The segmentation of
hippocampus is significant and accurate in detection of Alzheimer’s disease
using structural Magnetic Resonance Images. To segment the hippocampus from
magnetic resonance images a method is proposed by combining the region growing
and level set method. The first step is to segment the hippocampus using region
growing from the seed point. The second step is to fine tune the hippocampus
segmented images using level set method. To overcome the drawbacks of potential
leakage in region growing method for hippocampus segmentation. Quantative
analysis shows the correlation between proposed segmentation and human expert
segmentation. These two tools help to determine the volume of hippocampus as
well as to determine the memory loss in Alzheimer’s disease.

Best services for writing your paper according to Trustpilot

Premium Partner
From $18.00 per page
4,8 / 5
Writers Experience
Recommended Service
From $13.90 per page
4,6 / 5
Writers Experience
From $20.00 per page
4,5 / 5
Writers Experience
* All Partners were chosen among 50+ writing services by our Customer Satisfaction Team

Keywords: Hippocampus,
Level set method, Region growing, Segmentation, Magnetic Resonance Images.


The hippocampus is a
part of middle arc of the limbic system. It is located in the medial temporal
lobe inferior to the chorodial fissure and temporal horn. The gray matter of
the hippocampus is an extension of subiculum of the parahippocampal gyrus. The
hippocampus consist of two interlocking C-shaped structures are the cornus of
Ammonis and the dentate gyrus 1. The atrophy of hippocampus is one of the
symptoms of early stages of Alzheimer’s disease 2. High resolution
T1-weighted MRI allows precise assessment of these structures. In AD the
hippocampus volume is smaller than the healthy tissue and is associated with
greater severity of dementia 3. Accurate segmentation of hippocampus is
important for early diagnosis of Alzheimer’s disease patients. 4In literature
several approaches for automatic and semi automatic segmentation of hippocampus
have been proposed. Fully automatic segmentation often fails if the hippocampus
is relatively small and shape of object is highly varied from the process 5.
Compared to this semi automatic may provide a more naturalistic approach
because of human expertise and automatic techniques 6. In 7 a model which
contains the triangulated surface and the corresponding grey scale volume is
used. This requires the pre segmentation of objects of interest for
reconstructing the surface. In 8 a deformable model is generated from single
manually labeled volume to measure the size and shapes of hippocampus
segmentation. In this the disadvantage is manual segmentation task of
hippocampus and a low resolution of the model. In 9 a hierarchical algorithm
is used to segment neuro anatomical structures of the grey matter, such as
hippocampus. The algorithm is guided within a tree structure. In 10 fully
automatic method proposed using probabilistic and anatomical priors for
hippocampus segmentation. In this paper a new hippocampus segmentation scheme
is proposed. This involves the segmentation of hippocampus by region growing
and fine tuning the hippocampus by level set.


Data used in the
preparation of this paper are obtained from the department of Radiology at Advanced
KG Hospital, Madurai. The data set contains 30 brain T1 weighted MRI volumes
with hippocampus labels. Images were acquired from two different MRI units with
different field strengths. They have different resolution and contrasts.
Fifteen of the images have a higher resolution and were acquired with a
relatively new 3T MRI system. Ten of the images belong to patient’s temporal
lobe epilepsy that may have atrophic hippocampus. This makes the segmentation
challenging and evaluates the segmentation algorithms against practical
problems. The remaining 5 images belong to healthy subjects.

Seed point selection

To make the proposed
the system automatic the seed point using well known procedure adopted in a
versatile algorithm for the automatic segmentation of hippocampus based on
level set 11. This involves the pre processing of MRI using wave 12 atom
shrinkage and identification of lateral ventricle. Then approximate ROI is
selected based on the location of LV and the anatomical knowledge of
hippocampus. In coronal slice, from the superior portion of LV at 50mm
horizontal line is drawn towards cortex and 70mm vertical line is drawn towards
inferior portion of the slice. The rectangular box based on these two lines is
drawn and this is taken as approximate ROI. Then the bounding box is selected
with the dimension of 1.5cm size within the ROI. Initial bounding box is taken
at the extreme top left of ROI and subsequently the bounding box is moved from
left to right and top to bottom. The histogram of the bounding box is defined
to enable calculation of local statistics necessary for extraction of the
corresponding hippocampus region. Histograms have substantial local variations
that hamper determination of global peaks and valleys. Some smoothening has to
be applied to reduce the variations. To identify the seed point, the smoothed
histogram should be modeled by Gaussian. The peak of the third modeled curve is
taken as the seed point to segment the hippocampus.

Segmentation by region growing

Region growing is
simple and fast segmentation method to partition different region of an image.
It involves the selection of seed points and a cost function. The seed point
selection is explained in the previous section. The region growing starts from
the seed point to the adjacent pixels according to the image field such as grey
scale image. The newly grown points should satisfy the pre-determined cost
functions. For example, the grey level intensity value of the newly grown
points should be above or below some threshold. The threshold could be
determined by choosing the approximate area of hippocampus followed by taking
the histogram. Since middle peak belongs to grey matter hippocampus, the
intersection points between the middle and nearby peaks are taken as threshold.

The steps of region
growing could be summed up as follows:

1. Select a seed point
in the region of interest in the image.

2. Add adjacent
neighbor points of existing points of the region to potential region list.

3. Test every point in
the potential region list to see if it satisfies the cost function. Add the
points to the region of interest if they satisfy the cost function.

4. Repeat the steps
from step 2 until all neighbors are searched.

The segmented area
using region growing is given in figure 1.


Figure 1.Segmented area
using region growing in left hippocampus grown from seed. And right hippocampus
region grown from seed.

2.3. Fine tuning of
hippocampus by level set

The complex appearance
of regions surrounding the hippocampus and the low image contrast in the
hippocampus region helps the region growing method in giving the approximated
hippocampus region. To fine tune the hippocampus, level set is applied on the
approximate segmented hippocampus region. Level set methods can be formulated
as the zero level set {(x, y) | ? (x, y, t) = 0} of a time dependant function ?(x,
y, t) that evolves according to the following equations (2.1),

?? / ?t + F | ? ? | = 0                        (2.1)

Where F is the speed
function which depends on the image data and the level set function ?. In the
traditional level set methods the shape modeling and front propagation is used
for level set function can develop steep or flat shapes during its evolution,
which makes further computation inaccurate. To avoid this problem the function ?
is initialized as a signed distance function. Then it is reinitialized during
the evolution by solving the following equation (2.2),

?? / ?t = sign (?0)
(1 – | ? ? |)        (2.2)

Where ? is the
functions that is reinitialized and sign(?) is the sign function. Figure 2
shows the segmented hippocampus using level set method (marked in red color).

Figure 2. Segmented
hippocampus using level set in normal case and Alzheimer’s case.


Two groups of T1
weighted MR brain images from a total of 5 subjects were used to test the
proposed method. The first group consists of three data sets belonging to
patients with temporal lobe epilepsy, who may have atrophic hippocampus. The
second group consists of two data sets belonging to healthy subjects.

For validation the left
hippocampus of cone coronal slice is chosen from each dataset of voxel size
0.781 × 0.781 × 2 mm3. To evaluate the segmentation accuracy,
automated results were compared with manual labeling available on data set is
tested and the coefficient of similarity ? and spatial overlap ? are computed
using equations (3.1) and (3.2) 13.

manual – proposed

        ? = 1 –   ————————      (3.1)




          2 ×

            ?      =     
——————–     (3.2)

manual + proposed


Table A1 lists the hippocampus
segmentation results from expert segmentation and the proposed segmentation
runs for 5 test data sets, 3 of which were acquired from individuals with
temporal lobe epilepsy, who may have atrophic hippocampus. Voxel size in all
datasets is 0.781 × 0.781 × 2 mm3.

The mean and the standard deviation of
the coefficient of similarity between the expert and the proposed method for
the hippocampus extractions are 0.9400 and 0.8932 respectively for the 5 tested
datasets listed in table A1.

Spatial overlap is more sensitive to
small unmatched segmentation errors and is a more accurate measure of agreement
than coefficient of similarity, because the approach takes into account the
spatial properties of segmented regions derived by the two methods. The spatial
overlap measure is also more sensitive to difference between methods, since
both denominator and numerator change with increasing or decreasing overlap.
For 5 cases the best overlap metric is 0.9065 and the worst is 0.8429. The mean
and standard deviation of overlap for 5 cases are 0.4442 and 0.0088, respectively.

Visual inspection of the hippocampus
segmentation results are performed by comparing composite images of the
intersection of the extracted hippocampus with the underlying MRI brain and
determining to what degree the extracted boundaries represent the true
boundaries of the hippocampus.

































To understand the reason for the
differences between expert and the proposed hippocampus segmentation methods,
we created images showing the visual overlap of the two extractions
superimposed upon the underlying MRI brain volume. Figure 3 shows one such
composite image of the intersection of the extracted hippocampus ROIs from the
coronally oriented images. Red voxels in hippocampus represent the spatial
intersection of the expert and algorthim; Yellow voxels represent hippocampus
missed by the algorithm (false negative); green voxels represent the voxels
mistakenly classified as hippocampus as determined from the expert tracing
(false positive). It is noticeable that the highest segmentation differences
occur at the boundary of ROIs, where the boundaries are not clear in MRI. On
the basis of spatial overlap, the automatic algorithm extraction is shown to be
highly comparable to the manual extraction method.