UNSUPERVISED 3D SEGMENTATION OF HIPPOCAMPUS IN

BRAIN MR IMAGES

Keywords: Hippocampus, MRI segmentation, Water flow, deep brain structure, surface evolution, Region growing, 3D
segmentation, hippocampal volumetry.

Abstract: The most widely followed procedure for diagnosis and prognosis of dementia is structural neuroimaging of
hippocampus by means of MR. Hippocampus segmentation is of wide interest as it enables quantitative as-
sessment of the structure. In this paper, we propose an algorithm for hippocampus segmentation that is un-
supervised and image driven. It is based on a hybrid approach which combines a coarse segmentation and
surface evolution. A coarse solution is derived using region growing which is further refined using a modified
version of the physics based water flow model (Liu and Nixon, 2007). The proposed method has been tested
on a publicly available dataset. The performance of this method is assessed using Dice coefficient against the
ground truth provided for 25 volume images. It is consistent across volumes and the average Dice values are
comparable to a multi-atlas based method reported on a subset of the same dataset.

1 INTRODUCTION tern matching and brain boundary shift integral mea-

surements. (Schmidt et al., 2009). Hippocampus is
Dementia is a clinical syndrome that affects memory the structure responsible for long term memory in the
and cognitive ability of a person. It is known to be brain and its atrophy is an early and specific marker
caused due to traumatic brain injury, neurodegenera- of dementia. The severity of hippocampal atrophy is
tive diseases, bacterial infections, prolonged epilep- directly related to progress of the underlying disease
tic seizures, and so on. Some common types of de- and hence reflects the extent of cognitive impairment.
mentia are mild cognitive impairment (MCI) or incip- Although atrophy rates have been observed to be
ient dementia, Alzheimer disease (AD) or dementia of larger in entorhinal cortex than in hippocampus, diffi-
the Alzheimer type (DAT), dementia with Lewy bod- culty in unambiguously defining this structure makes
ies (DLB) and fronto-temporal dementia (FTD). De- its measurement highly variable. Thus, hippocampus
mentia is usually diagnosed based on clinical obser- is the region of interest in differential diagnosis of AD
vations, presence of characteristic neurological and and mild cognitive impairment (MCI).
neuropsychological features. Onset of dementia is Manual labelling by experts is considered as gold
difficult to diagnose with these methods. Therefore, standard of hippocampal segmentation. But this is
along with clinical diagnosis, structural neuroimag- laborious, time consuming, prone to inconsistency
ing is used to enable early diagnosis so that it can and inter-expert labelling variability. Advanced im-
be treated at the onset. Apart from diagnosis, struc- age analysis techniques have been used to minimise
tural neuroimaging helps to distinguish between dif- these problems by developing semi-automatic hip-
ferent types of dementia, to differentiate between nor- pocampal segmentation methods or further, by ob-
mal aging and dementia and in differential diagnosis. taining reliable initialisation, fully automatic segmen-
Some of the methods to assess change in brain vol- tation methods. From medical imaging viewpoint,
ume by means of structural neuroimaging are volume- research in segmentation of deep brain structures in
try, voxel based morphometry (VBM), cortical pat- general and hippocampus in particular is motivated
by various factors like i) development of computer or combination of these labelled data (image) can be
aided diagnosis solution by methods. The attraction used as an atlas. The disadvantage of using cohort at-
here is the potential for high precision and consistency las is the requirement of an expert and one or more
over manual segmentation which in turn offers reli- labelled data. Both of these methods have the com-
able volumetry (Haller et al., 1997; Juottonen et al., mon issue of inability to handle variation in data over
1999; Bernasconi et al., 2003); ii) study of the char- larger sets. Hence, atlases are more suitable for ini-
acteristic pattern of a particular disease or normal ag- tialisation than for complete segmentation.
ing (Chupin et al., 2008; Hsu et al., 2002; Dicker- An alternative way to use domain knowledge is
son et al., 2001); iii) development of generalised seg- with pattern classification and machine learning tech-
mentation algorithm applicable to various structures niques. Statistical and anatomical prior knowledge
(Pitiot et al., 2002; Morra et al., 2008b; Yang et al., are learnt and used to localise and segment hippocam-
2004); iv) acceleration of drug trials and population pus in (Morra et al., 2008b; Morra et al., 2008a;
study of diseases (Morra et al., 2008a). In general, Morra et al., 2010; Akselrod-Ballin et al., 2007;
the hippocampus, with weak edges and inhomogene- Sabuncu et al., 2009; Leemput et al., 2009). Appear-
ity, acts as a good test case for a neurosegmentation ance based features are used to build a context model
algorithm in both 2D and 3D (Morra et al., 2010; in (Morra et al., 2008b; Morra et al., 2008a; Morra
Pitiot et al., 2002; Yushkevich et al., 2006; Nain et al., et al., 2010), while segmentation is posed as an ex-
2007). Given the focus of this paper on hippocampus pectation maximisation problem in (Sabuncu et al.,
segmentation, next reviews the existing methods for 2009). Prior information from an atlas and segmen-
the same in some detail. tation via likelihood has been adopted in (Akselrod-
Ballin et al., 2007). Most of the techniques de-
1.1 Related Work pend upon the shape of hippocampus varying within
a given margin. Furthermore, since these techniques
are largely supervised, the choice of features and
The existing methods can be classified broadly based
modelling of the problem is critical. The performance
on the approach taken to include the domain knowl-
may not be consistent across varied real data if the
edge as follows: i) atlas and cohort atlas based, ii)
condition (hence shape) of hippocampus varies due
pattern classification and machine learning methods
to different diseases.
based and iii) deformable models or template based.
Segmentation via registration to a brain atlas is Template and deformable models help use domain
class of methods followed in (Haller et al., 1997; knowledge in a more flexible way. These methods are
Pruessner et al., 2000; Lord et al., 2007; Carmichael derived by using prior anatomical knowledge. The
et al., 2005; Akhondi-Asl et al., 2010; Ltjnen et al., structure to be segmented is modelled and parame-
2010). In these methods, a neuroanatomic atlas is terised by using B-splines, medial loci (skeletons),
registered, i.e., spatially aligned with the given sub- spherical wavelets and PCA based model in (Pitiot
ject image to delineate the boundary of the region of et al., 2002; Yushkevich et al., 2006; Nain et al.,
interest. The registration is carried out after geomet- 2007; Wang et al., 2007) respectively. (Yang et al.,
ric transformation to correct alignment and enhance- 2004) describe a maximum a posteriori estimation-
ment to correct intensity variations. Atlas-based ap- based method which uses configurations and context
proaches have the advantage that advances in regis- provided by neighbouring structures which have con-
tration techniques can be exploited. Thus, most of the sistent locations and shapes that aid in segmentation.
available standard registration algorithms and meth- This model is formulated in terms of level set func-
ods such as FreeSurfer, HAMMER, AIR, SPM, and tions. Minimising an energy function via iterated con-
FSL packages can be used for segmentation. Stan- ditional modes (ICM), initialised by prior informa-
dard atlases available are Harvard whole brain atlas tion from probabilistic atlas is described in (Chupin
and MNI-Talairach atlas. They contain the labelled et al., 2008). Although these methods have been ob-
structures of the whole brain and hence can be used to served to perform well, their design and implementa-
segment any structure without prior knowledge about tion are very complex. Shape encoding for segmenta-
the structure. However, if the subject image varies tion has to consider all possible variations of shape of
largely from the atlas in terms of age of the subject or the structure. This requires a large amount of super-
physical dimensions of the voxels, segmentation per- vision to train the model. In some of these methods,
formance suffers. To overcome this issue, a cohort anatomical information derived depends on landmark
atlas is built. In cohort atlas, experts manually seg- detection.
ment the structure of interest on one or more subject In general, while prior knowledge about brain struc-
images selected from a set of reference images. Each tures can aid the segmentation problem, there are
some costs involved in acquiring this knowledge and 2 METHOD
obtaining a good and consistent segmentation perfor-
mance: i) adequate amount of labelled data for train-
ing, ii) age, race and disease-specific matched infor- The water flow model directs the progress of a water
mation. An alternative is to use an unsupervised, hy- front based on local image properties derived from re-
brid approach. A combination of region growing and gion and edge based forces. In this paper, this model
deformable model or curve evolution is one such op- is chosen because of the following features: i) it is a
tion. Region growing is a relatively fast algorithm physics based model which does not depend on image
which can extract all parts of the region like curves topology and ii) it makes use of both gradient forces
and narrow parts. In most of the imaging modali- and region based forces which is attractive given the
ties like optical, CT and MRI, with appropriate fea- weak edged boundary of the hippocampus.
tures used for determining homogeneity criteria, re- Since the water flow algorithm acts on a water
gion growing proves to be one of the simplest meth- front, one of the simplest methods to obtain a front
ods of segmentation. However, this alone does not re- is by coarse segmentation via region growing. The
sult in an exact boundary as region growing is vulner- coarse segmentation stage works on intensity features
able especially to weak edges and non-homogeneous which are sensitive to intensity variation due to a bias
regions. The second curve evolution stage is em- field that can be present. In order to address this prob-
ployed to address this problem. The resulting surface lem, the image volume is pre-processed for bias field
is taken as input to surface evolution. This is more ac- correction. A modified fuzzy C-means algorithm pro-
curate than the conventional simple geometric surface posed in (Ahmed et al., 2002) is used for this pur-
initialisation and much faster than manual surface ini- pose. Here, the intensity inhomogeneity is estimated
tialisation. Surface evolution merely needs to adjust using fuzzy logic and used to correct the slowly vary-
the surface to exact boundaries which results in quick ing shading artefact over the image.
convergence. An image foresting transform (IFT) for The flow chart in Figure 1 depicts the proposed
coarse segmentation together with an active surface hippocampus segmentation algorithm.
solution has been proposed for liver segmentation in
CT (Pohle et al., 2003). Coarse segmentation via
region growing followed by topology-adaptive snake
evolution has been demonstrated for extraction of the
whole brain and a tumour from MRI and CT scans.
The refinement stage of snake mesh evolution is based
on topology adaptive technique proposed in (McIner-
ney and Terzopoulos, 1999).

In this paper, a non-parametric, hybrid method

is proposed for hippocampus segmentation which
utilises only local image information. The method
follows the coarse segmentation and surface evolution
framework with the key differences being: the surface
is not parameterised and is evolved using a physics
based water flow model (Liu and Nixon, 2007). The
water flow model has been applied to 2D image seg-
mentation of synthetic images. It has been demon-
strated (but not rigorously assessed) on some 2D real
cases such as grey-white matter interface, femur from
MRI, carotid artery in MR angiogram and retinal ves-
sels; and 3D lateral ventricles. Most of these struc-
tures are homogeneous with well defined boundaries.
In this paper, it is demonstrated how the water flow
model can be adapted to a relatively inhomogeneous,
partially weak edged hippocampus. It is demon-
strated that, in spite of its simplicity, our hybrid algo-
rithm yields results comparable to a recently reported Figure 1: Flow chart of the proposed method
method. The following section describes this method
in detail.
2.1 Coarse Segmentation set to the gradient magnitude (edge strength) other-
wise. This results in a force which is always directed
This stage begins with few seeds initialised by the away from the region. In a cross-sectional area A (set
user. The protocol of choosing multiple seeds, in- to unity), the flow velocity gained by a voxel is
stead of one, is used so that longitudinal extremes
of hippocampus can be approximated and an initial F f (ri )
v= (4)
modelling of intensity variation can be obtained. In AR
the process of region growing, these seeds are used where R is the flow resistance given by
to check for homogeneity in their immediate neigh-
bourhood and decide whether or not to include those R = ekE(ri ) (5)
voxels into the region. The parameters used in ho- where E(ri ) is the edge strength (magnitude of the
mogeneity verification are: deviation of a candidate’s gradient) at location ri and k is the resistance control-
voxel value from the mean voxel value of the region ling factor. Higher values of k leads to higher sensi-
and the gradient magnitude at the candidate voxel’s tivity to edges.
location. In addition to the static field force F f (ri ), image
The upper and lower thresholds for deviation of forces such as potential force, statistical force con-
candidate’s voxel value are Th and Tl , defined as tribute to the element flow. The potential force is cal-
culated based on the gradient of the edge map E. This
Th = µI + k1 ∗ σI (1) force is largely concentrated around the edges in the
image and hence acts as a barrier to the water flow.
In the known direction of flow i, given a target posi-
Tl = µI − k1 ∗ σI (2)
tion rt , the potential force experienced by the flowing
where µI and σI are, respectively, mean and standard element is given by
deviation of intensity within the region and k1 acts as
a weight for the permissible amount of deviation. Fp,i = ∇E(rt ) (6)
The threshold for gradient magnitude is set as Tg Water flow model also uses a region based statis-
given by tical force based on the Mumford-Shah functional. It
Tg = µG + k2 ∗ σG reflects the change in internal and external intensities
where µG and σG are, respectively, mean and standard of a closed region and its surroundings. When the tar-
deviation of gradient magnitude within the region and get voxel intensity deviates largely from the equilib-
k2 acts as a weight for the permissible amount of de- rium, this force turns negative and hence checks the
viation of gradient magnitude. flow. Statistical force due to a target voxel at position
In this paper, to maintain a low deviation, both k1 rt is given by
and k2 have been set to 1. Hence, the region grow-
ing process needs no external parameter tuning and is next nint
Fs,i = (I(rt ) − µext )2 − (I(rt ) − µint )2
adaptive. next + 1 nint + 1
(7)
where I(rt ) is the intensity value at voxel position rt ,
2.2 Fine Segmentation
n is the number of voxels and µ is the mean of inten-
sity of the internal and external regions denoted by the
The original water flow algorithm (Liu and Nixon, subscripts int and ext respectively. To maintain a bal-
2007) computes a resultant force on a voxel due to ance in the simultaneous contribution of these forces
its neighbouring voxels and decides the flow action. in the element flow, they are added by a convex equa-
Three types of image forces, namely, field, potential tion as
and statistical forces aid or oppose the flow, based on
their direction with respect to the resultant field force. Fi = αFp,i + (1 − α)Fs,i (8)
The Field force experienced by a voxel at position
ri due to its neighbourhood W is given by the force In the original water flow algorithm, the flow of an in-
field transformation (Hurley et al., 2005) as dividual waterfront voxel is not constrained explicitly
by other voxels in the front. However, its movement
(r j − ri ) should be dependent on the mobility of its surround-
F f (ri ) = ∑ L(r j )
|r j − ri |3
(3) ing voxels on the surface and the tension between
j∈W, j6=i
these voxels. Hence, a new tensile force component is
The elements of L matrix are set to −1 when a introduced which has the effect of opposing the flow.
voxel is filled with water (it is inside the region) and This force controls the stretching of the surface and
ensures a continuous surface without holes. The ten- Hippocampus is a longitudinally asymmetric
sile force experienced by a voxel at ri is calculated structure. It can be approximated to a spherical ob-
using the umbrella operator approximation of Lapla- ject near the head and narrowing cone as it progresses
cian given by (McInerney and Terzopoulos, 1999) as towards the body and tail. To make sure the surface
evolves properly to match both these shapes, two sets
1 of parameters of surface evolution - α, η and S are
Ft,i =
n ∑ (r j − ri ) (9) used, one to suit head region and one to suit the body-
j∈N
tail region. Empirically, the ratio of head region to
where r j is the position of the jth surface voxel within body-tail region is set to 0.35 : 0.65 of the total length
N neighbourhood of ri and n is the total number of of hippocampus as seen in coronal view.
voxels in that neighbourhood.
The work done by a voxel under the influence of
all the forces combined is given by 4 EXPERIMENTS AND RESULTS
m|vi |2 The proposed method is tested on a publicly available
J= + (Fi − ηFt,i )S (10)
2 brain MR dataset (Jafari-Khouzani, 2010). It con-
sists of 25 volumes of training data with associated
where m is mass of the element analogous to inten-
ground truth and 25 volumes of testing data without
sity of the voxel and S is the preset displacement. If
any ground truth. The first (training) set has been cho-
work done is positive, the element moves in the direc-
sen as the test data. This set consists of 15 volumes
tion of the resultant force for a predetermined (fixed)
of T1-weighted MR image data of each slice 256x256
distance. η controls the influence of the tensile force
pixels with voxel size 0.781x0.781x2.00mm3 and 10
experienced by the surface. The flowchart in Figure 1
volumes of T1-weighted MR image data of each slice
show the procedure for surface evolution with water
512x512 pixels with voxel size 0.39x0.39x2.00mm3 .
flow model.
Assessment is done by computing the Dice coef-
ficient. Given two sets, A and B, the Dice coefficient
between them is defined as
3 IMPLEMENTATION
2|A ∩ B|
D=
|A| + |B|
User initialised seed points (minimum 1 per slice)
are collected along the length of hippocampus in the Since a new tensile component was introduced in
coronal view of a brain MR volume image. Based the water flow algorithm, in order to assess its effec-
on the extreme points, with reasonable margin on all tiveness, the Dice coefficient was calculated for the
sides, a sub volume is extracted for faster process- proposed method with and without this component.
ing. The cluster centres for bias field correction us- Table 1 shows the maximum, minimum and the av-
ing fuzzy C-means are set to mean µ and µ + k1 Th erage Dice coefficient for 50 hippocampi in the test
and µ − k1 Tl according to equations (1) and (2) re- set. It can be observed from the table that the tensile
spectively on seed points collected. force component plays a positive role in segmenta-
The output of region growing results in a surface tion and contributes to 6% improvement, on average,
which is evolved by water flow based algorithm. in the Dice coefficient. The dataset (Jafari-Khouzani,
2010) has also been used for assessment of a multi-
while(J for all voxels != 0) atlas (derived from 10 subjects) based method pro-
for(each voxel on the surface) posed in (Akhondi-Asl et al., 2010) which reports the
calculate force F_ri;
average Dice coefficient for the testing part of the
calculate velocity v;
calculate F_i,F_ti in the direction i; dataset as (0.72 ± 0.09). This is comparable to that
calculate J; (0.68 ± 0.03) obtained on the training subset by the
if(J>0) proposed method despite the major difference in the
displace the voxel by preset distance; approaches behind the two methods: multiple atlases
endif based versus unsupervised method.
if(no movement on entire surface) Figure 2 provides a graphical comparison of the
algorithm converged;
break;
Dice coefficient for every volume. The trend of the
endif two plots is approximately similar. This trend is con-
endfor sistent with the fact that, by design, the tensile force
endwhile plays only a moderating role as per (10). In general,
it was observed that the segmentation performance performance is comparable to existing methods. The
drops when, even in absence of atrophy, size of hip- tight variance in the Dice coefficient is particularly
pocampus was very small. In such cases, the corti- encouraging. These results demonstrate that the
cal walls surrounding the hippocampus and the neigh- framework of hybrid segmentation can successfully
bouring amygdala make the true boundary difficult to tap advantages of each of the two stages.
detect resulting in an over-segmentation. The improvement in segmentation performance
Figure 3 shows some sample extracted hip- with the addition of the proposed tensile component
pocampi overlaid over the ground truth. The translu- to the water flow model indicates that the surface
cent regions of red and blue denote segmented left evolution is constrained yet yields more accurate sur-
and right hippocampus respectively by the proposed faces. A shortcoming of the proposed methods is that
method. The opaque white region under these corre- since the water flow model always evolves from in-
sponds to manual label which is used as the ground side the region to outside, the coarse segmentation
truth. Figure 4 shows the 3D surface view of seg- output has to be within the target boundary of the hip-
mented hippocampi as seen from head of the hip- pocampus. This limitation can be addressed by de-
pocampus in Figure 4(a) and as seen from its tail in signing techniques which permit a two-sided evolu-
Figure 4(b). tion of the water front.
Overall, it can be concluded that the obtained re-
Dice’s coefficient sults show promise and pave way for applying of
Water flow model Min. Max. Avg. the water flow model for segmenting other (partially)
With tensile force 0.63 0.72 0.68 ± 0.03 weak-edge structures as well. Future work will be
Without tensile force 0.49 0.68 0.62 ± 0.06 targeted at complete automation of this method by
Table 1: Comparison of performance with and without ten- reliable automatic initialisation of seed points or re-
sile force placing the region growing with any other method for
coarse segmentation.

REFERENCES
Ahmed, M., Yamany, S., Mohamed, N., Farag, A., and Mo-
riarty, T. (2002). A modified fuzzy c-means algorithm
for bias field estimation and segmentation of mri data.
Medical Imaging, IEEE Transactions on, 21(3):193 –
199.
Akhondi-Asl, A., Jafari-Khouzani, K., Elisevich, K., and
Soltanian-Zadeh, H. (2010). Hippocampal volume-
try for lateralization of temporal lobe epilepsy: Auto-
mated versus manual methods. NeuroImage, In Press,
Corrected Proof:–.
Akselrod-Ballin, A., Galun, M., Gomori, J. M., Brandt,
A., and Basri1, R. (2007). Prior knowledge driven
multiscale segmentation of brain mri. MICCAI 2007,
4792/2007:118–126.
Figure 2: Comparison of performance with and without ten- Bernasconi, N., Bernasconi, A., Caramanos, Z., Antel,
sile force S. B., Andermann, F., and Arnold, D. L. (2003).
Mesial temporal damage in temporal lobe epilepsy: a
volumetric MRI study of the hippocampus, amygdala
and parahippocampal region. Brain, 126(2):462–469.
Carmichael, O. T., Aizenstein, H. A., Davis, S. W., Becker,
5 CONCLUSION J. T., Thompson, P. M., Meltzer, C. C., and Liu, Y.
(2005). Atlas-based hippocampus segmentation in
alzheimer’s disease and mild cognitive impairment.
In this paper, a new method for segmentation of hip- NeuroImage, 27(4):979 – 990.
pocampus in MR volume images is proposed. The
Chupin, M., Cuingnet, R., Lemieux, L., Lehericy, S., Be-
advantage of this method is that it does not require an nali, H., Garnero, L., and Colliot, O. (2008). Fully au-
atlas or template and is free of parameterisation and tomatic hippocampus segmentation discriminates be-
supervision. In spite of its simplicity, this method’s tween alzheimer’s disease and normal aging - data
 (a) Tail (b) Body (c) Body

(d) Head

Figure 3: Overlay of algorithm output with manual label. Red and Blue translucent regions correspond to algorithm output
and white opaque regions correspond to manual label.

(a) 3D view from hippocampus head (b) 3D view from hippocampus tail

Figure 4: 3D surface view of the algorithm output.

from the adni database. In MICCAI Workshop on C., Wilson, R. S., Bennett, D. A., Beckett, L. A., and
CAPH. deToledo Morrell, L. (2001). Mri-derived entorhi-
nal and hippocampal atrophy in incipient and very
Dickerson, B. C., Goncharova, I., Sullivan, M. P., Forchetti,
 mild alzheimer’s disease. Neurobiology of Aging, Morra, J., Tu, Z., Apostolova, L., Green, A., Toga, A., and
22(5):747 – 754. Thompson, P. (2010). Comparison of adaboost and
Haller, J. W., Banerjee, A., Christensen, G. E., Gado, M., support vector machines for detecting alzheimer’s dis-
Joshi, S., Miller, M. I., Sheline, Y., Vannier, M. W., ease through automated hippocampal segmentation.
and Csernansky, J. G. (1997). Three-dimensional hip- Medical Imaging, IEEE Transactions on, 29(1):30 –
pocampal MR morphometry with high-dimensional 43.
transformation of a neuroanatomic atlas. Radiology, Nain, D., Haker, S., Bobick, A., and Tannenbaum, A.
202(2):504–510. (2007). Multiscale 3-d shape representation and seg-
Hsu, Y.-Y., Schuff, N., Du, A.-T., Mark, K., Zhu, X., mentation using spherical wavelets. Medical Imaging,
Hardin, D., and Weiner, M. W. (2002). Comparison of IEEE Transactions on, 26(4):598 –618.
automated and manual MRI volumetry of hippocam- Pitiot, A., Toga, A., and Thompson, P. (2002). Adaptive
pus in normal aging and dementia. Journal of Mag- elastic segmentation of brain mri via shape-model-
netic Resonance Imaging, 16(3):305–310. guided evolutionary programming. Medical Imaging,
IEEE Transactions on, 21(8):910 –923.
Hurley, D. J., Nixon, M. S., and Carter, J. N. (2005). Force
field feature extraction for ear biometrics. Computer Pohle, R., Behlau, T., and Toennies, K. D. (2003). Segmen-
Vision and Image Understanding, 98(3):491 – 512. tation of 3d medical image data sets with a combi-
nation of region-based initial segmentation and active
Jafari-Khouzani, K. (2010). Mri data set for hip-
surfaces. volume 5032, pages 1225–1232. SPIE.
pocampus segmentation by department of
diagnostic radiology at henry ford hospi- Pruessner, J., Li, L., Serles, W., Pruessner, M., Collins,
tal. http://www.radiologyresearch.org/ D., Kabani, N., Lupien, S., and Evans, A. (2000).
HippocampusSegmentationDatabase/. Volumetry of hippocampus and amygdala with high-
resolution mri and three-dimensional analysis soft-
Juottonen, K., Laakso, M. P., Partanen, K., and Soininen, H.
ware: Minimizing the discrepancies between labora-
(1999). Comparative MR Analysis of the Entorhinal
tories. Cereb. Cortex, 10(4):433–442.
Cortex and Hippocampus in Diagnosing Alzheimer
Disease. AJNR Am J Neuroradiol, 20(1):139–144. Sabuncu, M. R., Yeo, B. T., Leemput, K., Fischl, B., and
Golland, P. (2009). Supervised nonparametric im-
Leemput, K. V., Bakkour, A., Benner, T., Wiggins, G.,
age parcellation. In MICCAI 2009, pages 1075–1083,
Wald, L. L., Augustinack, J., Dickerson, B. C.,
Berlin, Heidelberg. Springer-Verlag.
Golland, P., and Fischl, B. (2009). Hippocampus,
19(6):549–557. Schmidt, R., Havas, D., Ropele, S., Enzinger, C., and
Fazekas, F. (2009). Mri in dementia. Neurologic Clin-
Liu, X. U. and Nixon, M. S. (2007). Image and volume ics, 27(1):221 – 236. Neuroimaging.
segmentation by water flow. In ISVC’07: Proceed-
ings of the 3rd international conference on Advances Wang, L., Beg, F., Ratnanather, T., Ceritoglu, C., Younes,
in visual computing, pages 62–74, Berlin, Heidelberg. L., Morris, J., Csernansky, J., and Miller, M. (2007).
Springer-Verlag. Large deformation diffeomorphism and momentum
based hippocampal shape discrimination in dementia
Lord, N., Ho, J., Vemuri, B., and Eisenschenk, S. (2007). of the alzheimer type. Medical Imaging, IEEE Trans-
Simultaneous registration and parcellation of bilateral actions on, 26(4):462 –470.
hippocampal surface pairs for local asymmetry quan-
tification. Medical Imaging, IEEE Transactions on, Yang, J., Staib, L., and Duncan, J. (2004). Neighbor-
26(4):471 –478. constrained segmentation with level set based 3-d de-
formable models. Medical Imaging, IEEE Transac-
Ltjnen, J. M., Wolz, R., Koikkalainen, J. R., Thurfjell, L., tions on, 23(8):940 –948.
Waldemar, G., Soininen, H., and Rueckert, D. (2010).
Fast and robust multi-atlas segmentation of brain mag- Yushkevich, P., Zhang, H., and Gee, J. (2006). Continuous
netic resonance images. NeuroImage, 49(3):2352 – medial representation for anatomical structures. Med-
2365. ical Imaging, IEEE Transactions on, 25(12):1547 –
1564.
McInerney, T. and Terzopoulos, D. (1999). Topology adap-
tive deformable surfaces for medical image volume
segmentation. Medical Imaging, IEEE Transactions
on, 18(10):840 –850.
Morra, J., Tu, Z., Apostolova, L., Green, A., Avedissian,
C., Madsen, S., Parikshak, N., Hua, X., Toga, A.,
Jack, C., Schufj, N., Weiner, M., and Thompson,
P. (2008a). Mapping hippocampal degeneration in
400 subjects with a novel automated segmentation ap-
proach. In Biomedical Imaging: From Nano to Macro,
2008. ISBI 2008. 5th IEEE International Symposium
on, pages 336 –339.
Morra, J., Tu, Z., Apostolova, L., Green, A., Toga, A., and
Thompson, P. (2008b). Automatic subcortical seg-
mentation using a novel contextual model. MICCAI.