BRAIN SEGMENTATION IN HEAD CT IMAGES

Ana Soﬁa Torres and Fernando C. Monteiro

Polytechnic Institute of Braganc¸a, Campus Santa Apol´onia, Apartado 1134, 5301-857 Braganc¸a, Portugal

Keywords:

Brain segmentation, Graph clustering, Head CT images, Watershed transform.

Abstract:

Brain segmentation in head computed tomography scans is essential for the development of computer-aided

diagnostic methods for identifying the brain diseases. In this paper we present a hybrid framework to brain

segmentation which joints region-based information based on watershed transform with clustering techniques.

A pre-processing step is used to reduce the spatial resolution without losing important image information. An

initial partitioning of the image into primitive regions is set by applying a rainfalling watershed algorithm on

the image gradient magnitude. This initial partition is the input to a computationally efﬁcient region segmenta-

tion process which produces the ﬁnal segmentation. We have applied our approach on several head CT images

and the results reveal the robustness and accuracy of this method.

1 INTRODUCTION

Image segmentation is one of the largest domains in

image analysis, and aims at identifying regions that

have a speciﬁc meaning within images. The role

of imaging as complementary mean of diagnosis has

been expanding beyond the techniques of visualiza-

tion and checkups in anatomical structures. This area

has become a very useful tool in planning of surgical

simulations and location of pathologies.

The Computed Tomography (CT) is an imaging

modality that allows the imaging of sections of the

human body, with almost no overlap of organs or

anatomical structures. Thus allowing us to actually

doing tests with a large number of sections quickly

and with high spatial resolution. The need for quanti-

tative analysis in tests with many sections has served

as a stimulus for the development of computational

methods for the detection, identiﬁcation and delin-

eation of anatomical structures. The segmentation of

the brain from CT scans is an important step before

the analysis of the brain. This analysis can be per-

formed by a specialist, which manually surrounds the

area of interest on each slice of the examination. This

requires very careful and attentive work and practi-

cal exams with a high number of slices, the identiﬁca-

tion of regions becomes a tedious and time consuming

task, subject to variability depending on the analyzer,

which makes it desirable to have automated methods.

However, if on one hand, manual segmentation has

the problems mentioned above, the automatic identiﬁ-

cation of structures from CT images becomes a tricky

task not only because of the volume of data associ-

ated with the imaging study, but also the complexity

and variability in the anatomical study, and that noisy

images can provide. So developing new accurate al-

gorithms with no human interaction to segment the

brain precisely is important.

The watershed algorithm is an example of a hy-

brid method, combining information about the inten-

sity and the image gradient. This algorithm is a pow-

erful edge-based method of segmentation, developed

within the framework of mathematical morphology

(Vincent and Soille, 1991; Grau et al., 2004). Some-

times, the use of the watershed over-segmentation re-

sults in unwanted regions. To circumvent this prob-

lem markers are applied to the image gradient in order

to avoid over-segmentation, thus abandoning the con-

ventional watershed algorithm (Shojaii et al., 2005).

This operation allows the reduction of regional min-

ima, grouping them in the region of interest.

The proposed methodology in this paper has three

major stages. First, from the gradient image we

create, based on the watershed transform, an over-

segmented image. The regions formed are atomic

regions. In the next step, the region similarity

graph (RSG) will be created (Monteiro and Campilho,

2008), from the over-segmented image, for apply a

graph clustering approach in the last station. This

framework integrates edges and region-based seg-

mentation with spectral based clustering through the

watershed transform. Figure 1 presents the stages of

434

Soﬁa Torres A. and C. Monteiro F..

BRAIN SEGMENTATION IN HEAD CT IMAGES.

DOI: 10.5220/0003794704340437

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 434-437

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Phases of the proposed method.

the entire process.

The combination of watershed and clustering

methods solves the weaknesses of each method.

Rather than clustering single feature points we will

cluster small regions, conﬁdent that the underlying

primitive regions are reliable. Our approach actu-

ally prefers the objects to be over-segmented into a

number of smaller regions to ensure that a minimal

amount of backgroundis connected to anyof the brain

regions.

The algorithm described in this paper can be well

classiﬁed into the category of hybrid techniques, since

it combines the edge-based, region-based and the

morphological techniques together through the spec-

tral based clustering approach. We propose that our

method can be considered as an image segmentation

framework within which existing image segmenta-

tion algorithms that produce over-segmentation may

be used in the preliminary segmentation step.

The remainder of this paper is organized as fol-

lows. Section 2 gives a description of the methods

used in this paper. Followed by the experimental re-

sults and discussion in Section 3. The concluding re-

marks are given in the last section.

2 BRAIN SEGMENTATION

The segmentation of an image is one of the most im-

portant factors in the analysis and identiﬁcation of the

brain on CT images. One of the objectives of devel-

oping new algorithms for image segmentation is to

increase the accuracy by reducing the computational

cost (Pham et al., 2000).

The watershed transform partitions the image in to

numerous regions depending on the number of local

minima of the gradient, usually the watershed tends

to produce an over-segmentation(Callaghan and Bull,

2005). In order to facilitate the calculations are over-

segmentation can be eliminated by incorporating a

pre-processing of the image. Many methods have

been proposed in order to reduce unwanted regions

and produce a meaningful segmentation.

In this work are provide some methods to over-

come this problem. For example, if the images con-

taining noise the ﬁrst step is to use non-linear ﬁlters

such as the bilateral anisotropic ﬁlter, which smooths

images while preserving its contours and structure,

since it only acts on neighbours who are part of the

same core region. The next step is to eliminate the

weaker contours through gradient minima suppres-

sion, the process known pre-ﬂooding (See Fig. 2).

This methodology uses a constant depth of certain

basin. Before to the transform each catchment basin is

ﬂooded up to a certain height above its bottom, this is,

the lowest gradient magnitude and it can be thought

as a ﬂooding of the topographic image at a certain

level (ﬂooding level). In the latest step, is made by a

segmentation fusion, merging atomic regions with a

graph-based clustering approach.

Segmentation result directly using watershed al-

gorithm is shown in Fig. 2.(a) and from it we can ﬁnd

that serious over-segmentation (1587 atomic regions)

makes the result meaningless even when we use pre-

ﬂooding as showed in Fig. 2.(b) and (c), with 1281

regions and 814 regions, respectively.

Spectral-based methods use the eigenvectors and

eigenvalues of a matrix derived from the pairwise

similarities of features (pixels or regions). This effect

is achieved by constructing a fully connected graph.

Considering all pairwise pixel relations in an image

may be too computational expensive. Unlike other

well known clustering methods (Shi and Malik, 2000)

which use down-sampling pixel-based to construct

the graph, our method is based on selecting links from

BRAIN SEGMENTATION IN HEAD CT IMAGES

435

(a) (b) (c)

Figure 2: Atomic regions from watershed. (a) No pre-ﬂooding. (b) Low level of pre-ﬂooding. (c) Medium level of pre-

ﬂooding.

a weighted undirected graph G = (V, E, W) based on a

region similarity graph where each node corresponds

to an atomic region (Monteiro and Campilho, 2008).

The proposed region similarity graph structure

takes advantage on region-based representation. The

set of nodes V are represented by the centroid of each

atomic region. The sets of links E and link weights

W represent, respectively, relationships and similarity

measures between pair of regions. Each region r

rep-

resents a small group of pixels where the centroid

is utilized as a node of the graph.

In almost all the graph-basedapproaches proposed

in the literature the spatial distance cue is also used to

compute the similarity between graph nodes. How-

ever, during our experiments, we noted that such cue

is responsible for the partition of homogeneous areas

in the image - an issue commonly associated to nor-

malized cut algorithm. Instead, we use intervening

contours (Leung and Malik, 2000) which are equiv-

alent to spatial distance without suffering from the

same problems. For each pair of nodes, node simi-

larity is inversely correlated with the maximum con-

tour energy encountered along the line connecting the

centroids of the regions. If there are strong contours

along a line connecting two centroids, these atomic

regions probably belong to different segments and

should be labeled as dissimilar.

Let i and j be two atomic regions and the orienta-

tion energy OE

∗

between them, then the intervening

contours contribution to the link weight is given by:

(i, j) = exp

−

max

t∈line(i, j)



∗

(

, x

)



(1)

where line(i, j) is the line between centroids

and x

formed by t pixels.

The mean intensity of each node contributes for

the link weight according to the following function:

(i, j) = exp

−



− I



. (2)

These cues are combined in a ﬁnal link weight

similarity function, with the values σ

and σ

selected

in order to maximize the dynamic range of W:

W(i, j) = w

(i, j) · w

(i, j) . (3)

To compute the similarity matrix the current ap-

proach uses only image brightness and magnitude

gradient. Additional features such as texture, could

be added to the similarity criterion. This may slow

the construction of the RSG but the rest of the algo-

rithm will proceed with no change.

3 EXPERIMENTAL RESULTS

The brain CT images used in this paper were pro-

viding by the database of IPB. The number of slices

varies from exam to exam, but not all slices of the

exam contain information regarding the brain, and for

that reason we used only a few images of a single scan

to his head. The images obtained in this condition are

stored with size of 512512 pixels. Each pixel is 16

bits in size and 0.85 mm resolution. The images are in

greyscale mode, stored in DICOM format. This group

of images selected includes the entire anatomy of the

brain from the top, the middle and the bottom. Figure

3 shows four slices selected from different parts of the

head to show the accuracy of this technique.

The brain segmentation results for the slices in

Fig. 3 are shown in Fig. 4. Comparing the segmented

brain regions with the original image conﬁrms that

our approach separates accurately the brain regions.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

436

(a) (b) (c) (d)

Figure 3: Original CT head images.

(a) (b) (c) (d)

Figure 4: Brain segmented images corresponding to the slices in Fig. 3, respectively.

4 CONCLUSIONS

In this paper we have proposed an image segmenta-

tion methodwhich combines edge- and region-based

information with spectral techniques through the mor-

phological algorithm of watersheds. An initial parti-

tioning of the image into primitive regions is set by

applying a watershed simulation on the image gradi-

ent magnitude. This initial partition is the input to a

computationally efﬁcient graph partition process that

produces the ﬁnal segmentation. The latter process

uses a region similarity graph representation of the

image regions.

Using small atomic regions instead of pixels leads

to a more natural image representation - the pixels

are merely the result of the digital image discretiza-

tion process and they do not occur in the real world.

Besides producing smoother segmentations than pixel

based partitioning methods, it also reduces the com-

putational cost in several orders of magnitude.

As future work it would be interesting to obtain

brain contours in the region carried out by experts, to

be subsequently developed a methodology for evalu-

ating the accuracyof brain contours resulting from the

segmentation algorithms. This assessment method-

ology can understand some criteria to overcome the

subjectivity barrier between the silhouette of the brain

performed by different specialists.

REFERENCES

Callaghan, R. O. and Bull, D. (2005). Combined

morphological-spectral unsupervised image segmen-

tation. IEEE Trans. Image Processing, 14(1):49–62.

Grau, V., Mewes, A., Alcaniz, M., Kikinis, R., and Warﬁeld,

S. (2004). Improved watershed transform for medi-

cal image segmentation using prior information. IEEE

Transactions on Medical Imaging, 23(4):447 –458.

Leung, T. and Malik, J. (2000). Contour continuity in

region-based image segmentation. In Procedings Eu-

ropean Conference in Computer Vision, volume I,

pages 544–549, Freiburg, Germany.

Monteiro, F. and Campilho, A. (2008). Watershed frame-

work to region-based image segmentation. In Pro-

ceedings of the International Conference on Pattern

Recognition, pages 1–4, Tampa, USA.

Pham, D., Xu, C., and Prince, J. (2000). A survey of cur-

rent methods in medical image segmentation. Annual

Review of Biomedical Engineering, 2:315–337.

Shi, J. and Malik, J. (2000). Normalized cuts and image

segmentation. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 22(8):888–905.

Shojaii, R., Alirezaie, J., and Babyn, P. (2005). Auto-

matic lung segmentation using watershed transform.

In Proc. of the International Conference of Image Pro-

cessing, volume II, pages 1270–1273, Genova, Italy.

Vincent, L. and Soille, P. (1991). Watersheds in digital

spaces: an efﬁcient algorithm based on immersion

simulations. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 13(6):583–589.

BRAIN SEGMENTATION IN HEAD CT IMAGES

437