Mandibular Image Segmentation on the CT-Scan of the head

using the Active Contour Method

Amillia Kartika Sari

, Riries Rulaningtyas

and Khusnul Ain

Postgraduate School, Airlangga University, Surabaya, Indonesia

Science and Technology Faculty, Airlangga University, Surabaya, Indonesia

Keywords : Image Segmentation, Active Contour Method, Mandibular, CT-Scan

Abstract : Image segmentation is one of the image processing methods with the goal of sharing the image based on

uniformity, one of which is the active contour method. This method is to detect objects on a particular

image by using curve evolution techniques, and can also overcome the deficiencies in the boundary method.

In this study image segmentation was carried out using the active contour method to evaluate the mandible

on the head CT scan. It started with a CT-Scan of the head as input data, and saved with BMP (Bitmap)

format. Then initial contour mandible, and after that the next step is image segmentation with active

contour chan-vese method. From the analysis and evaluation of 108 images of the mandible with *BMP

(Bitmap) format we get to the average accuracy values which were 99.809%, and sensitivity value of

99,806%. The conclusion of this study is that the active contour method gives accurate results of mandibular

bone segmentation on the CT scan of the head.

1 INTRODUCTION

The mandible is the bone that forms the face of a

erson, especially the lower third. Like other organs,

the mandible may develop abnormalities such as

tumors, fractures,

or dislocations. Tumor

abnormalities in the mandible may result in bone

defects. Bone defect is a state of partial or complete

loss of bone, which can cause changes in bone

function and anatomy that negatively impact by

psychological weakness and reduced confidence in

social relations

1,2

Therefore, mandibular reconstruction surgery is

recommended

immediately. O

ne thing that can be

done to optimize surgical operations is to use a 3D

prototype of the dissected organ. It aims to assess

the severity of bone defects, improve the accuracy of

marginal resection, as an implant pre-contour plate,

and can reduce surgical time

3,4

3D prototypes are the result of the printing

technology of 3-dimensional objects from

combining several materials such as plastics,

polymers, ceramics, liquids and living cells. Stages

to obtain 3D prototypes are image acquisition, image

processing, and prototype printing. For image

acquisition, data input is a digital image obtained

from radiology as a CT-scan image. In this study a

CT scan of the head is used.

After obtaining digital image data, image

processing is carried out, namely the segmentation

process. The image segmentation is the process of

dividing an image into a number of parts

. Many

methods are used in the image segmentation

processes, one of which is Active Contour. This

method uses evolutionary curve techniques to detect

objects in images

. The nature of this method is

finding the boundary or edge of the object becomes

segmented from the influence of internal energy and

external energy. Internal energy regulates continuity

while external energy functions to draw a curve to

the edge of the target

The Active Contour Method is divided into two

groups: parametric and geometric. Parametric

methods commonly known as deformable can

segment objects with a clear boundary, one of which

is the Snake Active Contour Model. While

geometric method is the method that has the ability

to segment objects with unclear boundaries, one of

which is the Active Contour Level set model. In this

study we used an active contour geometric with

Chan-Vese model.

Kartika Sari, A., Rulaningtyas, R. and Ain, K.

Mandibular Image Segmentation on the CT-Scan of the head using the Active Contour Method.

DOI: 10.5220/0007541703010305

In Proceedings of the 2nd International Conference Postgraduate School (ICPS 2018), pages 301-305

ISBN: 978-989-758-348-3

301

1.1 Active Contour Chan-Vese Model

One of the geometric active contours is the Chan-

Vese Method, in this method the process to

minimize energy is based on Region Base models

Introduced by T. Chan and L. Vese in 2001, this

model is an improvement from the edge based

model because edge detection in an image based on

curve evolution technique, Mumford-shah function

for segmentation and level set

. This is because

detection using an image gradient is less effective,

because discrete gradient is limited and the stop

function is not null at an edge and allows the curve

to cross the existing boundary.

In Chan-Vese (C-V) models, contour initiation

can be done anywhere in the image and the C-V

model will automatically detect the entire contour,

regardless of the location of contour initials

Chan-Vese active contour method assumes images

are formed from two parts. Assumed f (x, y) as a

gray-level image with the domain W. Domain W

can be divided into several parts W

along the

boundary C. If it is assumed that there are only

two parts, those in C are called W

and those

outside C are called W

. So the energy model

equation is shown in the equation 1

() () |(,) | |(,) |

C F C Ixy I dxdy Ixy I dxdy+= − + −

∫∫

(1)

In the equation above what is needed is the right

C value, so that the Pas Energy value becomes

minimal (equation 2)

as in Figure 1,

inf{() ()} 0FC FC+=

(2)

Figure 1: the C curve is outside the object if F1 (C)> 0 and

F2 (C) = 0.

If the C curve is inside the object it means F1 (C)

= 0 and F2 (C)> 0, whereas if Curve C is inside and

outside the object means F1 (C)> 0 and F2 (C)> 0.

Until finally the energy fitting will be minimized if C

= 0

For the practicality of the Fitting Energy model,

equation (2) is minimized by a regular equation such

as the length (length) of the C curve and the area in

the C curve to become equation (3)

(,) . () . ( ) |(,) |

|(,) |

s s

I C Length C Area W I x y I dxdy

I x y I dxdy

μυλ

=++−+

−

∫

(3)

With μ≥0, ʋ≥0, and λ1, λ2> 0 is a fixed

parameter. In the first term, functional energy F

controls the length of the contour that describes the

target object. The weight of μ in this equation helps

form a smooth boundary. Same with weight ʋ forces

the total area of the foreground part of W

to be

found in the segmentation process. Whereas the

third and fourth are proportions for the gray level

difference in the foreground and the background to

be detected.

The solution of the Is value can be approximated

by the Mumford-Shah segmentation function shown

in equation (4)

average( )of

average( )of W

{

I =

(4)

In the Chan-Vese active contour model, level-set

functions are introduced to reduce functional energy

F (c1, c2, C). The level-set function Φ is defined as

the zero level-set in the Lipschitz function (equation

5-7)

(){(,) :(,)0},Inside W x y W x y=∈Φ<

(5)

( ) {(,) : (,) 0},Inside W x y W x y=∈Φ>

(6)

{(,) : (,) 0}CxyWxy=∈Φ=

(7)

According to Crandall (2009), the completion of

the Fitting Energy function is changed to a level-set

function Φ using the Euler-Langgrange equation so

that the iteration calculation can be done by solving

equation (8)

() ( )[ ( ,( , ) )

(,(, ) )

td di Ixyc

Ixy c

φφ

φδφμυυλ

⎛⎞

∇

== −− −+

⎜⎟

∇

⎝⎠

−

(8)

To solve the equation (2.8) it must be discredited

first, e.g.

denoted as the value of the level-set

function Φ in the pixel (i, j) in the iteration n

(equation (9))

ICPS 2018 - 2nd International Conference Postgraduate School

302

,1,, ,,1,

,,1, ,,,1

xn n n xn n n

ij i j ij ij ij i j

yn n n yn n n

ij ij ij ij ij ij

φφ φ φφφ

++ − −

++− −

Δ= −Δ=−

(9)

2 MATERIALS AND METHOD

A computer with AMD Turion (tm II) Dual Core

2.30 Gz / 2.0 GB RAM and the Matlab application

(R2017a, MathWorks Inc., USA) was used in this

study for simulation and analysis data. The image

used is the image of the CT scan head with a

512x512 matrix size from the Brilliance CT 64

Channel (Philips medical system, Netherlands). The

parameters used on the CT scan head are slice

thickness of 2 mm, DFOV 360 mm, a voltage of 120

kVp and a tube current of 297 mA.

This study stage begins by collecting CT Scan

images, then converts them to files with BMP

extension (Bitmap), the data is called input data. The

next stage is image processing where normalization

will be carried out by converting input data into gray

scale images, with the aim of achieving uniformity

of pixel intensity. And the last stage is the process

of segmentation with the chan-vese model active

contour. Image results from the segmentation

process will be tested with images that are manually

segmented by experts (doctors). A flowchart of this

study can be seen in Figure 2.

Figure 2: Flowchart of the study.

.In this study the work system is divided into

several stages, namely input system, image

processing and segmentation process.

2.1 Input Data

In the input system, CT scan data is obtained from

scanning the head with the upper boundary of the

vertex and lower limit of the base of the cranial

using the MDCT 64 Slice modality (Figure 3). Data

formats from CT scan have extensions *DICOM and

are then converted into a file with *BMP format

with a matrix size of 512 x 512. The process of

converting *DICOM format to *BMP is done

through the Radiant Dicom application program.

Figure 3: (a) Input data of Mandible CT, (b) Image data after

normalization.

2.2 Image Processing

After obtaining an image with BMP file format, the

next process is image normalization. This process

aims to change the pixel intensity of values. The

initial pixel intensity value ranges from 22-255 to 0-

255 (Figure 3).

The result of normalization is an image that has a

gray level with an intensity range of 0-255. Image

segmentation is the next step.

2.3 Segmentation Process

The segmentation system stages using Chan-Vese

active contours are as follows:

1. Determine the control parameter initialization,

weights or constants of both internal energy and

external energy.

2. Determine the edge image as external energy by

determining the image gradient.

3. The next step is to determine the initial contour,

adjust the initial contour according to the original

image (Figure 4a-4b).

4. And determine total energy value.

5. If the total energy has reached the minimum, the

iteration will stop and will form a contour

according to the input data (Figure 4c).

6. Segmentation results are output data.

Mandibular Image Segmentation on the CT-Scan of the head using the Active Contour Method

303

Figure 4: (a) select initial contour, (b) result initial

contour, (c) result post segmentation process.

3 RESULT

The image data used in this segmentation process is

108 images. The analysis used in this study is by

comparing program segmentation results with results

of expert segmentation (doctor). From this

comparison value will be obtained, TP: True

Positive, FN: False Negative, TN: True Negative,

and FP: Wrong Positive, which is calculated based

on the number of pixels (Figure 5). Based on this

analysis we can measure the value of sensitivity and

accuracy with the equation (9-10) (Table 1).

Accuracy=

()

100%

TP TN

Acc x

TP TN FP FN

+++

(9)

Sensitivity:

()

100%

TPR x

TP FN

(10)

Figure 5. Analysis method of segmentation results data

Table 1: Result of accuracy and sensitivity data

From the above calculations it can be seen that

with image segmentation using the Chan-Vese

active contour method, the average accuracy value is

99.809% and the sensitivity value is 96.806%.

4 DISCUSSION

Image segmentation method is proposed using the

method of active contour, as active contours always

provide sustainable sub-region boundaries, so as to

increase the final results of the image analysis.

Mathematical implementation of the proposed active

contour models is achieved using the level set

method. The proposed image segmentation of this

method allows us to increase the detection of

mandibular bone in the image of the CT scan of the

head.

In this study the segmentation program has not

been able to distinguish solid and hollow bone tissue

considering the bone structure is not all solid. And

as this study is still using only one patient, it is

expected that further research can use more data.

5 CONCLUSION

Referring to the level of accuracy segmentation

results, the results of this study are quite accurate

and can be used as an alternative method of

mandibular bone segmentation on CT scan images to

be used as data input in the subsequent process of

making a 3-dimensional mandibular prototype for

mandibular reconstructive surgery.

ICPS 2018 - 2nd International Conference Postgraduate School

304

REFERENCES

Szplaski C, Wetterau M, Barr J, Saadeh P, Warren S.

Cranial bone defect: Current and Future Strategies.

Neurosurg Focus. 2010;29:1-11.

doi:10.3171/2010.9.FOCUS10201

Aydin S, Kucukyuruk B, Sanuz G. Cranioplasty: Review

of Material and Techniques. Neurosci Rural Pr.

2011;2:162-167. doi:10.4103/0976-3147.83586

Oliveira GDE, Silva DN, Meurer E, Ine M. Dimensional

error in selective laser sintering and 3D-printing of

models for craniomaxillary anatomy reconstruction *.

2008:443-449. doi:10.1016/j.jcms.2008.04.003

Yang L, Shang X, Fan J, et al. Application of 3D Printing

in the Surgical Planning of Trimalleolar Fracture and

Doctor-Patient Communication. 2016;2016.

Gonzalez R., Woods R. Digital Image Processing Using

MATLAB. (Prentice-Hall P, ed.). New Jersey,USA:

Pearson Education, Inc.; 2004.

Airouche M, Bentabet L, Zelmat M. Image Segmentation

Using Active Contour Model and Level Set Method

Applied to Detect Oil Spills. Proc World Congr Eng.

2009;1(1):1–3.

Bakos. Active Contour and Their Utilization at Image

Segmentation Technical. Slovakian-Hungarian Jt

Symp Appl Mach Intell ang Inform. 2007:313-317.

Apriliani D, Murinto. Analisis Perbandingan Teknik

Segmentasi Citra Digital Menggunakan Metode Level

Set. 2013;7(2):802-810.

Cohen R. The Chan-Vese Algorithm. 2010:1-18.

Li S, Z. Markov Random Field Modelling in Image

Analysis. Springer. 2009.

Sohn K. Segmentation of Lung Fields Using Chan-Vese

Active Contour Model in Chest Radiographs. In: SPIE

7963.

Chan TF, Vese LA. Active Contours Without Edges.

2001;10(2):266-277.

Crandall R. Image Segmentaion Using the Chan-Vese

Algorithm. In: ; 2009.

Mandibular Image Segmentation on the CT-Scan of the head using the Active Contour Method

305