Essay: Fuzzy C Mean Thresholding based Level Set for Segmentation of brain stroke using CT Images

Abstract
Accurate segmentation is a very important task in any computer vision systems. It also plays a vital role in computerized analysis of brain stroke CT images. The objective is to present a new segmentation method for brain stroke detection that combines the advantages of fuzzy c-means (FCM), thresholding and the level set method. 3-class fuzzy c-means based thresholding method is applied so that level set evolution get initialised automatically and also the controlling parameters of level set evolution are derived from the results of fuzzy clustering. In the proposed method, the contour of the image is obtained by FCM thresholding method which serves as initial contour for level set method and the final segmentation is achieved using level set method. As well as, the segmentation results of proposed method are compared with single seeded region growing algorithm, Otsu’s thresholding and 3-class FCM thresholding.
Keywords Brain stroke . Segmentation . Diagnosis . Fuzzy . Thresholding . Level sets . Seed region growing algorithm
1 Introduction
Stroke formerly known as cerebrovascular accident (CVA) is a leading cause of death and disability in the world. According to World Health Organization (WHO), 15 million people suffer from stroke, out of which 5 million die and another 5 million get permanently disabled. Stroke is diagnosed through various imaging technologies and these are: Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET) and Single-Photon Emission Computed Tomography (SPECT). Stroke occurs when the brain stops functioning due to disturbance in the artery that carries blood to the brain. In medical term, it is defined as sudden death of brain cells due to lack of oxygen [1]. There must be a continuous supply of blood, oxygen, glucose (blood sugar) to the nerve cells within the brain for it to function properly. The parts of the brain stop functioning temporarily if the supply is impaired and if the impairment is critical or continuous for the longest time, the brain cells die and damage permanently. The topology of stroke is divided into two categories:
Hemorrhagic stroke: It is due to rupture of weakened blood vessels.
Ischemic stroke or infarct: It is due to interruption of blood supply i.e. when the blood vessel that carries blood to the brain is blocked due to a thrombus.
Of these, ischemic stroke occurs usually. About 85% are of ischemic type. There is also a possibility that both ischemic and hemorrhagic stroke can occur at the same time [1]. Stroke results in serious disability or death. Though both being fatal in nature, complete recovery can be achieved in a fair number of patients suffering from hemorrhagic strokes whereas the chances of complete recovery in the number of patients suffering from ischemic stroke are considerably less. The recovery from ischemic stroke depends on the timing of treatment which can only be performed within the first few hours [0-4 hours] from the onset. With the passage of each second of time, more and more brain tissues suffer irreparable damage. The symptoms of stroke are: weakness or paralysis altered feeling on one side of the body, difficulty in speech, blurred vision, confusion and severe headache [1][2]. The risk factors for
stroke are: old age, high blood pressure, Transient Ischemic Attack (TIA), diabetes, high cholesterol, tobacco smoking etc. The most important risk factor for the stroke to happen is the high blood pressure.
1.1 Brain Stroke
Since the brain is the control centre of the human body so it requires a large amount of body resources in order to carry out important functions of the body properly. Thus, there must be a continuous and proper supply of blood, oxygen, glucose (blood sugar) to the nerve cells within the brain for it to function properly. Any disturbance in the supply of these resources can lead to complications which in many cases may prove fatal. Stroke is one such complication which can occur in form of ischemic and hemorrhagic stroke.
For the Ischemic stroke, around 85% of strokes are ischemic in nature. It occurs when the blood vessel or artery that carries blood to the brain is blocked due to thrombus or clot. Ischemic stroke is further divided into three categories on the basis of time passed since the onset of symptoms including hyperacute, acute and chronic. Hyperacute being the earliest stage and chronic being the latter stage at which point of time most of the tissues have undergone irreparable damage. The most common ischemic stroke includes:
Thrombotic stroke: A thrombotic stroke occurs when a thrombus or blood clot is formed in one of the arteries that supplies blood to the brain. A clot may be formed by the fatty deposits (plaque) that build up in arteries that reduced the flow of blood or other arteries conditions.
Embolic stroke: An embolic stroke occurs when a blood clot is formed in the artery that is away from the brain commonly in heart and is swept away through bloodstream to lodge in narrower brain arteries. This type of blood clot is known as embolus.
For the Hemorrhagic stroke, it occurs due to rupturing of weakened blood vessels in the brain. It is also possible that both the strokes can occur at the same time. Types of hemorrhagic stroke include:
Intracerebral haemorrhage: In an intracerebral haemorrhage, a blood vessel in the brain bursts and spills over the brain into the surrounding brain tissues which damages brain cells. Brain cells beyond the leak are deprived of blood and get damaged.
Subarachnoid haemorrhage: In a subarachnoid haemorrhage, a blood vessel on or near the surface of brain bursts and spills into the space between the surface of brain and skull. This bleeding is often signalled by an abrupt, severe headache.
The diagnostic strategies of ischemic and hemorrhagic differ in the principle and the treatment of one may prove fatal for the other. Hemorrhagic stroke requires the addition of coagulating agents in the blood in order to stop blood flow through the damaged artery where as ischemic stroke requires the addition of blood thinning agents in order to increase the flow of blood through the blocked artery. The main treatment for the ischemic stroke requires the use of intravenous (IV) recombinant tissue plasminogen activator (t-PA) but it increases the risk the risk of haemorrhage by 10 times so it is recommended where a substantial amount of recovery is possible. Fig.1 shows the appearance of different types of stroke and normal brain tissue as seen on a CT scan.
(a) Haemorrhage (b) Chronic (c) Acute (d) Hyperacute (e) Normal
(Left centre) (Right centre) (Top left half)
Fig.1 Representation of various types of stroke and normal brain on CT images [4].
1.2 Brain Imaging Techniques
Brain imaging techniques allow doctors and researchers to view activity or problems within the human brain without invasive
neurosurgery. Some of the widely used imaging technologies are: Computed Tomography (CT), Magnetic Resonance Imaging (MRI),
Positron Emission Tomography (PET) and Single – Photon Emission Computed Tomography (SPECT) as shown in Fig.2. CT and
MRI give physical aspects of the brain where as PET and SPECT shows actual working of brain. Out of these, CT and MRI are
usually used in the identification of stroke and are discussed in more detail in the following sections.
(a) CT (b) MRI (c) PET (d) SPECT
Fig.2 Representation of brain under various modalities [4]
For Magnetic Resonance Imaging (MRI), it uses magnetic fields and radio waves to produce detailed 3-D images.MRI scanners and a
computer produce images of the internal structures of the body including the brain and spinal cord, bones and joints, the heart and
blood vessels, breast tissue and other internal organs. As seen from figure (Fig.2 (b)), MRI provides a very detailed image having
good contrast between various types of tissues. MRI has a much greater range of available soft tissue. Contrast depicts anatomy in
greater detail. Unlike CT, it doesn’t use ionising radiation.
For computed Tomography (CT), CT uses X-rays to create detail cross-sectional images of the inside of body. CT scan can produce
images of every type of body structure including organs, bones and blood vessels and is used by health professionals to help diagnosis
and manage many health conditions but precise details of soft tissue especially the brain, some organs and around joints are less
visible on CT scans. Although MRI gives more detailed picture of the brain but CT is the preferred procedure in the acute stroke unit
due to the advantages of high speed, low cost, reduced exclusion criteria to MR imaging and sensitiveness to early stroke.
Disadvantages of MRI include high cost, lack of ready availability at most centres and insensitivity for detecting early haemorrhages
[15]. In CT images, a haemorrhage appears as a bright region (hyperdense) displaying sharp (bright) contrast against its surrounding.
Conversely, an ischemic stroke appears as a dark region (hypodense) with the contrast relative to its surroundings depends on the time
elapsed since the stroke occurred [5] as shown in Fig.3.
(a) Haemorrhage (left centre) (b) Chronic (right centre)
Fig.3 Representation of strokes in CT brain images [4]
1.4 Brain Stroke Segmentation Methods
Image segmentation is a broad and active field in medical imaging. Image segmentation is a procedure by which an image is divided
into the constituent parts or objects that are present in the image. Thus the main of subdividing an image into its constituent parts or
objects present in the image is that each of the constituent parts or objects present in the image can be analysed further to extract some
meaningful information so that extracted information is useful for high level machine vision applications. The level of this subdivision
depends on the problems which are being solved. Medical image segmentation is an important task for identification and location of
strokes, diagnosis and computer guided surgery etc. [12]. Nowadays, brain stroke segmentation methods can be classified into
different categories based on different principles. Brain stroke segmentation methods are usually classified into three main categories
including manual, semiautomatic and full automatic segmentation based on the degree of required human interaction [12-14]. They all
have their pros and cons.
For manual brain stroke segmentation, the experts of brain stroke must master the information present in the brain stroke images and
some additional knowledge such as anatomy because manual brain stroke segmentation aims to draw the boundaries of brain stroke
manually [12]. Manual delineation is widely used to clinic trial to-date. Manual delineation of stroke takes up to 30min. per patient in
clinic and large population studies contains hundreds to thousands of patients [16]. Thus, manual segmentation of stroke is the most
accurate one by the domain experts but time consuming. More advanced segmentation methods such as semiautomatic and fully
automatic segmentation methods are required.
For semiautomatic brain stroke segmentation, it mainly consists of user interaction and software computing. In the semiautomatic
brain stroke segmentation methods, the user needs to input some parameters and is responsible for analyzing the visual information
and providing feedback response for the software computing. The software computing is selected at the realisation of brain stroke
segmentation algorithms. The interaction is in charge of adjusting segmentation information between the user and software computing
[17]. Although semiautomatic brain stroke segmentation methods can give better results as compare to manual segmentation but it
also comes into being different results by the different users or by the same user at different times.
For fully automatic brain stroke segmentation, the computing software determines the segmentation of brain stroke without any human interaction. Generally, fully automatic segmentation algorithm combines artificial intelligence and prior knowledge. Since automatic segmentation doesn’t require any input from the user, thus it is much more difficult to obtain accurate results.
CT based brain stroke segmentation methods are divided into four major categories: Threshold based methods, Region growing based methods, neural network based methods and Deformable model based methods.
2 Survey on Previous Works
In the literature survey, we found that several algorithms have been developed for brain stroke segmentation in MR images, only a few methods have been proposed for the segmentation of stroke lesion in CT images. Most of the existing work with CT images has been directed towards the segmentation of hemorrhagic stroke.
S.Neethu and D.Venkataraman [1] proposed a method for the detection of stroke region in the absence of doctors or radiologists. In their method, they have applied region growing technique to segment the stroke part and also used histogram equalization for better visualization. This method gives better results for haemorrhagic stroke as compared to ischemic stroke. In the ischemic stroke images, boundaries are not clearly separated.
Li et al. [9] proposed a method for automatic subarachnoid space segmentation (SAS) and haemorrhage detection in clinical CT scans. In their method, they have applied morphological operations and fuzzy c means clustering to extract the brain part. After that five land mark of brain were extracted on the basis of distance. Classification based on Bayesian distance framework was used for each pixel to classify it as a SAS or non SAS.
Chawla et al. [7] has addressed the problem of detecting both hemorrhagic stroke and ischemic stroke in a given CT volume. In this study, lesion tissue was identified by comparing image intensities in the two hemispheres under the assumption that the intensity of the abnormal region in one hemisphere will be different as compared to the intensity of other hemisphere but if the symmetrical abnormalities occur on both sides with respect to the brain midline, this approach is imperfect. Clinical data indicates that symmetrical abnormalities on both the sides of the hemisphere are possible, but the numbers of such cases are limited.
Jeena and Kumar [3] proposed an algorithm using which they compared the stroke on CT and MRI images. In their method, they have used digital processing tools for the identification of infract and haemorrhage in human brain. After that seeded region growing algorithm along with Gabor filter is used for the segmentation. Rai and Nair [10] discussed about seeded region grow methodology for segmenting anatomical structures in CT Angiography images. They have proposed a gradient based homogeneity criteria to control the region grow process while segmenting CT Angiography images.
3 Methodology
3.1 Problem Formulation
Since doctors need to know the volume or area of stroke region for the treatment planning so the boundaries of stroke region should be clearly defined. The introduction of the level set method in medical image segmentation is shown to be effective for tracking the
boundaries that use dynamic implicit interfaces. The evolution of this method is needed to be initialized with manual selection or intensity thresholding. But this type of initialization is not a reliable choice for an optimal level set segmentation. After analyzing the advantages and disadvantages of image segmentation techniques, it is observed that fuzzy c-mean (FCM) clustering method performs well for providing rough clustering of pixels which can be used for getting better threshold values futher for the image segmentation. In addition to this, if the thresholded image is used as initial estimate for the level set evolution, it will result in more accurate segmentation with very fine and tuned boundaries as compared to standard level set method. We, hereby presented a technique, based on combining the threshold method, the fuzzy c-means (FCM) clustering method and the level set method to achieve higher accuracy in segmentation of brain stroke in CT images. Hence, the level set segmentation technique will be implemented using a 3-class fuzzy thresholding algorithm.
3.2 Proposed Methodology
The aim is to present a new segmentation method for brain stroke detection that combines the advantages of fuzzy c-means (FCM), thresholding and the level set method. 3-class fuzzy c-means based thresholding method is applied so that level set evolution get initialised automatically and also the controlling parameters of level set evolution are derived from the results of fuzzy clustering. In the proposed method, the contour of the image is obtained by FCM thresholding method which serves as initial contour for level set method and the final segmentation is achieved using level set method. Framework of proposed method is shown in Fig.5
Filtered image
Preprocessing
Input brain CT image
Fig. 5 Framework of proposed method
The main parts of our proposed algorithm are as follows:
Fuzzy c-means (FCM) thresholding algorithm
Initialise level set function
Calculate edge function (g)
Determines controlling parameters for regularizing level set (LS) evolution process automatically (λ, μ, τ, ν)
Level set (LS) evolution
Image Preprocessing
The preprocessing is used to smoothness, enhance and remove noise that is caused by defects of CT scanner. Thus, pre-processing improves the quality of the image and emphasizes certain features so that segmentation would be easier and more effective. Study shows that CT images contain Gaussian noise. A Gaussian filter is used for noise removal. Gaussian filter is a linear smoothing filter having a very unique property that that the function remains real and Gaussian in both spatial and frequency domain. A Gaussian filter in the frequency domain is represented as,
G (u) (1 )
Where A – constant
– Standard deviation of the Gaussian functions.
In the spatial domain, it is given as,
g(x) = ( 2 )
Both the equations constitute a Fourier transform pair that remains real and Gaussian.
3-class Fuzzy C-Means (FCM) Thresholding
Generally, FCM clustering is used to divide N objects into C classes (cluster). Since we are interested in segmenting only three major tissues in the human brain: the cerebrospinal fluid (CSF), the gray matter (GM) and the white matter (WM), thus in our method N is equal to the number of pixels in the image i.e. N = Nx × Ny and C = 3 for 3-class FCM clustering. The FCM algorithm uses iterative optimization of an objective function and it works by assigning membership to each data point corresponding to each cluster centre on the basis of the distance between the cluster centre and data point. Clearly, Summation of all the membership degrees of every data point to all clusters must be one. The algorithm of FCM clustering with thresholding is given sequentially as follows:
1. Define the number of Clusters C =3 and initialize the initial fuzzy partition matrix U = [uij] with i = 1, 2…..N and j = 1, 2………….C where uij indicates the degree of membership of ith pixel to the jth centre whose values lie between 0 and 1. Also define exponent m and termination tolerance . The best choice for m is in the interval [1.5, 2.5] so m = 2 is used here as it is widely accepted as a good choice of fuzzy partition matrix U = [uij].
2. Calculate C fuzzy cluster centre Cj; j = 1, 2, 3…………….C using equation:
Cj ; j = 1, 2, 3……………C ( 1 )
3. After calculating the centre of each cluster, the fuzzy partition matrix uij is updated by the following equation:
Uij = ; 1 ≤ i ≤ N and 1 ≤ j ≤ N ( 2 )
4. Compute the energy function by using the following equation:
JFCM (U, C1, C2………….Cc) = ( 3 )
Membership functions uij must be non negative and satisfy the constraint i = 1, 2……….N.
5. The process stops when the maximum number of iterations is attained or change in membership matrix is lower than a predefined threshold or when the improvement in the objective function between two consecutive iterations is less than the minimum amount of improvement specified.
i.e. if < then stop, otherwise return to step 2.
Where is termination iteration between 0 and 1.
K is the maximum number of iterations.
The FCM is based on the minimization of an objective function mentioned in step 4. The algorithm starts with defining the number of clusters = 3 and initializing the membership matrix U. This matrix contains the membership degrees for all data points for each cluster. The initialization of U is done randomly and the cluster centres are computed by using the membership matrix U. The cluster centres are calculated in such a way that the centre is closer to those points that have greater membership value to one cluster.
Once the cluster centres have been computed, the membership matrix is updated according the location of cluster centres. To calculate the new membership value of a point with respect to a particular cluster, the distance of that point from that cluster centre and the distance of that point from all other cluster centres are taken into account. Now the change in membership matrix is computed. If the change in membership matrix is lower than a predefined threshold or energy function J gets minimized, then the process is stopped otherwise, new cluster centres are calculated and membership matrix is updated with respect to the new cluster centres. The iteration continues till the change in membership matrix is minimized.
Fuzzy Threshold Selection
Fuzzy C-means (FCM) clustering is a soft segmentation method that has been used extensively to improve the compactness of the regions due to its cluster validity and simple implementation depends on the Euclidean distance between the data points on the basis of the assumption that each feature is equally important but in most of the real world applications, features are not of equal importance. Thus, this assumption affects the performance of clustering and the segmented part cannot be seen visibly by the FCM clustering. Hence, to improve the performance of clustering, FCM clustering is combined with Otsu thresholding on the basis of the advantages of both FCM clustering and thresholding. It works better than Otsu’s method. Finally after performing FCM clustering, each pixel is assigned to the cluster with highest membership value and the threshold value is calculated by taking the mean of maximum of cluster 1 and minimum of cluster 2 or mean of maximum of cluster 2 and minimum of cluster 3 on the basis of the intensity distribution obtained by using the histogram of the image. The output of this stage is a binary image which is used in the followed steps.
Fuzzy C-Means Thresholding based Level Set Segmentation
Level set method (LSM) is one of the automatic process approaches that have shown effective results for medical image segmentation. The basic idea behind the level set method (LSM) is to imbed a curve within a surface. In 2-D segmentation, the LSM represents a closed curve as the zero level set of called the level set function. In LSM, the movement of the zero level set is driven by the level set equation (LSE) which is a partial differential equation (PDE). Thus, LS method is a very complex and time consuming due to intensive computational requirements and regulation of controlling parameters. To overcome such problem, fuzzy c-means thresholding based method is used to facilitate the level set segmentation for automatic segmentation of brain stroke.
Level set method has been extensively used for the problem of curve evolution. Level set method utilizes dynamic implicit interfaces and partial differential equation (PDEs). The counter denoted by C is represented by the zero level set = 0 of a level set function in traditional level set (LS) formulation. The evolving equation of the level set function is called as level set equation (LSE) which is written in the general form as-
( 1 )
Where, the function F is the speed function and it represents the comprehensive forces including, the internal force from the interface geometry and the external force from image gradient. In order to stop the level set evolution near the optimal solution the advancing forces have to be regularized by an edge indication function g. The edge indication function g is given by as-
( 2 )
Where, – filtered image
Thus, the intensive computational requirements make the traditional LS method a complex method and it has a certain limitation of re-initialization of level set function to signed distance function that is needed for stable curve evolution. In order to overcome such limitation, fast level set formulation has been used in this thesis which was proposed by Li. Et al. [27]. This method is more efficient in computation and it can be implemented by using finite difference scheme. In order to segment the brain stroke CT image, the overall iterative process for level set evolution is given as-
( 3 )
Where, is the term for attracting towards the variational boundary.
& – the penalty term that forces to approach the signed distance function automatically.
This fast level set formulation proposed in [24] provides a benefit of flexible initialization where rough region obtained from fuzzy c-means thresholding algorithm is used to construct the initial set function. Hence, binary image (Bi) obtained from fuzzy c-mean thresholding algorithm that is discussed above is used for automatic initialization of the level set function here. The initial level set function is given by-
( 4 )
Where, ϵ – regulator for Dirac function which is defined as –
( 5 )
Binary image (Bi) obtained from fuzzy c-means thresholding operation determines the controlling parameters for regularizing the LS evolution process automatically.
Controlling parameters for the level set evolution are:
 Weighting coefficient μ of penalty term ζ = Area of pixels in binary image perimeter of pixels
Here, μ = 0.1 / time step
 The time step τ = 0.1 so that product time step * μ remains less than 0.25 which is the necessary condition in order to achieve stable evolution.
 λ– Coefficient of contour length for regulation of smoothness and its value taken here =1. The value of λ can be increased for accelerating the evolution process but it results in smoother contour. Hence essential care must be taken especially for the segmentation brain stroke CT image where over smoothened details about boundary of stroke lesion which is an important feature for correct diagnosis.
 ν – The balloon force that determines the speed and advancing direction of the evolving curve is given by-
( 6 )
Where, – modulating argument.
4 Experiment and Results
Here, fuzzy c-means thresholding based algorithm and the hybrid proposed algorithm for the brain stroke detection in CT images by the integration of thresholding, fuzzy c-means and level set method described in the section 3 are implemented using MATLAB software and the results of proposed method are compared with Otsu’s method, 3-class FCM thresholding method as well as with the results of base paper.
4.1 Detection of haemorrhagic stroke
Fig.7 shows the original input brain stroke CT images for the detection of haemorrhagic stroke. Each image has different shape, size and location of stroke lesion.
(i) (ii) (iii) (iv) (v)
Fig.7 Input Images
Implementation of 3-class FCM thresholding
In this section, simulation results for the 3-class FCM with thresholding algorithm are presented. As well as, the segmentation results of 3-class FCM thresholding algorithm are compared with Otsu thresholding. Segmentation results of 3-class FCM thresholding algorithm on all the above presented five input images are shown in Fig. 8.
(i) (ii)
(iii) (iv)
Fig. 8 Segmented Results of 3-class FCM Thresholding and Otsu Thresholding.
It can be seen from results shown in Figure 4.2 that 3-class FCM with thresholding method generates two output images correspond to each threshold level and this method worked better than Otsu thresholding. Otsu’s method failed to detect the object.
Implementation of 3-class FCM Thresholding with Level Set (Proposed Method)
This section evaluates the performance of proposed method in the segmentation of haemorrhagic stroke in various CT images. Each image has different shape, size and location of stroke lesion. Segmentation results of proposed method which is formed by the integration of the thresholding method, the fuzzy c-means method and the level set method are shown from Fig.9 – 12. It is an iterative process which continues until the maximum number of iterations is reached or the improvement in the objective function between two consecutive iterations is less than the minimum amount of improvement specified.
Case 1
(a) (b) (c)
(d) (e) (f) (g)
Fig.9 Segmentation results of haemorrhagic stroke (a) original image, (b-f) intermediate images, (g) final detected segment.
Case 2
(a) (b) (c)
(d) (e) (f) (g)
Fig.10 Segmentation results of haemorrhagic stroke (a) original image, (b-f) intermediate images, (g) final detected segment.
Case 3
(a) (b) (c)
(d) (e) (f) (g)
Fig.11 Segmentation results of haemorrhagic stroke (a) original image, (b-f) intermediate images, (g) final detected segment.
Case 4
(a) (b) (c)
(d) (e) (f) (g)
Fig.12 Segmentation results of haemorrhagic stroke (a) original image, (b-f) intermediate images, (g) final detected segment.
Table1 shows comparative analysis of segmentation results of base paper (Single Seeded Region Growing) with Otsu’s thresholding, 3-class FCM thresholding and Proposed method.
Table 1 Comparative analysis of base paper results with Otsu’s Thresholding, FCM Thresholding and Proposed Method
Original image
Single Seeded Region Growing Segmentation Results (base paper )
Otsu’s Thresholding Segmentation Results
FCM Thresholding Segmentation Results
FCM Thresholding based Level Set (Proposed method) Segmentation Results
On the basis of the results shown for the detection of haemorrhagic stroke, it is observed that by single seeded region growing technique (base paper) the boundaries of stroke region are not clearly defined. Otsu’s method completely failed to detect the stroke area. After that the segmentation results obtained using 3-class fuzzy thresholding algorithm are adequate but not satisfactory as the boundaries of the haemorrhagic stroke area are not well defined by the 3-class fuzzy thresholding algorithm. However, the segmentation results of the proposed algorithm show that the proposed algorithm achieves reasonably good accuracy with very fine and tuned boundaries of the haemorrhagic stroke area which is formed by the integration of thresholding, FCM means and level set techniques. Fuzzy level set gives better segmentation results as compared to the FCM thresholding for the haemorrhagic stroke detection in CT images. Table 2 shows comparison of FCM thresholding and Proposed Method.
Original Image
FCM Thresholding
Number of Iterations
FCM Thresholding based Level Set (Proposed Method )
Number of iterations
FCM Thresholding
Elapsed Time
FCM Thresholding based Level Set (Proposed Method)
Elapsed Time
22
70
4.282464
18.0018
26
70
5.186856
20.5250
29
100
10.789322
39.1854
18
170
5.579075
68.7439
It is concluded that Fuzzy level set gives better segmentation results as compared to the FCM thresholding for the haemorrhagic stroke detection in CT images but it requires more iterations count to achieve the desired optimal result. Hence, it consumes more elapsed time than FCM thresholding method.
4.2 Detection of Ischemic Stroke
Implementation of 3-class FCM thresholding
Implementation of 3-class FCM Thresholding with Level Set (Proposed Method)
Results of Single Seede Region Growing Technique (base paper)
(a) Original Image (b) Output of Region Growing Process
On the basis of results shown for the detection of ischemic stroke, it is observed that all the segmentation methds including single seeded region growing , Otsu’s method , 3-class fuzzy thresholding algorithm and proposed algorithm failed to detect ischemic stroke in CT images.
Since the topology of stroke is classified in two categories: haemorrhagic stroke due to rupture of weakened blood vessels and ischemic stroke due to interruption of blood supply. In CT images, haemorrhagic stroke appears as a bright region against its sorroundings and ischemic stroke appears as a dark region against its sorroundings and the most important property of the ischemic stroke is its asymmetry in two hemispheres. Thus, there occurs variation in intensity and also the nature of ischemic stroke is more challenging so same segmentation technique can not be used to detect both haemorrhagic stroke and ischemic stroke. Hence, proposed algorithm is not suited for all types of medical images.
5 Conclusions and Future Scope
Brain stroke is classified into two categories: Haemorrhagic stroke and Ischemic stroke. There are many algorithms for the detection of brain stroke in CT images. Each algorithm has some advantages and disadvantages. It is not necessary that an algorithm suggested for the segmentation of a particular image will be suited for all types of images. Generally, the segmentation of medical images is not easy as proper segmented image can be obtained only by applying the series of suitable algorithms. In this thesis, an optimal method for the brain stroke detection in CT images is explored by the hybridization of FCM thresholding and level set method. It combines the advantages of thresholding, FCM clustering and level set method for getting more accurate results. Now, it can be seen from the simulation results that proposed method achieved reasonably good accuracy for the segmentation of haemorrhagic stroke in CT images. However, it failed to achieve desired segmentation of CT image having ischemic stroke. FCM thresholding results are good when compared to Otsu’s method but are not satisfactory. When FCM clustering technique is combined with local thresholding, the optimization property of FCM gets improved. The image obtained from the FCM thresholding consists of some deblurred edges and some incorrect segmented part. To improve the segmentation, the level set method is used. The application of fuzzy level set gives enhanced results. The level set evolution will start from a region which is close to the genuine boundaries. Controlling parameters for the proposed algorithm are estimated from the FCM clustering automatically. This has reduced the manual intervention. Since fuzzy level set gives better segmentation results as compared to the FCM thresholding but it requires more computation time than FCM thresholding method. The proposed model can be useful for segmenting the images in medical and other field.
In the future work, ischemic stroke detection in CT images will be carried out. As well as to increase the efficiency of fuzzy level set method, future research will strive towards improving the computational speed of this method. The given method presented here is semiautomatic which present different results by the different users or by the same user at different times. In future we try to convert it into fully automatic segmentation method which will help the physician for better treatment. For these types of images, there is a very big scope of improvement in future.
References
1. S. Neethu and D. Venkataraman, Stroke detection in brain using CT images in Springer. (2015), 324, pp. 379-386.
2. S. Roy, K. Chatterjee, S.K Bandyopadhyay, Segmentation of acute brain stroke from MRI of brain image using power law transformation with accuracy estimation. Springer International Publishing. (2014), 27, pp. 453-461.
3. R.S. Jeena, S. Kumar, A Comparative analysis of MRI and CT brain images for stroke diagnosis. IEEE International Conference on Microelectronics, Communication and Renewable Energy. ICMICR (2013), 6, pp. 1–5.
4. Saurabh Sharma, M.S in Computer Science, International Institute of Information Technology, Hyderabad, October (2013).
5. J. Vymazal, A.M Rulesh, J.Keller, L. Janouskova, Comparison of CT and MR imaging in ischemic stroke. Springer Insights into Imaging. (2012), 3(6), pp. 619-627.
6. C.R Gillebert, G.W Humphreys, D. Mantini, Automated delineation of stroke lesions using brain CT images. Elsevier. (2014), 4, pp. 540-548.
7. C. Huang, B. Yan, H. Jiang, D. Wang, MR image segmentation based on fuzzy c-means clustering and the level set method. IEEE Fifth International Conference on Fuzzy Systems and Knowledge Discovery (FKSD). (2008), 1, pp. 67-71.
8. M. Chawla, S. Sharma, J. Sivaswamy, L.T. Kishore, A method for automatic detection and classification of stroke from brain CT images Engineering in medicine and biology society. Annual International Conference of the IEEE. (2009), pp. 3581–3584
9. Y.-H. Li, L. Zhang, Q.-M. Hu, H.-W. Li, F.-C. Jia, J.-H. Wu, Automatic subarachnoid space segmentation and haemorrhage detection in clinical head CT scans in Springer. (2012) 7, pp. 507–516
10. G.H. Rai, T.R. Nair, Gradient based seeded region grow method for CT angiographic image segmentation. Int. JRI Computer Science and Networking. (2009), 1(1), pp. 1-6
11. Darius Grigatis, Doctoral, Electrical Engineering and Electronics, Vilnius Gediminas Technical University, 2003-2007
12. D.L. Pham, C. Xu, J.L. Prince, Current methods in medical image segmentation. Annual Review of Biomedical Engineering. (2000), 2, pp. 315-337
13.N. Sharma, L.M. Aggarwal, Automated medical image segmentation technique. Journal of Medical Physics / Association of Medical Physicits of India. (2010), 35, pp. 35
14. S. Bauer, R. Wiest, L.P. Nolte, M.Reyes, A survey of MRI based medical image analysis for brain tumor studies. Physics in Medicine and Biology. (2013), 58 (13), pp. R97
15. A. Usinkas, R.A. Dobrovolskis, B.F. Tomandl, Ischemic stroke segmentation on CT images using joint feature. Informatica. (2004), 15 (2), pp. 283-290
16. A.V. Dalca, R. Sridharan, L. Cloonan, K.M. Fitzpatrick, A. Kanakis, K.L. Furie, J. Rosand, O. Wu, M. Sabuncu, N.S. Rost, P. Golland, Segmentation of cerebrovascular pathologies in stroke patients with spatial and shape priors. Med Image Comput Comput Assist Interv. (2014), 17(0 2), pp. 773–780
17. H. Yang, L. Zhao, S. Tang, Y. Wang, Survey on brain tumor segmentation method. IEEE International Conference on Medical Imaging Physics and Engineering (ICMIPE). (2013), pp. 140-145
18. [online]. Available: http://en.wikipedia.org/wiki/Region_growing
19. K-P. Wong, Medical image segmentation: Methods and application in functional imging. Handbook of Biomedical Image Analysis, Springer. (2005), pp. 111-182
20. P. Suetens, R. Verbeeck, D. Delaere, J. Nuyts, B. Bijnens, Model based image segmentation: Methods and applications. Springer. (1991), 44, pp. 3-24 21. H. Tek and B. B. Kimia, Shock-based reaction-diffusion bubbles for image segmentation, in Computer Vision,Virtual Reality and Robotics in Medicine. Springer. (1995), pp. 434-438.
22. T. McInerney and D. Terzopoulos, Deformable models in medical image analysis: A survey. Medical Image Analysis. (1996), 1, no. 2, pp. 91-108
23. A. Nyma, M. Kang, Y.K Kwon, C.H Kim, J.M Kim, A hybrid technique for medical image segmentation. Journal of Biomedicine and Biotechnology. (2012).
24. A. Masood, A.A Al-Jumaily, Fuzzy c-means thresholding based level set for automated segmentation of skin lesions. Journal of Signal and Information Processing. (2013), 4, pp. 66-71
25. Chunming Li, Chenyang Xu, Changfeng Gui, M.D. fox, Level set evolution without reinitialization: a new variational formulation. IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR). (2005), 1, pp. 430-436
26. J. Unamaheswari, G. Radhamani, An optimal approach for DICOM image segmentation based on fuzzy techniques. Springer, 2012, 166, pp. 799-807
27. Y. Yang, C. Zheng, P. Lin, Image thresholding via a modified fuzzy c-means algorithm. Springer. (2004), 3287, pp. 589-596
28. E. Abdel-Maksoud, M. Elmogy, R. Al-Awadi, Brain tumor segmentation based on a hybrid clustering technique. Egyptian Informatics Journal , ScienceDirect. (2015), 16 (1), pp. 71-81.
29.T. Saikumar, M. Nagarani, K. Yojana, B. Shashidhar, Image segmentation of an adaptive threshold algorithm using watershed transform and fuzzy c-means clustering on level set method. IEEE International Conference on Advances in Engineering , Science and Management (ICAESM). (2012), pp. 33-38
30. Y. Li, Q. Hu, J. Wu, Z. Chen, A hybrid approach to detection of brain hemorrhagic candidates from clinical head CT scans. IEEE Sixth International Conference on Fuzzy Systems and Knowledge Discovery (FKSD). (2009), 1, pp. 361-365