Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of King Saud University –

Computer and Information Sciences
journal homepage: www.sciencedirect.com

Automatic Brain Tumor Segmentation from MRI using Greedy Snake

Model and Fuzzy C-Means Optimization
C. Jaspin Jeba Sheela a,⇑, G. Suganthi b
a
Reg.No 17221282162010, St. Xaviers Autonomous College, Palayamkotai affiliated to Manonmaniam Sunadaranar University, Abishekapatti, Tirunelveli 627012, Tamil Nadu, India
b
Department of Computer Science, W.C College, Nagercoil affiliated to Manonmaniam Sunadaranar University, Abishekapatti, Tirunelveli 627012, Tamil Nadu, India

a r t i c l e i n f o a b s t r a c t

Article history: The automatic brain tumor segmentation in MRI (Magnetic Resonance Images) is becoming a challenging
Received 9 December 2018 task in the field of medicine, since the brain tumor occurs in different shapes, intensities and sizes. This
Revised 4 April 2019 paper proposes an efficient automatic brain tumor segmentation using Greedy Snake Model and Fuzzy C-
Accepted 10 April 2019
Means optimization. This method initially identifies the approximate Region Of Interest (ROI), by remov-
Available online xxxx
ing the non-tumor part by two level morphological reconstruction such as dilation and erosion. A mask is
formed by thresholding the reconstructed image and is eroded to improve the accuracy of segmentation
Keywords:
in Greedy Snake algorithm. Using the mask boundary as initial contour of the snake, the greedy snake
MRI images
Image segmentation
model estimates the new boundaries of tumor. These boundaries are accurate in regions where there
Fuzzy C-means algorithm is sharp edge and are less accurate where there are ramp edges. The inaccurate boundaries are further
Greedy snake model optimized by using Fuzzy C-Means algorithm to obtain the accurate segmentation output. The region that
has large perimeter is finally chosen, to eliminate the in- accurate segmented regions. The experimental
verification were done on T1-weighted contrast-enhanced image data set, using the metrics such as dice
score, specificity, sensitivity and Hausdorff distance. The proposed method outperforms when compared
with the traditional brain tumor segmentation methods in MRI images.
Ó 2019 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction Benign does not contain cancer cells which can spread to the adja-
cent tissues. Malignant is a type of brain tumor that contains can-
The uncontrolled growth of cells which can destroy healthy cer cell which easily spreads to the adjacent tissues.
cells in brain is termed as brain tumor. Magnetic Resonance Imag- The brain tumor segmentation algorithm can be grouped into
ing (MRI) (Bauer et al., 2011, 2013) is an imaging technique that four classes (i) Threshold based segmentation (ii) Edge based seg-
provides useful information about brain, since it provides high con- mentation (iii) Region based segmentation and (iv) Clustering
trast between soft tissues. In the past few decades the death caused based segmentation. Threshold based segmentation (Dass and
by brain tumor is increasing. Early detection and accurate segmen- Devi, 2012) partition the image into different region based on the
tation (Bengio, 2012) of brain tumor can reduce the death rate intensity of the pixel present in the image. The thresholding algo-
caused by brain tumor. Manual diagnosis for segmenting the brain rithm can be sub-divided as global thresholding and local thresh-
tumor can cause incorrect segmentation results. Therefore, there is olding. Global thresholding uses a single threshold value
an immediate need for an efficient (computer aided diagnosis) throughout the image. This method is simple if the intensity of
automatic brain tumor segmentation algorithm. Brain tumor may the image is uniform and there is a high contrast between fore-
be benign or malignant (Clark et al., 1998; Cobzas et al., 2007). ground and background regions. Otsu’s threshold (Selvaraj and
Dhanasekaran, 2013) sub-divides the image into two classes of
intensity. Local thresholding uses different threshold values for
⇑ Corresponding author.
each part of the image where the image is divided into sub-
E-mail address: [email protected] (C. Jaspin Jeba Sheela).
images. Global thresholding is faster than local thresholding. Based
Peer review under responsibility of King Saud University.
on the abrupt change in intensity of pixel near the edges, the edges
segmentation algorithm (Sujji et al., 2013) sub-divided the image
into foreground and background. The gradient based segmentation
(Aswathy et al., 2014) uses the difference obtained between the
Production and hosting by Elsevier

https://doi.org/10.1016/j.jksuci.2019.04.006
1319-1578/Ó 2019 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
2 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

particular pixel and its neighbour pixel. The edge segmentation segmentation results. The segmentation methods such as thresh-
method uses the operators such as Sobel operator, Laplacian oper- olding, edge based segmentation, region based segmentation are
ator, Canny operator and Prewits operators etc. also commonly used. The region based segmentation methods
In region based segmentation (Kandwal et al., 2014), the seg- includes region growing, region splitting and merging, watershed
mentation is done on the basis of certain criteria. This region based segmentation and active contour models. In active contour model
segmentation can be classified as region growing, region splitting or snake model, the snake is initially placed near the contour of
and merging and watershed segmentation methods. The region region of Interest. The snake moves due to internal and external
growing approach (Angel Viji and Jayakumari, 2013) initially forces and it stops moving when the energy function is minimum.
chooses a seed pixel and region begins to grow based on the inten- Ivan Cabria et al., proposed a Potential Field Segmentation (PFS)
sity of the region and the neighbouring pixel of the region. The where it uses an ensemble approaches that fuse the segmentation
growth of the region stops if the growth condition fails this method results of PFS. In this approach (Cabria and Gondra, 2017) the
can be applied in segmentation in small regions such as lesions and intensity of pixel is viewed as the mass, and the potential field is
tumour. In region splitting and merging method the image is sub- estimated for each pixel. If the potential field is less than adaptive
divided into different region until no further splitting is possible potential threshold, the corresponding pixel is viewed as the tumor
and it is merged based on certain criterion. Watershed segmenta- pixel.
tion can be applied in segmenting the regions that has uniform Mohammad Havaei et al. proposed an brain tumor segmenta-
contrast distribution. The clustering approach (Shen et al., 2005) tion using Deep Neural Networks. In this they proposed a Convolu-
groups the pixel into different classes and it does not require any tional Neural Network that was of both global contextual features
training. The pixel having largest probability are grouped in the and local features simultaneously (Havaei et al., 2017). The final
same category. layer of the Neural Network is the conventional implementation
The commonly used clustering approaches are (i) K-means clus- of a fully connected layer which uses a two-phase training system.
tering (ii) Fuzzy c-means clustering (iii) Hierarchal clustering. The Elisee Ilunga-Mbuyamba et al. proposed a Localized Active Contour
K-means clustering algorithm cluster the pixels into different clus- Model (LACM) for brain tumor segmentation. This method auto-
ter based on the properties of pixels. K-means performs hard clus- matically balances the mean intensity distance between regions
tering, where each pixel has a membership function corresponding of interest and background (Ilunga-Mbuyamba et al., 2017). This
to a particular cluster. Fuzzy- C-Means (FCM) (Ahmed et al., 2002) aims to reduce the attraction of active contour in the undesired
clustering algorithm assigns a membership function to each pixel boundaries. Moreover it also uses Hierarchical Centroid Shape
of the image corresponding to centre of the cluster. The pixels that Descriptor (HCSD) to detect the tumor region.
are nearer to the centre of the cluster have higher membership Morales et al. (2017) proposed a computer-aided segmentation
towards the corresponding cluster. The hierarchal clustering tech- which computes brain segmentation and volumetric analyses.
nique groups the pixels of the image into a three of cluster and it Essadike et al. (2018) proposed a segmentation method, that seg-
does not need to specify the number of cluster before the cluster- ments the abnormal tissue region. The initial active contour is
ing starts. detected using optical correlation. The active contour model move
The brain tumor can be accurately segmented usually in three towards the actual tumor borders. Angulakshmi and Lakshmi Priya
steps as follows, (2017) proposed this segmentation method based on spectral clus-
tering which identifies the tumor region using super pixels. The
i) Pre-processing the MRI image tumor tissues are segmented from the non-tumor tissues based
ii) Detecting the Region Of Interest (ROI) on spectral clustering.
iii) Optimizing the segmented results Umit Ilhan et al. developed a segmentation method that clearly
distinguish the tissues affected by cancer. This method makes the
The brain tumor image contains unwanted distortions and arti- segmented tumour region more clear for the medical practitioner
facts such as noise. Such distortions and artifacts must be removed so that it will be easy for him/her during diagnosis. It uses a collec-
before the actual segmentation starts to achieve good segmenta- tion of technique such as morphological operation, pixel subtrac-
tion results. Normally, Median Filter is used to filter such noises. tion, threshold based segmentation and image filtering
The Median Filter replaces a pixel value by the median intensity technique. Guotai Wang et al. proposed an MRI brain tumor seg-
of immediate neighborhood pixels which preserves the edges of fil- mentation using deep learning with image specific fine tuning. This
tered MRI image. After pre-processing the image, the ROI is method uses deep learning based interactive segmentation with
approximately segmented and optimized using any suitable seg- Convolutional Neural Networks (CNNs). The images specific fine-
mentation method. tuning may be supervised or unsupervised. The supervised
The remaining part of the paper is organized as follows, section required additional scribbles while the unsupervised method does
II presents the proposed automatic brain tumor segmentation not required additional user interaction. The convolutional neural
method. Section III discusses the experimental results of the pro- networks can be designed by using the weighted loss function
posed segmentation method. Finally section IV concludes the and fine-tuning can be done using interaction based uncertainty.
paper. This method can be applied for segmenting the multiple organs
from 2D fetal magnetic resonance images. The 3D brain tumor seg-
mentation can be done using this method in two ways ie. Segmen-
2. Related works tation of brain tumor core and whole brain tumour segmentation.
The tumor core segmentation exclude the edema while the whole
Various segmentation methods have been proposed such as brain tumour segmentation includes the edema region. The next
Gaussian mixture model (Wu and Corso, 2012), K-Means (Paul section shows the proposed MRI brain tumor segmentation algo-
and Bandhyopadhyay, 2012) and Fuzzy C-Means (Aparajeeta rithm (See Fig. 1).
et al., 2016).
The Gaussian mixture model (Wu and Corso, 2012) filters by 3. Proposed MRI segmentation method
estimating the maximum likely-hood parameters. In k-means
algorithm, seed point selection is a major factor for image segmen- The proposed automatic brain tumor segmentation in MRI
tation initially. The FCM (Aparajeeta et al., 2016) provides accurate images can be done in five steps,

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx 3

Fig 1. Block diagram of Proposed brain tumor segmentation method.

i) Pre-processing region. Here the background is eliminated by using two levels of

ii) Snake Contour detection morphological operation. The first level morphological reconstruc-
iii) Segmentation using greedy snake algorithm tion is by erosion while the second level morphological reconstruc-
iv) Optimization of segmented results tion is by dilation. Let I2 ðx; yÞ be the median filtered image which is
v) Selecting the accurate segmented tumor region used as a mask image. For simplicity we represent the mask image
I2 ðx; yÞ as ‘X’.
The block diagram of the proposed brain tumor segmentation The morphology reconstruction by erosion of a mask image ‘X’
method is shown in figure. from a marker image ‘M’, which is represented by reX ðMÞ.
ðn Þ
reX ðMÞ ¼ eX ðM Þ ð3:2Þ
3.1. Pre-processing
ðnÞ ðnþ1Þ
With ‘n’ so that eX ðMÞ ¼ eX ðM Þ
The MRI images usually consist of distortion and artifacts. For Let S1 ðx; yÞ be the output of first level morphological recon-
computer aided segmentation the distortion and artifacts must structed image. For simplicity of equations, we represent the level
be removed before segmentation to obtain accurate segmentation 1 morphological reconstructed image S1 ðx; yÞ as Y. The morpholog-
output. It is necessary to maintain the essential feature while ical reconstruction by dilation of a mask image ‘Y’ from a marker
removing such artifacts. Median filtering is the most commonly image ‘M’ is represented as r dY ðM Þ.
used filtering technique that filters the distortion and artifacts
ðn Þ
without the loss of essential features such as edge information. rdY ðMÞ ¼ dY ðMÞ ð3:3Þ
Median filter replaces the pixel value by the median intensity of ðnÞ ðnþ1Þ
the nearest neighborhood pixels. Let I1 ðx; yÞ be the original MRI With ‘n’ so that dY ð M Þ
¼ dY ðM Þ.
The image dilation uses a
image and I2 ðx; yÞ the median filtered image or the Pre-processed 3  3 structuring element which is represented as structuring ele-
2 3
output represented by. 0 1 0
ment 1 1 1 5. Let S2 ðx; yÞ be the level 2 morphological recon-
4
I2 ðx; yÞ ¼ medianfI1 ði; jÞ; ði; jÞ 2 wg ð3:1Þ 0 1 0
structed image, which is the background minimized image.
where w represents a 3  3 neighbourhood centered around the
location ðx; yÞ in the image. 3.3. Thresholding and dilation

3.2. Snake contour detection The background minimized output minimizes the background
information. The output of background minimized image S2 ðx; yÞ
The initial snake contour can be detected in two steps, is segmented using thresholding method where the threshold
value is the maximum intensity presents in the background elim-
(i) Background minimization inated image. Therefore the pixels whose intensity equals the max-
imum intensity are converted as foreground and the pixels whose
(ii) Thresholding and Dilation intensity other than the maximum intensity is set as background.
Let S3 ðx; yÞ be the thresholding output represented as
3.2.1. Background minimization 
1 if S2 ðx; yÞ ¼ maxðS2 ðx; yÞÞ
Background Minimization aims to minimize the non-tumorous S3 ðx; yÞ ¼ ð3:4Þ
0 otherwise
region information while maintaining the features of tumors

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
4 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

The internal energy can be represented as,

aðsÞjv 0 ðsÞj2 bjv 00 ðsÞj2

Ein ¼ þ ð3:11Þ
2 2
Here the first term makes the snake to act like membrane.
Higher value of aðsÞ increases the internal energy as it stretches.
The image energy is the combination of energy of edge, line and
Fig 2. Greedy snake segmentation starting point without and with dilation (a)
terminations.
Without dilation (b) Greedy snake starting point without dilation (c) Dilation Eim ¼ we Ee þ wl El þ wt Et ð3:12Þ
output (d) Greedy snake starting point with dilation.
where we , wl and wt are he weights of edge, line and termination.
Energy in edges can be found out using gradient. The snake will
The boundaries of this segmented image can be used as the ini- be attracted towards the contours with large image gradients. If a
tial snake contour for the greedy snake algorithm. In order to particular edge has low edge energy, the snake will try to move the
increase the segmentation accuracy in greedy snake algorithm, contour in the edge which has higher edge energy. Based on the
the boundaries of the segmented image is enlarged by dilating sign of wl , the snake will be attracted towards white lines or dark
the segmented image. Fig. 2 shows an example of choosing the ini- lines. The constraint energy ðEcon Þ will move the snake away or
tial contour for the greedy snake algorithm without and with dila- towards from certain features of the region. Let I3 ðx; yÞ be the
tion. If the initial snake contour is chosen as Fig. 2(b) without greedy snake segmented image which is applied to the input of
dilatingthe segmented image as shown in Fig. 2(a), then the greedy fuzzy C-means clustering algorithm.
snake output may miss some of the tumor boundaries, since the
initial contour lies inside the tumor region. Therefore dilation as 3.5. Fuzzy -C-means clustering optimization
Fig. 2(c) makes the initial contour lies outside the tumor region
as shown in Fig. 2(d). This dilation process will increase the bound- Let I3 ðx; yÞ be the fuzzy c-means clustering input which is seg-
ary of the segmented image, which further increases the accuracy mented to obtain the optimized image I4 ðx; yÞ. Fuzzy- C- means
of segmentation in greedy snake algorithm. Let S4 ðx; yÞ be the (FCM) algorithm partition a set of data ‘X’ from the image I3 ðx; yÞ
dilated output for thresholded image S3 ðx; yÞ, where the image into two or more clusters. The regions (pixels) on boundaries does
dilation uses a 3  3 structuring element which is represented as
2 3 not belong to a particular classe (foreground or background). Such
0 1 0 pixels have membership degrees between ‘00 and ‘10 , which can be
structuring element 4 1 1 1 5.
a member of background or foreground. The FCM algorithm was
0 1 0
initially proposed by Dunn and it was extended by Bezdek. Let
the data set of image pixel be,
3.4. Greedy snake algorithm
X ¼ fX 1 ; X 2 ; . . . . . . X N g ð3:13Þ
Snakes or active contour is a curve that moves within a image,
The data ‘X’ contains ‘N’ number of elements. The number of
where there is minimum energy. The snake moves and tries to find
cluster is represented as ‘c’ which is greater than ‘10 . ie 2  c < N.
the boundaries of a region of interest under the influence of exter-
Let M ij be the degree of membership of X i in any cluster ‘j’. Let Rj
nal and internal forces. The initial contour of the snake can be esti-
be the jth cluster center. Let ‘s’ be the termination value which is
mated from the boundary of dilated image S4 ðx; yÞ: The image
a constant whose value lies between ‘00 and ‘10 . Let ‘t’ be the num-
I2 ðx; yÞ is segmented using the initial contour estimated from the
ber of iteration and ‘m’ be the fuzziness index, which is a real num-
boundary of the region in dilated image S4 ðx; yÞ to obtain the
ber greater than ‘10 . The FCM algorithm work as follows,
greedy snake segmented image I3 ðx; yÞ
The image forces the snake contour to move towards the image  
1. Initialization the membership matrix M ¼ M ij . Let it be M ð0Þ .
features such as edges and lines. Let Ein be the internal energy and
Eex represents the external snake force. Therefore the total energy 2. The cluster center matrix at any step ‘t’ can be calculated as
 
of the snake is represented as RðtÞ ¼ Rj .
Z E PN
ðv ðsÞÞds
m
E¼ ð3:5Þ j¼1 M ij X i
snake Rj ¼ PN m
ð3:14Þ
i¼1 M ij
Z E
E¼ ðv ðsÞÞ þ Eex ðv ðsÞÞds ð3:6Þ 3. Update M ðtÞ ,M ðtþ1Þ as,
in

The external force Eex is the sum of the image force Eim and 1
M ij ¼ ð3:15Þ
external constraint force is Eco . Pc Dij ðm1
2
Þ
t¼1 Dtj
Eex ¼ Eim þ Eco ð3:7Þ
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Z u N
½Ein ðv ðsÞÞþEim ðv ðsÞÞþEco ðv ðsÞÞ uX
E¼ ds ð3:8Þ Dij ¼ t ðX i  ci Þ ð3:16Þ
i¼1

where v ðsÞ is the parametric representation of the curve. 4. Stop the iteration if the absolute difference between the mem-
v ðsÞ ¼ ½xðsÞ; yðsÞ ð3:9Þ bership of any two successive iteration is less than the termina-
tion value s.
The internal energy of the snake is the sum of the energy of the
snake contour ðEcr Þ and the energy of spline curvature ðEcur Þ.
M ðtþ1Þ  M ðtÞ < s ð3:17Þ
Ein ¼ Ecr þ Ecur ð3:10Þ
Else repeat step 2 and 3 until Eq. (3.17) satisfies.

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx 5

The FCM clustering works by assigning membership to each 0.7 respectively. The maximum number of iteration t, termination
data based on the distance between the data and the center of value s and fuzziness index m in FCM algorithm are set as 100,
the cluster. If the distance between the center of a cluster and 0.001 and 2 respectively.
the data point is less, the data point have more membership with The size of each image is 512  512 which is initially filtered
that cluster. Since the proposed method aims to segment the image using a median filter having a mask size of 3  3, before actual seg-
into two classes (foreground and background), the number of clus- mentation. Fig. 3 shows the output images obtained during pre-
ters cis set to 2. The FCM algorithm uses the pixels that are processing, background elimination (level 1 and level 2 morpho-
enclosed in the region segmented by Greedy snake segmentation logical reconstruction), thresholding, dilation, greedy snake seg-
algorithm. The edges of the Greedy snake segmentation algorithm mentation and FCM optimization. Here we have measured the
are accurate since the edge pixels may or may not belong to the metrics such as specificity, sensitivity, and dice-score using Eqs.
tumor port. The FCM algorithm classifies the complete pixels (4.1), (4.2) and (4.3) respectively. Specificity also called as true neg-
enclosed inside the region segmented by Greedy snake algorithm ative rate that measures the actual negatives that are correctly
into two clusters, where one cluster contains the pixels of the identified as such. Sensitivity, specificity and Dice score are the
tumor tissues while the other cluster contains the pixels of the performance of a binary classification test such as segmentation.
non tumor region. Let I4 ðx; yÞ be the optimized segmented image Sensitivity also called as true positive rate is a measure of actual
obtained through FCM. positives that are correctly identified as such. Dice score is used
to measure how the proposed output matches with the ground
truth segmentation results is illustrated in Fig. 4.
3.6. Accurate region selection
TN
The segmented output of FCM optimization has more than one Specificity ¼ ð4:1Þ
TN þ FP
segmented region as depicted in Fig. 2(a)-(e). The exact boundary
can be chosen by choosing the region that has maximum perime- TP
ter. Therefore the perimeter of all regions are estimated and the Sensitiv ity ¼ ð4:2Þ
TP þ FN
region with maximum perimeter is chosen as the accurate region
which is shown in Fig. 2(f)-(j). The segmented output by the pro- 2  TP
posed segmentation algorithm is Zðx; yÞ. Dice score ¼ ð4:3Þ
FN þ 2  TP þ FP
The segmented result was compared with the ground truth out-
4. Experimental results put to find TN (true negative); TP (true positive); FN (false nega-
tive) and FP (false positive).
This section shows the experimental result for the proposed TP (True Positive): If the ground truth is tumour and the output
segmentation algorithm. The data set is taken from T1-weighted is also tumor.
contrast-enhanced image database that contains 233 patient TN (True Negative): If the ground truth is non-tumour and the
images with three kinds of brain tumor such as Meningioma, output is tumor.
Glioma and Pitutary tumor as shown in Table 1. The value of a, FP (False Positive): If the ground truth is non-tumour and the
b, we , wl , wt in Greedy snake algorithm are 0.7, 0.2, 0.3, 0.4 and output is tumor.
FN (False Negative): If the ground truth is tumour and the output
is non-tumor.
Table 1
Description of dataset. Probabilistic Rand Index is the metric that compares two parti-
tions such as tumor region and non tumor region. PRI estimates the
Type No of slices
fraction of pairs whose labeling are consistent between the ground
Meningioma 708 truth and obtained segmentation result. The PRI ranges from 0 to 1.
Glioma 1426
PRI ¼ 0 resembles that there is no similarities and PRI ¼ 1 resem-
Pitutary tumor 930
bles all the segmentations are identical with the ground truth.

Fig. 3. Accurate region selection based on maximum perimeter (a)-(e) FCM output (f)-(j) selected accurate region.

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
6 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Fig. 4. Experimental results for the proposed algorithm (a) Input MRI image (b) Pre-processed Image (c) Level 1 Morphological reconstructed image (d) Level 2 Morphological
reconstructed image (e) Thresholding output (f) Dilated output (g) Greedy snake model input with initial contour (h) Greedy snake output (i) Segmented FCM output (j)
Proposed segmentation output.

a1 þ a2 using a median filter and the preprocessed image is shown in

PRI ¼ ð4:4Þ
a1 þ a2 þ b1 þ b2 Fig. 5(d)–(f). The preprocessed image is background minimized,
a1 þ a2 is the similarity between segmented pixel result and thresholded and dilated to obtain the dilated image shown in
ground truth, b1 þ b2 is the difference between segmented pixel Fig. 5(g)–(i). The dilated image is used as the initial contour for
result and ground truth. greedy snake algorithm where the greedy snake segmentation out-
The performance of the proposed brain tumor segmentation put is shown in Fig. 5(j)–(l). The optimization is performed on
method was compared with the performance o the conventional greedy snake output to obtain the optimized result shown in
brain tumor segmentation methods such as K-means algorithm Fig. 5(m)–(o). Finally the proposed segmentation algorithm is
(Paul and Bandhyopadhyay, 2012), Fuzzy-C-Means algorithm shown in Fig. 5(p)–(r).
(Aparajeeta et al., 2016), Gaussian Mixture Model(Wu and Corso, Fig. 6 shows the dice-score comparison for the proposed
2012), New threshold approach (Ilhan and Ilhan, 2017) and Inter- method with the conventional methods such as GMM, FCM, K-
active segmentation (Guotai Wang et al., 2018) means, New threshold approach and Interactive segmentation for
Fig. 5 shows the images for pre-processed output, dilation, the tumors such as Meningioma, Glioma and Pituitary tumor.
greedy snake output, FCM output and final segmentation output. Table 2 shows the dice-score estimation for the proposed and
Fig. 5(a)–(c) shows the input MRI image where it is preprocessed conventional methods. The dice-score appears to be high for all

Fig. 5. Experimental results for the proposed algorithm for the type Meningioma (a)-(c) Input MRI images (d)-(f) Pre-processed Image (g)-(i) Dilated output (j)-(l) Greedy
snake output (m)-(o) Segmented FCM output (p)-(r) Proposed segmentation output.

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx 7

Dice Score Analysis Fig. 8 shows the graphical comparison for sensitivity and speci-
ficity for proposed and conventional methods for the tumor type
0..9
0..8 Meningioma.
0..7 The sensitivity of the proposed method was found to be high in
Meningioma
0..6 all the three types of tumor when compared to the traditional
0..5 methods such as Interactive, New threshold, GMM, FCM and K-
Dice Score Glioma
0..4
means. Table 3 shows the sensitivity values for proposed and con-
0..3
Pitutary tumoor ventional methods for the types of tumors. The sensitivity was
0..2
0..1
0
Meningioma Analysis
1.2
1
Methods 0.8
Performance
0.6
Fig 6. Comparison of dice score for different types of tumor such as Meningioma, Score
0.4
Glioma and Pitutary tumor for proposed method and conventional method. Sensitivity
0.2
0 Specificity

Table 2
Comparison of Dice-score between proposed and traditional Brain tumor segmenta-
tion methods.

Method Meningioma Glioma Pituitary tumor Methods

K-means 0.65 0.49 0.4
Fig 8. Performance comparison of Sensitivity and Specificity for proposed method
GMM 0.67 0.52 0.42
and conventional methods for Meningioma.
FCM 0.69 0.52 0.42
New threshold 0.72 0.54 0.45
Interactive 0.75 0.57 0.47
Proposed 0.78 0.59 0.49
Table 3
Comparison of sensitivity between proposed and traditional Brain tumor segmenta-
tion methods.
the three types of tumor. For Meningioma, Glioma and Pituitary Method Meningioma Glioma Pitutary tumor
tumor the dice-score was calculated as 0.78, 0.59 and 0.49 respec-
K-means 0.58 0.41 0.35
tively. The dice-score for Meningioma type provides a higher value GMM 0.6 0.43 0.36
where the Pitutary type provides the lower value. The Dice-score of FCM 0.6 0.45 0.38
the proposed method is greater than the traditional methods in New threshold 0.62 0.47 0.4
three different types of tumor. Figs. 5 and 7 shows the output Interactive 0.64 0.49 0.42
Proposed 0.67 0.51 0.44
images for the tumor type Meningioma and Glioma respectively.

Fig 7. Experimental results for the proposed algorithm for the type Glioma (a)-(c) Input MRI images (d)-(f) Pre-processed Image (g)-(i) Dilated output (j)-(l) Greedy snake
output (m)-(o) Segmented FCM output (p)-(r) Proposed segmentation output.

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
8 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

Fig 9. Experimental results for the proposed algorithm for the type Pituitary (a)-(c) Input MRI images (d)-(f) Pre-processed Image (g)-(i) Dilated output (j)-(l) Greedy snake
output (m)-(o) Segmented FCM output (p)-(r) Proposed segmentation output.

Table 4
Glioma Analyysis
Comparison of specificity between proposed and traditional Brain tumor segmenta-
1 tion methods.
0.9
0.8 Method Meningioma Glioma Pituitary tumor
0.7 K-means 0.87 0.88 0.87
0.6 GMM 0.9 0.9 0.89
Performance
0.5 FCM 0.92 0.92 0.9
Score
0.4 New threshold 0.94 0.92 0.9
0.3 Sensitivity Interactive 0.94 0.93 0.91
0.2
Specificity Proposed 0.96 0.94 0.91
0.1
0

Table 5
Average PRI for the proposed method without and with FCM.

Methods Method Meningioma Glioma Pituitary tumor

PRI without FCM 0.91247 0.923412 0.89321
Fig 10. Performance comparison of Sensitivity and Specificity for proposed method
PRI with FCM 0.96843 0.953242 0.93214
and conventional methods for Glioma.

Average Analysis

Pituitary tumor Analysis 1

1 0.95
0.9 0.9
0.8 0.85
0.7 0.8
0.6 Average 0.75
Performance Score 0.5 0.7
0.4 0.65 PRI without FCM
0.3 Sensitivity
0.6 PRI with FCM
0.2 Specificity 0.55
0.1
0.5
0

Methods
Methods
Fig 11. Performance comparison of Sensitivity and Specificity for proposed method
and conventional methods for Pituitary tumor. Fig 12. . Average PRI comparison for the proposed method without and with FCM.

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx 9

Fig 13. Comparison of Proposed segmented output with the ground truth for some of the test images (Row 1 shows the input images, Row 2 shows the Ground truth Results,
Row 3 shows the Proposed segmented output, Row 4 shows the corresponding HD values in mm).

calculated as 0.67, 0.51 and 0.44 for Meningioma, Glioma and Pitu- HðP; GÞ ¼ max min jx  yj ð4:6Þ
x2P y2G
itary tumor respectively. The sensitivity of the proposed method
was higher than the traditional methods for all the types of tumor. Fig. 13 shows the Comparison of Proposed segmentation output
Fig. 9 shows the output images obtained at various stages for with the ground truth results along with the HD values for few of
Pituitary tumor. Fig. 10 shows the graphically representation of the test images. The HD values are obtained by comparing the out-
Sensitivity and Specificity for the tumor type Glioma. The sensitiv- put (P) obtained with the ground truth results (G) using Eq. (4.5).
ity and specificity appears to be high for the proposed method than The HD values were found to be less than 5 for all the test images.
the conventional method for the tumor type Glioma are illustrated This indicates that there is only a small change in segmented out-
in Fig. 11. put from the ground truth results. The next section shows the con-
The specificity of the proposed method was found to be high clusion of the proposed segmentation algorithm.
(0.94) for the tumor Glioma when compared to the proposed and
traditional methods.
5. Conclusion
The specificity of the proposed method and Gaussian mixture
model is the same for the tumor types Meningioma and Pituitary
This paper proposes an automatic brain tumor segmentation
tumor. The specificity was found to be 0.96 and 0.91 for
method in MRI images. Using greedy snake algorithm and Fuzzy-
Menignioma and Pituitary tumor which is shown in Table 4.The
C-Means optimization. This method initially detects the initial
specificity of the proposed method was high than the traditional
snake contour using two level morphological reconstruction such
Interactive segmentation method for Meningioma and Pituitary
as dilation and erosion. A mask is formed by thresholding and
tumor, but the specificity similar for the tumor type Glioma.
the image is reconstructed. The initial contour of the snake is esti-
It is observed that the specificity of Menigioma and Pituitary
mated from the boundary of the eroded mask. The greedy snake
tumor is found to be the same for GMM and proposed method.
algorithm segments the region of interest using the initial contour.
The dice-score and Sensitivity of the proposed method outper-
The segmented result is optimized using FCM algorithm. Finally
forms the conventional methods for all the types of tumor such
the accurate segmented region is selected using maximum perime-
as Menigioma, Glioma and Pituitary tumor.
ter. Experimental results show that the proposed algorithm out-
The average PRI is calculated for all the test images with and
performs the conventional method in terms of dice-score,
without FCM for the tumor types Menigioma, Glioma and Pituitary
specificity and sensitivity, for different types of tumor such as
tumor. The average PRI is tabulated in Table 5
Meningioma, Glioma and Pitutary tumors. Also the proposed seg-
The PRI is less for the proposed method without including the
mentation output was compared with the Ground truth results
FCM block. The PRI increases in three different types of tumor with
using Hausdorff distance. The Hausdorff distance provides a lesser
the use of FCM algorithm which is depicted in the graph shown in
values and this shows that, there is only small changes in seg-
Fig. 12. This shows that the FCM optimizes the boundaries of the
mented output when compared to Ground truth results.
tumor tissues.
The Hausdorff distance (HD) gives the maximum surface dis-
tance measures. It represents the maximum difference between Conflict of interest
the tumor size. Let G represents the ground truth segmentation
output and P represent the proposed segmentation output. Then, There is no conflict of interest.
the maximum surface distance can be measured using Hausdorff
distance as,
References
HD ¼ maxðHðP; GÞ; HðG; PÞÞ ð4:5Þ
Ahmed, M.N., Yamany, S.M., Mohamed, N., Farag, A.A., Moriarty, T., 2002. A modified
where fuzzy c-means algorithm for bias field estimation and segmentation of MRI

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006
10 C. Jaspin Jeba Sheela, G. Suganthi / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx

data. IEEE Trans. Med. Imaging 21, 193–199. https://doi.org/10.1109/ Essadike, A., Ouabida, E., Bouzid, A., 2018. Brain tumor segmentation with Vander
42.996338. Lugt correlator based active contour. Comput. Methods Programs Biomed. 160,
Angel Viji, K.S., Jayakumari, J., 2013. Modified texture based region growing 103–117. https://doi.org/10.1016/j.cmpb.2018.04.004.
segmentation of MR brain images. (:unav). https://doi.org/ Havaei, M., Davy, A., Warde-Farley, D., Biard, A., Courville, A., Bengio, Y., Pal, C.,
10.1109/cict.2013.6558183 Jodoin, P.-M., Larochelle, H., 2017. Brain tumor segmentation with Deep Neural
Angulakshmi, M., Lakshmi Priya, G.G., 2017. Automated brain tumour segmentation Networks. Med. Image Anal. 35, 18–31. https://doi.org/10.1016/
techniques- a review. Int. J. Imaging Syst. Technol. 27, 66–77. https://doi.org/ j.media.2016.05.004.
10.1002/ima.22211. Ilhan, U., Ilhan, A., 2017. Brain tumor segmentation based on a new threshold
Aparajeeta, J., Nanda, P.K., Das, N., 2016. Modified possibilistic fuzzy C-means approach. Procedia Comput. Sci. 120, 580–587. https://doi.org/10.1016/j.
algorithms for segmentation of magnetic resonance image. Appl. Soft Comput. procs.2017.11.282.
41, 104–119. https://doi.org/10.1016/j.asoc.2015.12.003. Ilunga-Mbuyamba, E., Avina-Cervantes, J.G., Garcia-Perez, A., de Romero-Troncoso,
Aswathy, S.U., Glan Deva Dhas, G., Kumar, S.S., 2014. A survey on detection of brain R., Aguirre-Ramos, J., Cruz-Aceves, H., Chalopin, C, I., 2017. Localized active
tumor from MRI brain images. (:unav). https://doi.org/10.1109/ contour model with background intensity compensation applied on automatic
iccicct.2014.6993081 MR brain tumor segmentation. Neurocomputing 220, 84–97. https://doi.org/
Bauer, S., Nolte, L.-P., Reyes, M., uer et al. 2011. Fully Automatic Segmentation of 10.1016/j.neucom.2016.07.057.
Brain Tumor Images Using Support Vector Machine Classification in Kandwal, R., Kumar, A., Bhargava, S., 2014. Review: Existing Image Segmentation
Combination with Hierarchical Conditional Random Field Regularization. In: Techniques. Int. J. Adv. Res. Comput. Sci. Software Eng. 4.
Fichtinger, G., Martel, A., Peters, T. (Eds.), Medical Image Computing and Morales, S., Bernabeu-Sanz, A., López-Mir, F., González, P., Luna, L., Naranjo, V., 2017.
Computer-Assisted Intervention – MICCAI 2011. Springer, Berlin Heidelberg, BRAIM: a computer-aided diagnosis system for neurodegenerative diseases and
Berlin, Heidelberg, pp. 354–361. https://doi.org/10.1007/978-3-642-23626- brain lesion monitoring from volumetric analyses. Comput. Methods Programs
6_44. Biomed. 145, 167–179. https://doi.org/10.1016/j.cmpb.2017.04.006.
Bauer, S., Wiest, R., Nolte, L.-P., Reyes, M., 2013. A survey of MRI-based medical Paul, T.U., Bandhyopadhyay, S.K., 2012. Segmentation of Brain Tumor from Brain
image analysis for brain tumor studies. Phys. Med. Biol. 58, R97–R129. https:// MRI Images Reintroducing K - Means with advanced Dual Localization Method.
doi.org/10.1088/0031-9155/58/13/R97. Int. J. Eng. Res. 2, 6.
Bengio, Y., 2012. Practical Recommendations for Gradient-Based Training of Deep Selvaraj, D., Dhanasekaran, R., 2013. MRI Brain Image Segmentation Techniques - A
Architectures. In: Montavon, G., Orr, G.B., Müller, K.-.R. (Eds.), Neural Networks: Review 4, 18.
Tricks of the Trade. Springer, Berlin Heidelberg, Berlin, Heidelberg, pp. 437–478. Shen, S., Sandham, W., Granat, M., Sterr, A., 2005. MRI Fuzzy Segmentation of Brain
https://doi.org/10.1007/978-3-642-35289-8_26. Tissue Using Neighborhood Attraction With Neural-Network Optimization. IEEE
Cabria, I., Gondra, I., 2017. MRI segmentation fusion for brain tumor detection. Trans. Inf Technol. Biomed. 9, 459–467. https://doi.org/10.1109/
Information Fusion 36, 1–9. https://doi.org/10.1016/j.inffus.2016.10.003. TITB.2005.847500.
Clark, M.C., Hall, L.O., Goldgof, D.B., Velthuizen, R., Murtagh, F.R., Silbiger, M.S., 1998. Sujji, G.E., Lakshmi, Y.V.S., Jiji, G.W., 2013. MRI Brain Image Segmentation based on
Automatic tumor segmentation using knowledge-based techniques. IEEE Trans. Thresholding. Int. J. Adv. Comput. Res. 3, 5.
Med. Imaging 17, 187–201. https://doi.org/10.1109/42.700731. Wang, G., Li, W., Zuluaga, M.A., Pratt, R., Patel, P.A., Aertsen, M., Doel, T., David, A.L.,
Cobzas, D., Birkbeck, N., Schmidt, M., Jagersand, M., Murtha, A., 2007. 3D Variational Deprest, J., Ourselin, S., Vercauteren, T., 2018. Interactive Medical Image
Brain Tumor Segmentation using a High Dimensional Feature Set. In: 2007 IEEE Segmentation Using Deep Learning With Image-Specific Fine Tuning. IEEE
11th International Conference on Computer Vision. Presented at the 2007 IEEE Trans. Med. Imaging 37, 1562–1573. https://doi.org/10.1109/tmi.2018.2791721.
11th International Conference on Computer Vision, IEEE, Rio de Janeiro, Brazil, Wu, W., Corso, J.J., 2012. Brain tumor segmentation based on GMM and active
pp. 1–8. https://doi.org/10.1109/ICCV.2007.4409130 contour method with a model-aware edge map 4.
Dass, R., Devi, S., 2012. Image Segmentation Techniques 3, 5.

Please cite this article as: C. J. J. Sheela and G. Suganthi, Automatic Brain Tumor Segmentation from MRI using Greedy Snake Model and Fuzzy C-Means
Optimization, Journal of King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2019.04.006