian B. Avants a, John Pluta b, Sandhitsu Das a, David Minkoff b,Dawn Mechanic-Hamilton b, Simon Glynn e, Stephen Pickup c, Weixia Liu c, James C. Gee a,Murray Grossman d, John A. Detre b

a Penn Image Computing and Science Laboratory (PICSL), Department of Radiology, University of Pennsylvania, 3600 Market St., Ste 370, Philadelphia, PA 19104, USAb Center for Functional Neuroimaging, Departments of Radiology and Neurology, University of Pennsylvania, Philadelphia, PA, USAc Department of Radiology, University of Pennsylvania, Philadelphia, PA, USAd Department of Neurology, University of Pennsylvania, Philadelphia, PA, USAe Department of Neurology, University of Michigan, Ann Arbor, MI, USA

⁎ Corresponding author.E-mail address: [email protected] (P.A. Y

1053-8119/$ – see front matter © 2008 Elsevier Inc. Alldoi:10.1016/j.neuroimage.2008.08.042

a b s t r a c t
a r t i c l e i n f o
Article history:
This paper describes the co Received 8 April 2008Revised 12 August 2008Accepted 31 August 2008Available online 18 September 2008
nstruction of a computational anatomical atlas of the human hippocampus. Theatlas is derived from high-resolution 9.4 Tesla MRI of postmortem samples. The main subfields of thehippocampus (cornu ammonis fields CA1, CA2/3; the dentate gyrus; and the vestigial hippocampal sulcus)are labeled in the images manually using a combination of distinguishable image features and geometricalfeatures. A synthetic average image is derived from the MRI of the samples using shape and intensityaveraging in the diffeomorphic non-linear registration framework, and a consensus labeling of the templateis generated. The agreement of the consensus labeling with manual labeling of each sample is measured, andthe effect of aiding registration with landmarks and manually generated mask images is evaluated. The atlasis provided as an online resource with the aim of supporting subfield segmentation in emerging hippo-campus imaging and image analysis techniques. An example application examining subfield-levelhippocampal atrophy in temporal lobe epilepsy demonstrates the application of the atlas to in vivo studies.

© 2008 Elsevier Inc. All rights reserved.

Introduction

The hippocampus is a structure of acute interest in neuroimaging.It is primarily associated with encoding of episodic memory, but isalso believed to play an important role in both encoding and retrievalof other types of long term memory (Squire et al., 2004; Moscovitchet al., 2006). Hippocampal neuropathology is of vital interest in thestudy of dementia, epilepsy, schizophrenia and other neurologicaland psychiatric disorders. However, the complex anatomy of thehippocampus poses challenges to image-based computational mor-phometry techniques. The hippocampus is formed by two interlock-ing folded layers of neurons, the cornu ammonis (CA) and the dentategyrus (DG), which are further subdivided into subfields and arebelieved to subserve different functional roles. Non-uniform neuronloss across the hippocampal subfields has been reported inneurodegenerative disorders (West et al., 2004; Bobinski et al.,1997; Braak and Braak, 1991), as has a non-uniform rate ofneuroplasticity (Eriksson et al., 1998; Kuhn et al., 1996; van Praag etal., 2002). However, it is very difficult to distinguish the boundariesbetween hippocampal layers in clinical MRI modalities, since the

ushkevich).

rights reserved.

voxel resolution of ≈1 mm3 (isotropic) is larger than the thickness ofthe DG. The layered structure of the hippocampus can be appreciatedin specialized T2-weighted sequences with a limited field of view andhighly anisotropic voxels (Zeineh et al., 2003), although partialvolume effects are significant in the anterior and posterior of thehippocampus, where the main axis of the structure curves withrespect to the imaging plane. Emerging MRI technologies, includingmulti-coil designs and 7 Tesla human scanners, yield in vivo imageswith significantly improved resolution, and subfield differentiation isquickly becoming a reality (Van Leemput et al., 2008; Cho et al., 2008;Li et al., 2008; Bernasconi et al., 2008).

Traditionally, morphometry studies have represented the hippo-campus as a homogeneous gray matter structure with a blob-likeshape (Carmichael et al., 2005; Fischl et al., 2002; Hsu et al., 2002; Jacket al., 1995; Kelemen et al., 1999; Pitiot et al., 2004; Shen et al., 2002;Styner et al., 2004; Shenton et al., 2002). However, some of the morerecent studies have focused on detecting disease-related structuraland functional differences in the hippocampus at the subfield level(Zeineh et al., 2003;Wang et al., 2006; Apostolova et al., 2006; Kirwanet al., 2007; Mueller et al., 2007). In order to differentiate betweensubfields, investigators either label the subfields for each of thesubjects in a given study or, in the studies that employ large defor-mation diffeomorphic metric mapping (LDDMM) computational

mailto:[email protected]

http://dx.doi.org/10.1016/j.neuroimage.2008.08.042

http://www.sciencedirect.com/science/journal/10538119

1 As an alternative, placing the sample in Fomblin (Ausimount, Thorofare, NJ, USA)liquid was attempted; however because of the complex structure of the hippocampus,we were unable to prevent air bubbles from becoming trapped inside the sample,which would cause severe imaging artifacts.

2 There is some debate as to whether this innermost region of the hippocampusshould be considered a part of the dentate gyrus, as advocated by Amaral (1978), or apart of the cornu ammonis, as in Duvernoy (2005).

Table 1Acquisition parameters for the high-resolution postmortem scans of the hippocampus

ID TR/TE (ms) NT NS MAT FOV (mm) THK (mm) Voxel size (mm) SNR Time (h:min)

1R 4000/26 34 120 350×256 70×77 0.20 0.30×0.20×0.20 70 13:132L 5000/26 36 130 280×300 56×90 0.20 0.30×0.20×0.20 72 14:002R 5000/26 40 130 280×300 56×90 0.20 0.30×0.20×0.20 76 15:333L 5000/26 225 130 200×400 40×80 0.20 0.20×0.20×0.20 120 62:303R 5000/26 45 130 240×300 48×90 0.20 0.30×0.20×0.20 80 15:00

Column abbreviations are as follows; ID: scan identifier, TR/TE: relaxation and echo times, respectively, NT: number of averages, NS: number of slices, MAT: matrix size, FOV: field ofview, THK: slice thickness, SNR: relative signal-to-noise ratio, where the value of 1000 corresponds to a single-average imagewith 1mm isotropic voxels, Time: total acquisition time.

386 P.A. Yushkevich et al. / NeuroImage 44 (2009) 385–398

anatomy techniques (Miller et al., 2006), label the template to whichsubjects' images are normalized (Csernansky et al., 2005; Wang et al.,2006). In either case, the knowledge of where the subfields lie inimage space is based on low-resolution in vivo imagery and the use ofreferences, such as the Duvernoy (2005) atlas. However, emergingimaging modalities are making it possible to delineate hippocampalsubfields in vivo with greater reliability. Van Leemput et al. (2008)leverage 3 Tesla multi-coil imaging at 0.4×0.4×0.8 mm resolution toautomatically label major hippocampal subfields with excellentaccuracy. As 7 Tesla imaging becomes more common, even greateraccuracy in subfield resolution will be possible.

The present paper is premised on the belief that a detailed atlas ofthe hippocampus based on postmortem high-resolution MRI wouldprove highly complementary to morphological studies that leveragecomputational anatomy methods. Postmortem imaging allows thehippocampus to be imaged in laboratory conditions, with greater fieldstrength, smaller field of view and longer acquisition time, yieldinghigh-contrast images with voxels about 100 times smaller in volumethan in typical in vivo imaging. Indeed, Wieshmann et al. (1999) pointout that a twenty-fold increase in resolution over current in vivomodalities is needed to distinguish the layers of the hippocampusreliably. Our approach involves MRI of brain samples containing theintact whole hippocampus at 9.4 T, using overnight scans to achievegood contrast at high resolution.While others have imaged portions ofthe hippocampus at even higher resolution (Wieshmann et al., 1999;Fatterpekar et al., 2002; Chakeres et al., 2005), to the best of ourknowledge, this is the first work to collect images of the entirestructure with isotropic or nearly isotropic voxels 8–12 nl in volume.We manually label hippocampal subfields relying on a combination ofdiscernable intensity and shape features. To combine data fromdifferent samples, we apply an implementation of the unbiased imageaveraging algorithm by Avants and Gee (2004), which is based ondiffeomorphic Lagrangian frame image registration. Registrationaccuracy is evaluated in terms of subfield label alignment in atlasspace.

In addition to describing the postmortem atlas, this paperdemonstrates a possible approach for leveraging the atlas in in vivohippocampal morphometry studies. The atlas is used in combinationwith conventional geometric and image-based computational anat-omy techniques to assess hippocampal atrophy in temporal lobeepilepsy (TLE). Atrophy is evaluated by estimating the asymmetrybetween the hippocampus contralateral to the side of seizure in TLEand the ipsilateral hippocampus. The postmortem hippocampus atlasallows us to estimate atrophy on a subfield-specific basis, and ourexperimental results suggest that atrophy across subfields in non-uniform.

Methods and materials

Specimens and imaging

Formalin-fixed brain specimens (≥21 days) from three autopsycases with no abnormal neuropathological findings were studied.Hemisphereswere separated from the cerebellum and brain stem, and

samples containing the intact hippocampus and not larger than70 mm in diameter were extracted from each hemisphere by makingtwo incisions: the first, orthogonal to the midsagittal plane andparallel to the main axis of the hippocampus, passing through thecorpus callosum; the second, parallel to the midsagittal plane,removing the lateral-most third of the sample.

Images of five hippocampus samples (3 right, 2 left) wereacquired on a 9.4 Tesla Varian 31 cm horizontal bore scanner (VarianInc, Palo Alto, CA) using a 70 mm ID TEM transmit/receive volumecoil (Insight Neuroimaging Systems, Worchester, MA). Samples wereplaced in leak-proof bags and wrapped with plastic to fit snuglyinside the coil.1 Scanning parameters varied slightly across thesamples and are listed in Table 1. All acquisitions used the standardmulti-slice spin echo sequence with TR between 4 s and 5 s, and TE26 ms. An oblique slice plane was chosen to cover the hippocampuswith as few slices as possible, requiring around 130 slices with0.2 mm slice thickness for most images. The phase encode directionwas from left to right and the readout direction followed the longaxis of the hippocampus. Four of the samples were scanned at0.2 mm×0.3 mm×0.2 mm resolution using 34–45 averages toachieve good contrast; these scans were acquired overnight. Onesample was scanned at 0.2 mm×0.2 mm×0.2 mm isotropicresolution with 225 averages over 63 h.

Manual segmentation

Reconstructed MRI images of each of the samples were segmentedby one of the authors (JP) manually. The laminar structure of thehippocampus is clearly observable in these images (Fig. 1). With thehelp of the Duvernoy (2005) atlas and the labeling of the hippocam-pus in MRI slices by Fatterpekar et al. (2002), we were able toconsistently identify four layers in the hippocampus: the high-intensity layer in the CA formed by the stratum oriens and thepyramidal cell layer; the low-intensity layer in the CA formed by thestratum radiatum, stratum lacunosum-moleculare and the vestigialhippocampal sulcus; the stratummoleculare of the dentate gyrus, andthe dentate hilus (referred to as CA4 by some).2 The layers of thedentate gyrus are the most difficult to tell apart, while the contrastbetween bright layers of the CA, the dark layers of the CA andsurrounding structures is very well pronounced. The boundariesbetween the different subfields in the pyramidal cell layer of the cornuammonis (CA1, CA2 and CA3), as well as between CA and thesubiculum, are not visible, and separation have to be made based onposition and shape, rather than intensity. Perhaps the most difficultareas to segment were the transitions between subfields towards thehead and the tail of the hippocampus.

Fig. 1. Coronal and sagittal slices through each of the hippocampus images included in the atlas. The acquisition parameters for these images are listed in Table 1. Image 3L has0.2 mm3 isotropic resolution; other images have resolution 0.3×0.2×0.3 mm3.

387P.A. Yushkevich et al. / NeuroImage 44 (2009) 385–398

Based on our ability to consistently identify substructures in thehippocampus, we selected five labels for the segmentation: CA1; alabel that combines CA2 with CA3; the dentate hilus (i.e., CA4);stratum moleculare of the dentate gyrus; and the label combiningstratum radiatum, stratum lacunosum-moleculare and the vestigialhippocampal sulcus (i.e., the prominent ‘dark band’ in the MRIimages). A schematic of the labels is shown in Fig. 2.

Labelingwas performed using ITK-SNAP (Yushkevich et al., 2006b),an open-source image segmentation tool that allows simultaneoustracing in three orthogonal view planes and provides feedback viathree-dimensional rendering of the segmentation. Most of the timewas spent tracing the subfields in the coronal plane, with verificationin the sagittal plane. In most of our images the hippocampus occupiesover 200 coronal slices, making segmentation extremely labor-intensive, requiring about 40 h of tracing and editing per image.

Reliability analysis

Given the difficulty of segmenting the hippocampus at highresolution, it was deemed impractical to perform reliability analysis,as is commonly done, by having multiple raters make multiple

segmentation attempts. Instead, a limited reliability study was per-formed, in which a subset of five coronal slices was selected in eachhippocampus image for repeated labeling. Two slices were selected inthe head of the hippocampus, two in the midbody of the structure,and one in the tail. Rater JP segmented each of the slices a month afterthe completion of the five full hippocampus segmentations. Rater DMsegmented the same slices independently after having been trained byJP. Dice overlap between segmentation attempts was used to assessintra-rater and inter-rater reliability.

Atlas generation

Combining image data frommultiple samples into a common atlasallows us to generate a model of “average” hippocampal anatomy, toboost signal-to-noise ratio, and to study variations in hippocampalstructure across subjects. In building an atlas, we follow the work ofGuimond and others (Guimond et al., 2000; Davis et al., 2004; Avantsand Gee, 2004) by searching for a synthetic image that can benonlinearly registered to each of the input images with minimal totaldeformation. Our normalization algorithm belongs to the family oflarge deformation diffeomorphic registration techniques (Christensen

Fig. 2. A diagram of the partitioning of the hippocampus into labels used in this paper. The abbreviations used in subsequent figures are in boldface. The leaf nodes in the graphs arethe labels that were actually traced in the postmortemMR images. Later in the paper, segmentation repeatability and atlas consistency results are reported for the leaf nodes as wellas for the parent-level structures formed bymerging the labels corresponding to the leaf nodes. Abbreviations: SO = stratum oriens; PCL = pyramidal cell layer; SR = stratum radiatum;SLM = stratum lacunosum-moleculare; VHS = vestigial hippocampal sulcus.


et al., 1997; Joshi and Miller, 2000; Beg et al., 2005; Miller et al., 2006)and is called Symmetric Normalization (SyN) (Avants et al., 2008). Ituses a greedy symmetric approach, where, given an image matchmetric Π, registration of images I and J seeks a pair of diffeomorphicmaps ϕ1, ϕ2 that minimize

E I; J;ϕ1;ϕ 2ð Þ =Π I ϕ 1 x;12

� �� ; J ϕ2 x;

12

� �� +Z 1

2

0‖v1 x; tð Þ‖Ldt

+Z 1

2

0‖v2 x; tð Þ‖Ldt; subj: to

Aϕ i x; tð ÞAt

= vi ϕ i x; tð Þ; tð Þ; i = 1;2: ð1ÞSymmetric diffeomorphic maps from I to J and from J to I are

given, respectively, by ϕ2−1○ϕ1 and ϕ1

−1○ϕ2. By implementing asymmetric approach that searches for two coupled diffeomorphisms,SyN simultaneously generates forward and inverse maps between Iand J, while the greedy implementation allows for reduced memoryuse and relatively fast convergence. SyN implements a multi-resolution scheme and our experiments used three levels of resolu-tion, where images were smoothed and subsampled by factors of 4, 2and 1 in each dimension. The step size is determined by the Courant–Friedrichs–Lewy (CFL) condition which restricts changes to thediffeomorphism to be sub-voxel, in this case, half a voxel in size. Toreduce the number of parameters that influence the results andrequire tuning, optimization at each resolution level continues untilthere is no reduction in the metric value over several gradient descentiterations. In this way, we find a “best fit” diffeomorphism betweenthe images, given the similarity metric. This philosophy is the naturaldiffeomorphic extension of basic rigid registration. For speed, linearinterpolation of all data is used. Thus, the only parameters thatinfluence outcome are choice of similarity terms, time-step (restrictedunder the CFL condition) and multi-resolution strategy.

Registration of postmortem MRI is challenging because imageshave different fields of view and because there are many unique localintensity features in each image (vessels, hyperintensities, imagingartifacts) that make finding one-to-one correspondences difficult. Toboost registration accuracy, we introduce two kinds of manuallygenerated aids: landmarks and rough binary masks of the hippocam-pus. Landmarks are placed at the two ends of each digitation in thehead of the hippocampus (digitations are folds that can be seen in theright column of Fig.1), as well as at the anterior-most point of the headand posterior-most point of the tail. Landmarks are incorporated intoSyN by adding an additional term to the optimization criterion, asdescribed in Avants et al. (2008). SyN thus optimizes the prior

(landmark) correspondence term simultaneouslywith the appearanceterm, leading to a different solution than an unguided normalization,in particular in the vicinity of the landmarks. Additional details of thisapproach are given in the Appendix.

Binary masks are generated in a fraction of the time required forthe subfield segmentation because they are traced in the axial plane,where the hippocampus occupies the fewest slices, and because theyare drawn roughly andmade to include a layer of several pixels aroundthe hippocampus. Both the landmarks and the masks require about anhour each to generate for every sample. When building an atlas from alarger postmortem dataset, it would be much more cost-effective togenerate these registration aids than full manual segmentations. Toincorporate masks into the registration, we apply a dilation operationto the masks with the radius of a few (5) voxels, and multiply theimage by the mask. This causes the registration algorithm to line upmask boundaries, while also lining up the image intensity edgesassociated with the boundary between the hippocampus andsurrounding tissues which remain slightly inside the mask. Also, inthe presence of a mask, for all similarity metrics, the gradient value ismultiplied by the mask value. Thus, information outside of the mask(represented by value zero) has no direct influence on the registrationsolution.

Initial linear alignment of hippocampus images based on imageintensity alone proved difficult in postmortem data. Instead, affinealignment was performed by landmark matching, followed bymatching of binary masks. Subsequent atlas building involved thefollowing algorithm (Avants and Gee, 2004): (1) use SyN to registerevery image to a template; (2) average the intensity of all registeredimages; (3) average the initial velocity fields Yv x;0ð Þ of the maps fromthe template to each of the images and generate a shape distanceminimizing diffeomorphic update to the template shape; this step isknown as shape averaging (Avants and Gee, 2004); (4) repeat step 1using the new shape average as the template. To make group-wiseregistration unbiased, Avants et al. (Avants and Gee, 2004) suggestusing the intensity average of affine-registered images as the initialtemplate. However, this requires a substantial number of images, sothat the features common to all images are more prominent than thevariation across images. Since in our case the number of samples issmall, we choose one of the samples as the initial template.We use thenormalized cross-correlation image match metric, which is robust tointensity inhomogeneity present in our data.

A consensus segmentation of the atlas is generated by warping themanual segmentations to the atlas and using the STAPLE algorithm

Fig. 3. Coronal, sagittal and 3D views of hippocampus images and subfield labels for samples 1R and 2L, following affine alignment, and for atlases computed by unsuperviseddeformable registration (DEF) and mask-aided registration (DEF-MSK). Abbreviations: CA — cornu ammonis; DG:H — hilus of the dentate gyrus (a.k.a. CA4); DG:SM — stratummoleculare of the dentate gyrus; SR — stratum radiatum; SLM — stratum lacunosum-moleculare; VHS — vestigial hippocampal sulcus.


(Warfield et al., 2004) to assign a label to each voxel. To examine thequality of alignment, we report the overlap in template space betweenthe STAPLE segmentation and each of the individual segmentations.

Results

Fig. 1 shows coronal and sagittal slices throughMRI images of eachof the five samples used to construct the postmortem atlas. Theseimages have good contrast, and the edges separating the hippocampusfrom adjacent tissue are well pronounced (e.g., unlike in clinical MRI,the amygdala is clearly separated from the hippocampus). A numberof imaging artifacts are also evident: intensity inhomogeneity;hyperintensity and small amounts of distortion at tissue-air bound-aries; ringing artifacts are also present in some images.

Segmentations of two of the samples are shown in two and threedimensions in Fig. 3. In one of the samples, the amygdala and thesubiculum were segmented in addition to hippocampal subfields,shown in Fig. 4.

Overlap-based reliability analysis is given in Fig. 5. The repeat-ability of the segmentation by rater JP was high, while the inter-raterreliability was significantly lower for CA1 and CA2/3 and approxi-mately the same for the other subfields. Lower reliability for CA2/3 isprobably due to the differences in how far into the head of thehippocampus each rater would extend the CA2/3 label: JP did not labelCA2/3 in roughly the anterior two fifths of the hippocampus (as seenin Fig. 3), while DM did label CA2/3 in the anterior hippocampus.

Atlases generated by diffeomorphic averaging with and withoutmasks and landmarks are shown in Fig. 3. Notice that the atlases haveimage characteristics and shape similar to those of the input images.This is particularly noticeable in the 3D renderings of automatically

Fig. 4. Hippocampus segmentation, along with the adjacent amygdala and subiculum (tparahippocampal cortex) Subfield color labels are the same as in Fig. 3.

generated consensus segmentations, where the folds in the anteriorhippocampus are clearly visible.

To simultaneously analyze the agreement between subfield labelsacross segmentations and normalization quality, we measure theoverlap between the consensus segmentation of each atlas with thewarped segmentations of individual samples. We emphasize that thesubfield segmentations are not used during atlas building, althoughfor some of the atlases, we guide registrationwith binary masks of thewhole hippocampus, which are generated in an hour or less, and in theaxial plane. For each subfield, Fig. 6 plots the average Dice overlap(Dice, 1945) for each subfield between the segmentation of thesubfield in each individual, warped to atlas space, and the consensussegmentation of the subfield in the atlas, obtained using the STAPLEalgorithm (Warfield et al., 2004). Based on the overlap, the atlas builtusing deformable registration with masks but without landmarks hasthe best consistency. Average overlap is nearly 90% for CA1, which isthe largest subfield and nearly 80% for the dentate hilus (DG:H), whichis smaller, but also the least hollow of the subfields. When the dentategyrus is considered as a single structure by combining the dentatehilus with the stratum moleculare, the overlap increases to 83%. Forthe thinner layers, the dentate's stratum moleculare and the darkband formed by SR, SLM and VHS, the overlap is below 70%. However,since these layers are very thin, even a small misalignment results inlarge overlap error.

In addition to the overlaps, we plot in Fig. 7 the root mean squaredistance between the boundary surface of each subfield in theconsensus segmentation and the corresponding boundary surface ineach of the individual segmentations warped to atlas space. Thismeasure of boundary displacement error is less sensitive to thethickness of structures than Dice overlap, and indeed the difference in

he subiculum label includes the presubiculum, the parasubiculum and parts of the

Fig. 5. Analysis of segmentation repeatability, given in terms of Dice overlap between the full 3D segmentations used in constructing the atlas and repeated segmentations of aselected set of slices by two raters. Average Dice overlap for each subfield, as well as various combinations of subfields, is reported. See Fig. 2 for a diagram of the subfield labelsappearing on the x-axis.


this error is much less across all subfields than the overlap differencesin Fig. 6. Furthermore, the average boundary displacement error canbe evaluated point-wise on the boundary of each subfield in the atlas,resulting in a scalar map that can be helpful to identifying areas wheresubfield segmentation and normalization are least accurate. Fig. 8shows these maps for each of the subfields in the “DEF-MSK” atlas. Itillustrates that the greatest disagreement in subfield labels is at twolocations: the point where CA1 and the dentate gyrus merge towardsthe tail of the hippocampus, and the similar merging point in the headof the hippocampus. These points are very difficult to labelconsistently due to the lack of intensity features separating CA andthe dentate at the anterior and posterior ends of the dentate. However,in the midbody of the hippocampus, the agreement is remarkablygood, as is the alignment between the folds in the anteriorhippocampus.

Figs. 6 and 7 clearly illustrate the improvement given bydeformable registration over affine alignment, while the benefit ofmasks and landmarks used in deformable registration is notimmediately apparent. To test if the use of landmarks and maskinghas a significant effect on Dice overlap, we use a 4-way ANOVA withfactors ‘masking’, ‘landmarks’, ‘subfield’ and ‘sample’, allowingsubfield-sample andmasking-landmarks interactions. When applying

Fig. 6. Agreement between the consensus atlas labeling and segmentation of individual sabuilding approaches are compared: landmark-based and mask-based affine alignment (AFF-aided by landmarks (DEF-LM), masks (DEF-MSK) and both masks and landmarks (DEF-LM-

ANOVA to Dice overlaps, we first normalize overlaps using the logittransform: logit(α)= log(α)− log(1−α) As stipulated by Zou et al.(2004), this transform maps the overlaps to the entire real line,making Gaussian-based statistics more appropriate. However, in thisparticular experiment, this transform does not affect the significanceof the ANOVA findings. When using all labels in Fig. 6, the ANOVAreports significant positive effect of masking on overlap (F=28.3,pb0.0001) as well as significant negative effect of using landmarks(F=4.25, pb0.04). No significant interaction between landmarks andmasking is detected. Furthermore, when we exclude CA1 and itsparent labels (SO+PCL, H, see Fig. 2), the effect of masking losessignificance (F=2.82, pb0.10) while the negative effect of usinglandmarks remains (F=5.21, pb0.03). An equivalent ANOVA examina-tion of boundary distances yields similar results: significant positiveeffect of masking, (F=6.07, pb0.02) and negative effect of landmarks(F=4.61, pb0.03) when considering all labels in Fig. 6; no significanteffect of masking (F=0.001, pb1.00) and significant negative effect oflandmarks (F=5.45, pb0.02) when looking only at internal subfields.These results suggest that the improvement associated with maskingis greatest along the outer boundary of the hippocampus (as onewould expect, since the mask falls very close to this boundary) whilethe improvement for the internal subfield alignment is minimal.

mples expressed in terms of average Dice overlap computed in atlas space. Six atlas-LM/AFF-MSK), image-based deformable normalization (DEF), deformable normalizationMSK). See Fig. 2 for a diagram of the subfield labels appearing on the x-axis.

Fig. 7. Agreement between the consensus atlas labeling and segmentation of individual samples expressed in terms of average distance between the boundaries of subfields in theatlas and the boundaries of corresponding subfields inwarped individual segmentations. Refer to Fig. 6 for the explanation of different bars in the plot. See Fig. 2 for a diagram of thesubfield labels appearing on the x-axis.


Surprisingly, the use of landmarks had a counterproductive effect,perhaps because most of the mismatch occurs away from the anteriorhippocampus where the majority of the landmarks were placed.

Application to in vivo studies3

This section uses a small dataset from a temporal lobe epilepsy(TLE) imaging study to demonstrate how the postmortem hippocam-pus atlas may be used to analyze “standard” in vivo T1-weighted MRIdatawith 1mm3 resolution. In this experiment, we test the hypothesisthat structural asymmetry between the diseased and healthy hippocampiin unilateral TLE varies across the hippocampal subfields.

Our approach uses standard normalization techniques to estimatevolume change between the subfields in the hippocampus contral-ateral to the seizure focus (the “healthy” hippocampus) and theipsilateral (“diseased”) hippocampus. The approach combines twonormalization schemes: shape-based normalization is used to mapsubfield labels from the postmortem atlas to binary segmentations ofthe whole healthy-side hippocampus in T1 data; and deformableimage registration is used to estimate asymmetry between the healthyand diseased hippocampus regions. Shape-based normalization yieldsan approximate labeling of the subfields in the healthy hippocampusand registration produces a field of Jacobian determinant values. Byintegrating the Jacobian field over the subfield labels, we get anestimate of volume change for each subfield. A drawback of thisapproach is that it relies on manual or semi-automatic segmentationof the hippocampus in in vivo data and is sensitive to errors in thissegmentation, which can be substantial (Carmichael et al., 2005). ThePotential future applications section discusses more advancedpotential approaches for postmortem to in vivo mapping.

Subjects, imaging and semi-automatic segmentation

Our in vivo experiment uses structural MRI data from an ongoingfMRI study of TLE. The aims and design of the study, which wasoriginally implemented at 1.5 T, are described in Rabin et al. (2004). Anupdated protocol that uses 3 Tesla imaging is detailed in Narayan et al.(2005). Our experiment uses a subset of the structural MRI datacollected for the TLE study, according to the following inclusioncriteria: (1) diagnosis of medically refractory TLE with a unilateralseizure focus; (2) successful outcome, in terms of seizure control,

3 This section is new for the current revision. We do not use boldface for clarity.

following epilepsy surgery; (3) absence of other neurological orpsychiatric disorders; (4) imaging at 3 T. Of the 12 patients that metthese criteria, 5 had seizure focus in the right medial temporal lobe (4female/1 male; age=36.6(11.3)) and 7 had a left medial temporal lobeseizure focus (5 female/2 male; age=40.0(15.4)). For each subject, theimaging data used in our experiment consisted of a high resolution(voxel size 0.9375×0.9375×1.0 mm3) T1-weighted structural MRIscan obtained on a 3 Tesla Siemens Trio MRI scanner using an eight-channel head coil.

The hippocampus was segmented bilaterally by one of the co-authors (JP) using a combination of landmark-driven automatic atlas-based segmentation and manual editing in ITK-SNAP. The details ofthis semi-automatic segmentation protocol and reliability estimatesfor the TLE dataset appear in Pluta et al. (2008). The reportedreliability is high, with 85.9% average Dice overlap between manualand semi-automatic segmentation for the hippocampus contralateralto the seizure focus and 83.1% average overlap for the ipsilateralhippocampus.

Shape-based mapping of hippocampal subfields from postmortem atlasto the healthy hippocampus

Shape-based normalization is used to map hippocampal subfieldlabels from the postmortem hippocampus atlas to the segmentation ofthe healthy hippocampus in each image. Shape-based normalizationis appropriate for mapping anatomical data between structures insituations where image intensity does not provide sufficient cues forestablishing a mapping. This is the case for the mapping between thepostmortem atlas, where the boundaries between subfields are clearlyvisible, and the T1 images, where these boundaries are blurred out orinvisible because of insufficient resolution and partial volume effects.The assumption of shape-based mapping is that the position of thesubfields relative to the overall hippocampus shape is roughly thesame in the healthy in vivo hippocampus and the postmortem atlas.We recognize that this is a strong assumption that leads to an inherentlevel of error in subfield boundary estimation; however, some level oferror must be accepted when estimating subfields in clinicalresolution in vivo MRI, since the data itself does not provide sufficientclues for determining subfield boundaries.

Shape-based normalization uses the continuous medial representa-tion (cm-rep) method (Yushkevich et al., 2006c, 2007, 2008). Thismethod represents anatomical structures using deformable models(cm-reps, for short) that describe objects in terms of their geometricskeletons and derive boundary representations of objects as a function

Fig. 8. Surface maps of root mean square distance between each subfield in the atlas and the corresponding subfield segmentation in each of the input imageswarped to the atlas. Theatlas used in this figure was constructed using binary masks and no landmarks (DEF-MSK in Fig. 6). Larger values in the maps (yellow and red) indicate lesser coherence between theatlas and the individual segmentations. At the bottom left, all subfields in the atlas are shown together to provide a visual reference.


of the skeletons. A cm-rep is fitted to the binary segmentation of eachhealthy hippocampus in the study by maximizing the overlapbetween the boundary representation and the binary image. Duringdeformable modeling, cm-reps preserve the branching structure ofthe skeleton, making it possible to fit geometrically congruentskeletons to multiple instances of an anatomical structure, which inturn allows one to leverage geometrical features derived from theskeleton for normalization and shape analysis.

A cm-rep model is defined as follows. The skeleton of a cm-repmodel is a surface patch (i.e., a manifold with boundary home-omorphic to the unit disk in ℝ2) that is specified as a triangle mesh.Each node in this mesh consists of a pointm∈ℝ3 and a positive radiusvalue R. The mesh is constructed by recursive subdivision of a coarsemesh using Loop subdivision rules (Loop and DeRose, 1990). Theboundary of the model consists of two surface patches b+ and b−, oneon each side of the skeleton. These boundary patches are derived fromthe skeleton mesh using the inverse skeletonization formula:

b� =m + R −rmR�ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1−‖rmR‖2

pNm

� �; ð2Þ

where ▿mR denotes the Riemannian (surface) gradient of R on m andNm denotes the unit normal to m. These first-order properties arecomputed using finite difference approximations (Xu, 2004). Con-straints are in place to ensure that the two boundary patches share acommon edge and do not self-intersect, thus forming a boundary of asimple region in ℝ3. These constraints are derived in Yushkevich et al.

(2006c) and an approach towards enforcing them on an triangle meshis given in Yushkevich et al. (2008).

The interior of the cm-rep model is spanned by line segments,called spokes, which extend from points mi on the skeleton to thecorresponding points bi

+ and bi− on the model's boundary and which

are orthogonal to the boundary. As argued in Yushkevich et al. (2006c)and illustrated in Fig. 9, the spokes form a shape-based coordinatesystem on the interior of the cm-rep model. In this coordinate system,each point x on the interior of a cm-rep model is assigned a tuple (u, v,ξ), where (u, v) describes the position of the “tail” of the spoke passingthrough x on the medial surface m, and ξ describes the relativeposition of the point on the spoke between the medial surface and theboundary. The values (u, v) at x depend on the parameterization of mand can be changed using reparameterization. During cm-rep modelfitting, the parameterization is constrained to minimize the distortionin the area element of m(u, v) across the medial surface. Thisconstraint, analogous to uniform arc length parameterization of acurve, ensures that (u, v) coordinates establish a degree of geometriccorrespondence between different cm-rep models.

Shape-based normalization between the postmortem hippocam-pus atlas and the hippocampi in in vivo data involves fitting a cm-repmodel to the binary segmentation of the hippocampus in each of theimages. Then the normalization is simply the set of diffeomorphicmaps that take every point with coordinates (u, v, ξ) in the cm-repmodel fitted to the postmortem atlas to the points with the samecoordinates in models fitted to the in vivo hippocampi. Using these

Fig. 9. A conceptual 2D illustration of shape-based normalization via the cm-rep coordinate system. The central curve is the skeleton m; the radial lines from the skeleton to theboundary are the spokes; the shape-based mapping between the two models is given by the locations of corresponding grid vertices.


maps, subfield labels can be assigned to each of the in vivohippocampi.

Image-based estimation of healthy-diseased asymmetry

Atrophy associated with temporal lobe epilepsy is defined as thelocalized rate of change in volume estimated using image registration.The rationale for using image registration is that some featuresrevealing subfield boundaries may be present in the in vivo data.While it may be difficult to match these features to the postmortematlas, which has a different modality, resolution and field of view,matching these features between the left and right hippocampi in thesame subject is a less difficult registration problem.

Registration is performed between the healthy and diseasedhippocampi for each subject. To initialize the registration, we usecm-rep models. We fit a cm-rep model to both hippocampi, thus

Fig. 10. Illustration of cm-rep model fitting used for shape-based normalization. The top rowFig. 3), and the other rows showmodels fitted to segmentations of the “healthy” hippocampucolumn shows the target structure, i.e., the segmentation of the hippocampus. The secondplotting the radius function R on the skeleton. The third column illustrates the spoke field; spboundary and “span” the interior of the model (see Fig. 9 for the 2D illustration of spokes). Spshown in red. Correspondence on the basis of spoke fields is used to map subfield labels fromb=b+∪b−, which closely approximates the surface of the target hippocampus.

establishing a shape-based point correspondence. We then flip thediseased hippocampus model across the MR image left-right axis anduse Procrustes analysis to find a six-parameter rigid body transforma-tion of the flipped model that minimizes the sum of squared distancesbetween points in the flipped model and the corresponding points inthe model fitted to the healthy hippocampus. This transformationgives a very good initial mapping from the neighborhood of thediseased hippocampus to the neighborhood of the healthy hippo-campus (see the left two columns in Fig. 11). The SyN method,described above in the Atlas generation section, is then used tocompute a deformable registration between the two hippocampi. SyNgenerates a deformation vector field in the space of the healthyhippocampus. This field maps each voxel in the neighborhood of thehealthy hippocampus to a point in the neighborhood of the rigidlyaligned flipped diseased hippocampus. The Jacobian of this vectorfield estimates the point-wise local rate of atrophy between the

shows the cm-rep model fitted to the “DEF-MSK” postmortem hippocampus atlas (sees for three subjects in the TLE study (subjects L1, R2 and L4; selected arbitrarily). The leftcolumn shows the skeleton surface m of the fitted cm-rep model, with the color mapokes are line segments extend from the skeleton to the boundary, are orthogonal to theokes extending to boundary half b− are shown in blue, and spokes that extend to b− arethe postmortem atlas to the in vivo data. The last column shows the boundary surface


healthy and diseased hippocampi. By integrating this Jacobian fieldover each subfield label (computed for the healthy hippocampus inthe previous section), we get an estimate of atrophy for that subfield.We express atrophy using the asymmetry index (AI):

AI =Vcontralateral−Vipsilateral

Vcontralateral + Vipsilateral;

where V is the estimated volume of a subfield. Positive values of theasymmetry index indicate atrophy in the diseased hippocampus.

Experimental results

Fig. 10 illustrates the shape-based normalization approach. Itshows examples of cm-rep models fitted to the postmortem atlas andto individual hippocampus segmentations from the in vivo data. Theaverage Dice overlap between the segmentations of healthy hippo-campi and the cm-rep models fitted to them was 0.921, and theaverage distance from the cm-rep model boundary to the segmenta-tion boundary was 0.230 mm. For the postmortem atlas, the overlapbetween the segmentation and the model was 0.946 and the average

Fig. 11. Example images from the in vivo temporal lobe epilepsy hippocampal asymmetry ancolumn shows a sagittal cross-section of the region surrounding the hippocampus contra(“diseased”) hippocampus region after flipping across the midsagittal plane and rigid alignmthe transformation between the healthy hippocampus and diseased hippocampus computedin the healthy hippocampus maps to a smaller region in the diseased hippocampus; posihippocampus (i.e., to the manual segmentation). The last column shows the estimation of thfrom the postmortem atlas using shape-based normalization.

distance from the cm-rep model boundary to the segmentationboundary was 0.126 mm. The fitting accuracy is higher for thepostmortem hippocampus because its boundary is much smoother(see Fig. 10) and, thus, easier to fit using a deformable geometricmodel. The subfield mapping obtained using shape-based normal-ization is illustrated in the last column of Fig. 11. Image-basednormalization is illustrated in Fig. 11, which shows examples ofhealthy hippocampus images, rigidly aligned diseased hippocampusimages, and Jacobian determinant fields computed by SyN imageregistration.

The estimated asymmetry index for subfields CA1, CA2/3 and DGfor each subject is plotted in Fig. 13. Asymmetry in CA2/3 tends to begreater than in CA1; likewise, asymmetry is greater for DG than forCA1. These trends are confirmed by paired Student's t-tests: p=0.0008for the AI(CA1)bAI(CA2/3) comparison and p=0.0001 for the AI(CA1)bAI(DG) comparison. The test does not detect significantdifferences (p=0.31) between AI(CA2/3) and AI(DG). To help under-stand the difference between atrophy across subfields, Fig. 12 projects(using mappings computed by shape-based normalization) theJacobian maps from all subjects in the study into the space of thepostmortem hippocampus atlas. In the coronal slices, the Jacobian

alysis experiment. Each row shows data from one of the subjects in the study. The firstlateral to the seizure focus (“healthy” side). The second column shows the ipsilateralent to the healthy hippocampus. The third column plots the logarithm of the Jacobian ofby deformable image registration. Negative values (blue) indicate that a small region Rtive values (red) indicate the opposite. The Jacobian map is restricted to the healthye location of hippocampal subfields in the healthy hippocampus. Subfields are mapped

Fig. 12. This figure plots the healthy/diseased asymmetry in TLE in a reference space, i.e., in the space of the postmortem hippocampus atlas. The left column shows 5 coronal slicesand a sagittal slice through the postmortem atlas “DEF-MSK” (see Fig. 3), which is the atlas used in the TLE experiment. The second column shows the subfield labeling of the “DEF-MSK” atlas. The third column shows the logarithm of the Jacobian determinant of themapping from the healthy hippocampus to the diseased hippocampus, averaged over all subjects.This average Jacobian map is computed by using shape-based normalization to transform the log-Jacobian maps shown in Fig. 11 back into atlas space, followed by computing theaverage. Negative values (blue) indicate local volumetric decrease (disease-associated atrophy); positive values (red) indicate local increase in volume. The last column showscorresponding slices from the non-masked postmortem atlas “DEF”, for the purpose of visualizing the adjacent tissues.


appears lower along the middle section of the hippocampus, whereDG and CA2/3 are.

Discussion

Contributions

The main contributions of this work are (1) the acquisition andlabeling of high-resolution nearly isotropic images of the wholehippocampus; (2) a demonstration of the efficacy of large deformationdiffeomorphic mapping techniques normalizing high-resolution post-mortem imaging data; (3) a proof-of-concept experiment showing theapplication of the postmortem atlas to in vivo studies; and (4) open-source public release of a data set and an atlas that have the potential

to become valuable resources for investigators studying subfield-levelhippocampal morphometry using in vivo imaging.

From the imaging point of view, although a number of authorshave imaged themedial temporal lobe at high field andwith very highresolution (Wieshmann et al., 1999; Fatterpekar et al., 2002; Chakereset al., 2005), there has not been an emphasis on imaging thehippocampus as a whole at isotropic or nearly isotropic resolution,as we have done in this paper. Indeed, a common theme in priorpostmortem imaging work is the use of MRI as a neuropathology tool,with comparisons between structures visible in MRI slices andcorresponding histology slices. This led the authors to select protocolswith high in-plane resolution and thick slices. Our principal goal inconducting the postmortem imaging study was to build a three-dimensional model of hippocampal shape and of the spatial


distribution of hippocampal subfields. The main benefit of imagingwith isotropic voxels, which has been our approach, is the ability tolabel the structures in the hippocampus in three dimensions.

The computational anatomy aspect of this work, i.e., building anatlas from a set of images, draws on existing methodology. Theapproach that we use, i.e., normalization via large deformation diffeo-morphic registration and atlas generation via minimization ofgeodesic distance and shape averaging (Avants et al., 2008) is state-of-the-art and builds on some of the most highly acclaimedmethodology in the computational anatomy field (Miller et al.,2006; Beg et al., 2005; Joshi and Miller, 2000; Christensen et al.,1997; Guimond et al., 2000; Davis et al., 2004). Probably, the mostcommon application of diffeomorphic registration techniques is towhole-brain normalization, and the images collected in this studypose special challenges for these techniques due to higher resolution,limited and non-uniform field of view, and presence of variousartifacts. We demonstrate that with the help of basic aids, such asmasks outlining the structure of interest, diffeomorphic techniquescan perform reasonably well on challenging data, and that withoutthese aides, the loss in performance is moderate. Our experience maybe of benefit to other researchers seeking to apply diffeomorphictechniques to “non-standard” imaging data, in particular ultra high-resolution in vivo imaging of hippocampal subfields that is becomingincreasingly relevant with wider availability of 7 Tesla scanners.

The data collected and generated in this study are being madeavailable through the NIH-sponsored Neuroinformatics Tools andResources Clearinghouse (NITRC) under the name “Penn Hippocam-pus Atlas”. This resource includes raw imaging data, manualsegmentations, affine and deformable transforms from native imagespace to atlas space, atlases and consensus segmentations. As dataacquisition continues, new datasets will be continually released intothe public domain. All data thus distributed is fully anonymous. Weenvision Penn Hippocampus Atlas as a resource that will contribute tothe development of image analysis pipelines that allow automatic orsemi-automatic estimation of the location, size and shape of hippo-campal subfields in in vivo imaging data.

The experiment presented in the Application to in vivo studiessection demonstrates an application of the postmortem atlas on invivo data. Hippocampal volumetry has been widely used to studyatrophy in TLE patients (Cendes, 1993; Bernasconi et al., 2003).However, as different regions within the hippocampus are known tobe affected by the disease process differently, both in terms of theirhistopathology and seizure onset zones (King et al., 1997), the abilityto perform subfield-level morphometry may provide greater insight

Fig. 13. Plots of volumetric asymmetry between the subfields of the healthy and diseased hipplines representing subjects with the right side of seizure and red lines corresponding to thevolume between the subfield in the healthy hippocampus and the same subfield in the diseathan asymmetry in CA1 (paired Student t-test p=0.0008) and asymmetry in DG greater tha

into the associated atrophy patterns. Neuronal loss in DG, as well as inCA1 and CA3, is a common histological finding in sclerotic hippocampi(Swartz et al., 2006; Bote et al., 2008), which may lead to thevolumetric asymmetry in these subfields as observed here (Fig. 13).Greater relative atrophy in DG found here is consistent with the factthat the neurons in the hilus of DG are considered to be amongst themost vulnerable in TLE (Robbins et al., 1991), and is sometimes foundmore consistently than atrophy in CA1/CA3 (Swartz et al., 2006).However, other studies have found greater cellular neuropathology inCA1 (Van Paesschen et al., 1997). The use of subfield-level deformationbased morphometry in quantifying atrophy is novel in the study ofTLE, and further studies are required to fully evaluate the efficacy ofthis approach. Nonetheless, subfield-specific morphometric techni-ques have been shown to be useful in other clinical applications,including Alzheimer's disease (Mueller et al., 2007), and the proof-of-concept application presented here shows promising preliminaryresults.

Potential future applications

This section briefly outlines possible approaches for leveraging thehippocampus atlas in in vivo imaging studies. This section isspeculative, and its intention is to simply sketch out the range ofapproaches in which the atlas may be used. One of the mainmotivations for releasing the atlas publicly is to allow the developersof hippocampus-aimed approaches to leverage it in their algorithms.

We envision essentially three ways to map the postmortemhippocampus atlas to in vivo data: using shape correspondences,using image similarity, and using some combination of the two, as wedid in the Experimental results section. Shape correspondences can befound between manual segmentations of the hippocampus in in vivodata and segmentations in the postmortem atlas. Such correspon-dences, if interpolated over the interior of each segmentation, couldbe used tomap subfield labels from the atlas to image space. However,the accuracy of shape-based mapping is highly dependent on thequality of the manual segmentation, which in itself is a challengingproblem in clinical MRI, especially in the presence of neurodegenera-tion (Carmichael et al., 2005). Furthermore, the interpolation ofcorrespondences onto the interior of an object means that themapping of internal subfield boundaries (i.e., boundaries betweensubfields that are not curves on the hippocampus boundary, such asmost of the DG/CA boundary) would be dependent solely on boundaryshape cues, making inferences derived from the direct comparison ofinternal subfields (e.g., DG) across subjects questionable at the least. A

ocampi in the in vivoTLE study. Each subject is represented by a line segment, with blueleft side of seizure. The y-axis plots the asymmetry index, i.e., the relative decrease insed hippocampus. The plots illustrate the following trends: asymmetry in CA2/3 greatern asymmetry in CA1 (paired Student t-test pb0.0001).


more reliable approach, very similar in spirit to the approach taken inthe Experimental results section, is to use shape correspondences tomap the postmortem atlas onto the hippocampus segmentation in a invivo template, to which in vivo data are normalized using deformableimage registration. Integration of the Jacobian field associated withthis registration over each subfield defined in template space wouldyield subfield-specific volumetric features. Finally, for some lesscommon imaging modalities, such as T2-weighted in vivo hippocam-pus-specific imaging at 3 T, as used inMueller et al. (2007), Apostolovaet al. (2006) and Zeineh et al. (2003), or emerging 7 Tesla imaging (Choet al., 2008; Li et al., 2008; Bernasconi et al., 2008), it may becomepossible to directly map the subfield information from the post-mortem atlas to individual in vivo images on the basis of imagesimilarity, in particular if aids likemasks and landmarks are employed.This is the direction of our future work.

Another potential application of a detailed hippocampus atlas is touse it as a prior for deformable model segmentation of thehippocampus in vivo. One of the problems facing model-basedsegmentation is that the apparent topology of the hippocampus inclinical resolution images can change substantially in the presence ofneurodegenerative disease. Some of the topological differencesobserved in in vivo modalities are due to the boundaries betweenhippocampal layers not visible in the healthy adult hippocampusappearing in aging and dementia. By inferring the topology and shapeof the hippocampal layers from postmortem data and using it as ashape prior during segmentation it may be possible to reduce thenumber of possible topological configurations and thus enable betternormalization of the hippocampus between cohorts.

Limitations

One of the limitations of this work is that the current atlas is basedon a relatively small number of samples. This causes us to combinesamples from both hemispheres in a single atlas, whereas it may bemore appropriate to have separate atlases of the left and righthippocampus. Even though we believe that in its current state theatlas can be of potential benefit to other researchers, we intend tocontinue acquiring and labeling images of the postmortem hippocam-pus in different clinical populations. Specifically, we plan to image thehippocampus region in presence of Alzheimer's disease and semanticdementia and to build atlases specific to these neurodegenerativedisorders. Asmore samples are added to the atlas, itwill become feasibleto not onlymodel the average hippocampus shape, but also the variationin the shape and intensity of the hippocampus. Along with collectingmore data, we aim to expand the number of structures labeled in eachsample to include those shown in Fig. 4. Having a computational modelof the relationship between the hippocampal subfields, the amygdala,the fimbria and the subiculum will be of great benefit in in vivoapplications where these structures are partially discernable.

Furthermore, the atlas would benefit from an extensive validationof the current postmortem image segmentation protocol. In post-mortem data, a direct validation is possible by performing histolo-gical staining of the samples. Hippocampal strata and subfields canbe differentiated in histological slices, providing a good estimate ofground truth. However, matching 2D histological slices withpossible tearing and distortion to the 3D MRI atlas will posesubstantial challenges. We have previously developed similar 2D-3Dregistration techniques for the mouse brain (Yushkevich et al.,2006a), and plan to apply them to the human hippocampus in ourfuture work.

A possible limitation associated with applying the postmortematlas to in vivo data is that one must be cautious to account formorphological differences between the living brain and formalin-fixedsamples. Cutting of the samples may introduce additional deforma-tions, although we take care to leave at least 1 cm of tissue around thehippocampus to reduce such distortion. Nevertheless, some form of

global shape correction will likely be necessary to account for thesedifferences.

In the current atlas, the consensus segmentation of the hippo-campus does not provide guarantees on the topology of the subfields.Indeed, as seen in Fig. 8, some of the thinner subfields have multipleconnected components. This can be corrected by deforming atemplate with correct topology to the consensus segmentation usinga label-by-label overlap metric. Incorrect topology may pose difficultyfor applications that require geometrical modeling of the subfields,but does not preclude one from performingmorphometry, similarly tohow many brain tissue class segmentation algorithms can beleveraged without guarantees on the topology of the cortical mantle.

Acknowledgments

This work was supported by the NIH grants AG027785, NS061111,MH068066, NS058386 and NS045839, the McCabe Foundation PilotAward and the Pilot Project Award from the Penn ComprehensiveNeuroscience Center. We are deeply thankful to Prof. Nicholas K.Gonatas, M.D., Igor Tsimberg and Joseph DiRienzi of the Department ofPathology and Laboratory Medicine at the University of Pennsylvaniafor their help with obtaining specimens for this project. PostmortemMR studies described in this manuscript were performed in the SmallAnimal Imaging Facility in the Department of Radiology at theUniversity of Pennsylvania.

Appendix: Incorporating landmarks in SyN

Define {Q} as the set of points describing the landmark labelswithin the template space, I. That is, Qi∈ {Q} is the Euclidean positionof some labeled point in patient space. Define {q} as the set of pointsdescribing the landmark labels within the individual space, J. Notethat the cardinality of {Q}, denoted n, and the cardinality of {q}, m,may not be the same. In most cases, however, m≈n. Then, given I, J,{q}, {Q}, and some initial estimate for ϕ1, ϕ2, find the shortest ϕ∈Diff0such that:

1. ‖I ϕ1 x;0:5ð Þð Þ−J ϕ2 x;0:5ð Þð Þ‖ is minimal,2. ∑jw

jq‖ϕ2 qj;0:5

� �−CL ϕ1 Qf g;0:5ð Þð Þϕ2qj

‖ is minimal and3. ∑iwi

Q‖ϕ1 Q i;0:5ð Þ−CL ϕ2 qf g;0:5ð Þð Þϕ1Qi‖ is minimal.

The function, CL({p}q), represents the “closest point” operation.That is, CL returns the smallest distance between the point set, {p}with respect to the point q. This translates to the following variationalproblem, with linear operator L,

ESyN ðI; J; fqg; fQgÞ = inf�1

inf�2

∫0:5t = 0 fhLv1;v1i + hLv2;v2igdt + ∫Ω½ωjIðϕ1ðx;0:5ÞÞ �−Jðϕ2ðx;0:5ÞÞj2 +∑

jwj

q‖ϕ2 ðqj;0:5Þ−CLðϕ1 �ðfQg;0:5ÞÞϕ2qj‖2

+ ∑iwi

Q ‖ϕ1ðQ i;0:5Þ−CLðϕ2 ðfqg;0:5ÞÞϕ1Qi‖2�dX with ω a scalar and wq

and wQ scalar, spatially varying weights and each ϕi the solution of:

dϕi x; tð Þ=dt = vi ϕi x; tð Þ; tð Þ with ϕi x;0ð Þ = x: ð3Þ

In this work, we set ω=1 as a constant, wqj is a Gaussian smoothed

dirac delta function centered at each deformed landmark with value 3at the peak and wQ

i starts with the same value as wqj . We optimize this

problem by gradient descent on the total energy, in the same multi-resolution framework as used for non-landmarked SyN.

References

Amaral, D.G., 1978. A Golgi study of cell types in the hilar region of the hippocampus inthe rat. J. Comp. Neurol. 182 (4 Pt 2), 851–914.

Apostolova, L.G., Dinov, I.D., Dutton, R.A., Hayashi, K.M., Toga, A.W., Cummings, J.L.,Thompson, P.M., 2006. 3d comparison of hippocampal atrophy in amnestic mildcognitive impairment and Alzheimer's disease. Brain 129 (Pt 11), 2867–2873.


Avants, B., Gee, J.C., 2004. Geodesic estimation for large deformation anatomical shapeaveraging and interpolation. Neuroimage 23 (Suppl. 1), S139–S150.

Avants, B.B., Epstein, C.L., Grossman, M., Gee, J.C., 2008. Symmetric diffeomorphic imageregistration with cross-correlation: evaluating automated labeling of elderly andneurodegenerative brain. Med. Image Anal. 12 (1), 26–41.

Beg, M.F., Miller, M.I., Trouvé, A., Younes, L., 2005. Computing large deformation metricmappings via geodesic flows of diffeomorphisms. Int. J. Comput. Vision 61 (2),139–157 ISSN 0920-5691.

Bernasconi, N., Bernasconi, A., Caramanos, Z., Antel, S., Andermann, F., Arnold, D., 2003.Mesial temporal damage in temporal lobe epilepsy: a volumetric MRI study of thehippocampus, amygdala and parahippocampal region. Brain 126 (2), 462.

Bernasconi, N., Mansi, T., Postelnicu, G., Augustinack, J., Fischl, B., 2008. Quantitativecytoarchitectural mapping of the parahippocampal region. Proceedings 16thScientific Meeting, International Society for Magnetic Resonance in Medicine,Toronto, p. 2217.

Bobinski, M., Wegiel, J., Tarnawski, M., Bobinski, M., Reisberg, B., de Leon, M.J., Miller,D.C., Wisniewski, H.M., 1997. Relationships between regional neuronal loss andneurofibrillary changes in the hippocampal formation and duration and severityof Alzheimer disease. J. Neuropathol. Exp. Neurol. 56 (4), 414–420.

Bote, R., Blázquez-Llorca, L., Fernández-Gil, M., Alonso-Nanclares, L., Muñoz, A., DeFelipe, J., 2008. Hippocampal sclerosis: histopathology substrate and magneticresonance imaging. Semin. Ultrasound CT MR 29 (1), 2–14.

Braak, H., Braak, E., 1991. Neuropathological stageing of Alzheimer-related changes. ActaNeuropathol. 82 (4), 239–259.

Carmichael, O.T., Aizenstein, H.A., Davis, S.W., Becker, J.T., Thompson, P.M., Meltzer, C.C.,Liu, Y., 2005. Atlas-based hippocampus segmentation in Alzheimer's disease andmild cognitive impairment. Neuroimage 27 (4), 979–990.

Cendes, F., 1993. MRI volumetric measurement of amygdala and hippocampus intemporal lobe epilepsy. Neurology 43 (4), 719–725.

Chakeres, D.W., Whitaker, C.D.S., Dashner, R.A., Scharre, D.W., Beversdorf, D.Q.,Raychaudhury, A., Schmalbrock, P., 2005. High-resolution 8 Tesla imaging of theformalin-fixed normal human hippocampus. Clin. Anat. 18 (2), 88–91.

Cho, Z., Kim, K., Kim, Y., Bae, K., Moon, C., 2008. High-resolution in vivo MR imaging ofthe human hippocampus at 7 Tesla. Proceedings 16th Scientific Meeting,International Society for Magnetic Resonance in Medicine, Toronto, p. 2218.

Christensen, G., Joshi, S., Miller, M., 1997. Volumetric transformation of brain anatomy.IEEE Trans. Med. Imag. 16, 864–877.

Csernansky, J.G., Wang, L., Swank, J., Miller, J.P., Gado, M., McKeel, D., Miller, M.I., Morris,J.C., 2005. Preclinical detection of Alzheimer's disease: hippocampal shape andvolume predict dementia onset in the elderly. Neuroimage 25 (3), 783–792.

Davis, B., Lorenzen, P., Joshi, S.C., 2004. Large deformation minimum mean squarederror template estimation for computational anatomy. IEEE International Sympo-sium on Biomedical Imaging: Macro to Nano, pp. 173–176.

Dice, L.R., 1945. Measures of the amount of ecologic association between species.Ecology 26 (3), 297–302.

Duvernoy, H.M., 2005. The Human Hippocampus, Functional Anatomy, Vascularizationand Serial Sections with MRI, third ed. Springer.

Eriksson, P.S., Perfilieva, E., Björk-Eriksson, T., Alborn, A.M., Nordborg, C., Peterson, D.A.,Gage, F.H., 1998. Neurogenesis in the adult human hippocampus. Nat. Med. 4 (11),1313–1317.

Fatterpekar, G.M., Naidich, T.P., Delman, B.N., Aguinaldo, J.G., Gultekin, S.H., Sherwood,C.C., Hof, P.R., Drayer, B.P., Fayad, Z.A., 2002. Cytoarchitecture of the human cerebralcortex: MR microscopy of excised specimens at 9.4 Tesla. AJNR Am. J. Neuroradiol.23 (8), 1313–1321.

Fischl, B., Salat, D.H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., van der Kouwe, A.,Killiany, R., Kennedy, D., Klaveness, S., Montillo, A., Makris, N., Rosen, B., Dale, A.M.,2002. Whole brain segmentation: automated labeling of neuroanatomicalstructures in the human brain. Neuron 33 (3), 341–355.

Guimond, A., Meunier, J., Thirion, J.P., 2000. Average brain models: a convergence study.Comput. Vis. Image Underst. 77 (9), 192–210 ISSN 1077-3142.

Hsu, Y.Y., Schuff, N., Du, A.T., Mark, K., Zhu, X., Hardin, D., Weiner, M.W., 2002.Comparison of automated and manual MRI volumetry of hippocampus in normalaging and dementia. J. Magn. Reson. Imaging 16 (3), 305–310.

Jack, C.R., Theodore, W.H., Cook, M., McCarthy, G., 1995. MRI-based hippocampalvolumetrics: data acquisition, normal ranges, and optimal protocol. Magn. Reson.Imaging 13 (8), 1057–1064.

Joshi, S., Miller, M.I., 2000. Landmark matching via large deformation diffeomorphisms.IEEE Trans. Image Process. 9 (8), 1357–1370.

Kelemen, A., Székely, G., Gerig, G., 1999. Elastic model-based segmentation of 3-Dneuroradiological data sets. IEEE Trans. Med. Imag. 18 (10), 828–839.

King, D., Bronen, R., Spencer, D., Spencer, S., 1997. Topographic distribution of seizureonset and hippocampal atrophy: relationship between MRI and depth EEG.Electroencephalogr. Clin. Neurophysiol. 103 (6), 692–697.

Kirwan, C.B., Jones, C.K., Miller, M.I., Stark, C.E.L., 2007. High-resolution fMRIinvestigation of the medial temporal lobe. Hum. Brain Mapp. 28 (10), 959–966.

Kuhn, H.G., Dickinson-Anson, H., Gage, F.H., 1996. Neurogenesis in the dentate gyrus ofthe adult rat: age-related decrease of neuronal progenitor proliferation. J. Neurosci.16 (6), 2027–2033.

Li, T., Yao, B., van Gelderen, P., Merkle, H., Koresky, A., Duyn, J., 2008. Corticalarchitecture of the human hippocampus. Proceedings 16th Scientific Meeting,International Society for Magnetic Resonance in Medicine, Toronto, p. 2219.

Loop, C., DeRose, T., 1990. Generalized b-spline surfaces of arbitrary topology. ComputerGraphics (ACM SIGGRAPH Proceedings), pp. 347–356.

Miller, M.I., Trouvé, A., Younes, L., 2006. Geodesic shooting for computational anatomy.J. Math. Imaging Vis. 24 (2), 209–228 ISSN 0924-9907.

Moscovitch, M., Nadel, L., Winocur, G., Gilboa, A., Rosenbaum, R.S., 2006. Thecognitive neuroscience of remote episodic, semantic and spatial memory. Curr.Opin. Neurobiol. 16 (2), 179–190.

Mueller, S.G., Schuff, N., Raptentsetsang, S., Stables, L., Weiner, M.W., 2007. Distinctatrophy pattern in hippocampal subfields in Alzheimer's disease (AD) and mildcognitive impairment (MCI). Alzheimer's and Dementia 3 (3), S113.

Narayan, V.M., Kimberg, D.Y., Tang, K.Z., Detre, J.A., 2005. Experimental design forfunctional MRI of scene memory encoding. Epilepsy Behav. 6 (2), 242–249.

Pitiot, A., Delingette, H., Thompson, P.M., Ayache, N., 2004. Expert knowledge-guidedsegmentation system for brain MRI. Neuroimage 23 (Suppl. 1), S85–S96.

Pluta, J., Avants, B., Glynn, S., Awate, S., Gee, J., Detre, J., 2008. Appearance andincomplete label matching for diffeomorphic template based hippocampussegmentation. Workshop on the Computational Anatomy and Physiology of theHippocampus (CAPH). New York.

Rabin, M.L., Narayan, V.M., Kimberg, D.Y., Casasanto, D.J., Glosser, G., Tracy, J.I., French,J.A., Sperling, M.R., Detre, J.A., 2004. Functional MRI predicts post-surgicalmemory following temporal lobectomy. Brain 127 (Pt 10), 2286–2298.

Robbins, R., Brines, M., Kim, J., Adrian, T., de Lanerolle, N., Welsh, S., Spencer, D., 1991. Aselective loss of somatostatin in the hippocampus of patients with temporal lobeepilepsy. Ann. Neurol. 29 (3), 325–332.

Shen, D., Moffat, S., Resnick, S.M., Davatzikos, C., 2002. Measuring size and shape of thehippocampus in MR images using a deformable shape model. Neuroimage 15 (2),422–434.

Shenton, M.E., Gerig, G., McCarley, R.W., Székely, G., Kikinis, R., 2002. Amygdala-hippocampal shape differences in schizophrenia: the application of 3D shapemodels to volumetric MR data. Psychiatry Res. 115 (1–2), 15–35.

Squire, L.R., Stark, C.E.L., Clarkx, R.E., 2004. The medial temporal lobe. Annu. Rev.Neurosci. 27, 279–306.

Styner, M., Lieberman, J.A., Pantazis, D., Gerig, G., 2004. Boundary and medial shapeanalysis of the hippocampus in schizophrenia. Med. Image Anal. 8 (3), 197–203.

Swartz, B., Houser, C., Tomiyasu, U., Walsh, G., DeSalles, A., Rich, J., Delgado-Escueta, A.,2006. Hippocampal Cell Loss in Posttraumatic Human Epilepsy. Epilepsia 47 (8),1373–1382.

Van Leemput, K., Bakkour, A., Benner, T., Wiggins, G., Wald, L., Augustinack, J., Dickerson,B., Golland, P., Fischl, B., 2008. Model-based segmentation of hippocampal subfieldsin ultra-high resolution in vivo MRI. Medical Image Computing and Computer-Assisted Intervention, MICCAI.

Van Paesschen, W., Revesz, T., Duncan, J., King, M., Connelly, A., 1997. Quantitativeneuropathology and quantitative magnetic resonance imaging of the hippocampusin temporal lobe epilepsy. Ann. Neurol. 42 (5), 756–766.

van Praag, H., Schinder, A.F., Christie, B.R., Toni, N., Palmer, T.D., Gage, F.H.,2002. Functional neurogenesis in the adult hippocampus. Nature 415 (6875),1030–1034.

Wang, L., Miller, J.P., Gado, M.H., McKeel, D.W., Rothermich, M., Miller, M.I., Morris, J.C.,Csernansky, J.G., 2006. Abnormalities of hippocampal surface structure in verymilddementia of the Alzheimer type. Neuroimage 30 (1), 52–60.

Warfield, S.K., Zou, K.H., Wells, W.M., 2004. Simultaneous truth and performance levelestimation (STAPLE): an algorithm for the validation of image segmentation. IEEETrans. Med. Imag. 23 (7), 903–921.

West, M.J., Kawas, C.H., Stewart, W.F., Rudow, G.L., Troncoso, J.C., 2004. Hippocampalneurons in pre-clinical Alzheimer's disease. Neurobiol. Aging 25 (9), 1205–1212.

Wieshmann, U.C., Symms, M.R., Mottershead, J.P., MacManus, D.G., Barker, G.J., Tofts,P.S., Revesz, T., Stevens, J.M., Shorvon, S.D., 1999. Hippocampal layers on highresolution magnetic resonance images: real or imaginary? J. Anat. 195 (Pt 1),131–135.

Xu, G., 2004. Discrete Laplace-Beltrami operators and their convergence. Comput. AidedGeom. Des. 21 (8), 767–784.

Yushkevich, P.A., Avants, B.B., Ng, L., Hawrylycz, M., Burstein, P.D., Zhang, H., Gee, J.C.,2006a. 3D mouse brain reconstruction from histology using a coarse-to-fineapproach. Workshop on Biomedical Image Registration (WBIR).

Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., Gerig, G., 2006b.User-guided 3D active contour segmentation of anatomical structures: significantlyimproved efficiency and reliability. Neuroimage 31 (3), 1116–1128.

Yushkevich, P.A., Zhang, H., Gee, J., 2006c. Continuous medial representation foranatomical structures. IEEE Trans. Med. Imag. 25 (2), 1547–1564.

Yushkevich, P.A., Detre, J.A., Mechanic-Hamilton, D., Fernández-Seara, M.A., Tang, K.Z.,Hoang, A., Korczykowski, M., Zhang, H., Gee, J.C., 2007. Hippocampus-specific fMRIgroup activation analysis using the continuous medial representation. Neuroimage35 (4), 1516–1530.

Yushkevich, P.A., Zhang, H., Simon, T.J., Gee, J.C., 2008. Structure-specific statisticalmapping of white matter tracts. Neuroimage 41 (2), 448–461.

Zeineh, M.M., Engel, S.A., Thompson, P.M., Bookheimer, S.Y., 2003. Dynamics of thehippocampus during encoding and retrieval of face-name pairs. Science 299 (5606),577–580.

Zou, K.H., Warfield, S.K., Bharatha, A., Tempany, C.M.C., Kaus, M.R., Haker III, S.J., W.M.W.,Jolesz, F.A., Kikinis, R., 2004. Statistical validation of image segmentation qualitybased on a spatial overlap index. Acad. Radiol. 11 (2), 178–189.

a high-resolution computational atlas of the human hippocampus...

Documents