A Survey on Medical Image Segmentation

Send Orders for Reprints to [email protected] Current Medical Imaging Reviews, 2015, 11, 3-14

3

A Survey on Medical Image Segmentation Saleha Masood*, Muhammad Sharif, Afifa Masood, Mussarat Yasmin and Mudassar Raza Department of Computer Science, COMSATS Institute of Information Technology, Wah Cantt, Pakistan Abstract: Much work has been done in the field of Image segmentation but still there is a room for improvement. Medical image segmentation is a sub field of image segmentation in digital image processing that has many important applications in the prospect of medical image analysis and diagnostics. Here in this paper different approaches of medical image segmentation will be classified along with their sub fields and sub methods. Recent techniques proposed in each category will also be discussed followed by a comparison of these methods.

Keywords: Medical image segmentation, Modalities, Classifiers, Clustering, thresholding, Markov random field, Atlas guided methods, Bayesian method, Neural networks, Region growing, Deformable models. 1. INTRODUCTION Segmentation is a process in which an image is divided into several sub regions based on a specific feature in order to pick up a region of interest. Segmentation process has enormous applications in the medical field. In the field of research and development much work has been done to overcome the problems faced by the segmentation process and yet there is a need of more effective and efficient work. 1.1. Purpose of Medical Image Segmentation In the process of segmentation of a medical image, the details required by the segmentation process are highly dependent on clinical application of the problem [1]. The purpose of segmentation is to improve the process of visualization to handle the detection process more effectively and efficiently. Other reasons of medical image segmentation can be seen in (Fig. 1).

The analysis of functions of anatomy problems is carried out through the segmentation process [2]. It covers all the factors that influence the analysis of a disease. Through the process of segmentation one can analyze, diagnose, quantify, monitor and plan the navigation of a disease. 1.2. Basic Principles of Segmentation The process of segmentation is carried out on the basis of two central principles. These principles as shown in Figure 2 are classified on the basis of features that contain texture, intensity, sharpness of edges and all the significant features in this context [3].

Fig. (2). Basic principles of segmentation.

1.3. Problems in Medical Image Segmentation Segmentation of medical image faces many problems because of which the quality of segmentation process gets affected [4]. These problems can be analyzed in Figure 3 below.

Fig. (1). Purpose of segmentation.

*Address correspondence to this author at the Department of Computer Science, COMSATS Institute of Information Technology, Wah Cantt, Pakistan; E-mail: [email protected] 1875-6603/15 $58.00+.00

The problem of uncertainty arises when there is noise in the image which makes the classification of image difficult [5]. The reason is that intensity values of pixels are amended because of noise in the image. This alteration in the intensity ©2015 Bentham Science Publishers

4

Current Medical Imaging Reviews, 2015, Vol. 11, No. 1

Fig. (3). Problems in medical image segmentation.

values of pixels disturbs uniformity in the intensity range of image [6]. Noise can be in the image because of motion in the picture, blurring effect and lack of diverse features etc. The problem of partial volume averaging causes the issue of inconsistency in the intensity values of image pixels. So in order to handle this uncertainty in the medical diagnosis systems image segmentation is playing a vital role [7]. 1.4. AI Techniques and Medical Image Segmentation There is a need of automatic problem detection in the domain of medical image segmentation when the rate of making mistakes by the humans is increasing rapidly [8]. AI is a field that is closely related and has many applications in the prospect of medical image segmentation [9]. There are different reasons for using AI in medical image segmentation. These include: •

AI helps to analyze and classify data automatically using AI tools.

•

AI approaches are helpful for storage purpose.

•

AI techniques make it easier to retrieve data from a given medical image.

•

AI provides an optimal way out to analyze the information in order to solve a problem.

•

AI helps out in making effective decision in the field of medical image segmentation [10].

1.5. Modalities of Medical Image Segmentation Segmentation process has many applications in the medical field. There are different medical modalities which are handled through the segmentation process [11]. These modalities can be seen in Figure 4 below.

Masood et al.

Fig. (4). Medical image segmentation modalities.

MRI brain images. The reason is that these images contain high signal to noise ratio which requires enhancement and segmentation of image to find out the region of interest. Another issue regarding these images is that they contain a variety of resolution because of which segmenting the image with required level of contrast is a great problem. The main applications is this regard are extracting volume of brain, segmenting different issues in matter of grey, white cerebrospinal liquid and to outline precise brain formations [15]. 1.5.2. CT Segmentation process has many applications in the analysis of computed tomography images. The main use of segmentation process in this regard is in the analysis of bones, thoracic scans, and segmentation of heart, stomach, brain and liver images and demarcation of abdominal aortic aneurysms [16]. The contrast and resolution of these images is not as good as MRI images. Variety of methods is applicable in the segmentation process of CT images. These methods will be discussed later in this paper. 1.5.3. Ultrasound Ultrasound images are usually with high rate of imperfection which makes it difficult to segment out the region of interest accurately. This reason caused many methods inapplicable for the segmentation of ultrasound images. Regardless of this issue some work still has been done in this regard. In most of the cases manual segmentation is done but these images are also used for the estimation of motion involved together with identifying pathology by means of textural classifiers [17]. 1.5.4. Multimodal

Here in this section we will have a quick overview of medical modalities. These modalities are discussed in the prospect of reconstruction in [12] whereas detailed analysis of these modalities is presented in [13]. Similarly enhancement of these modalities is discussed in [14].

In this case different modalities are used simultaneously to discover a problem. The information provided by different modalities is utilized to segment out a specific region of interest. The problem with this modality is that it is not always possible to gather multimodal data. Another issue in this regard is that they mostly require alignment process [18].

1.5.1. MRI

1.5.5. Digital Mammography

If we analyze the applications of segmentation in the medical field we can say that most work is carried out on the

Detection of different tumors is basically carried out in the segmentation of digital mammography. Most common

A Survey on Medical Image Segmentation


5

Fig. (5). Methods of medical image segmentation.

methods utilized for the mammography segmentation are variations in the process of thresholding. 1.5.6. Chest Radiography In chest radiography, radiograph is projected to analyze or diagnose predicament of the chest area and its structure. The main problems under chest infection analysis are categorized as ABCDEF where each letter represents a different problem. 2. LITERATURE REVIEW Looking back at the history of methods and techniques proposed in the context of medical image segmentation, we can say that there is a great improvement in this regard. With the passage of time more effective and efficient processes have been found out as shown in Figure 5. Here in this section we will analyze different methods that have been developed and utilized in the process of medical image segmentation. Recent work will also be taken into consideration. 2.1. Thresholding Thresholding is one of the most common methods used for image segmentation. The reason is that it is the most effective way when we want to analyze the foreground context by eliminating the image background. The basic working of this method is dependent on the intensity values of pixels in the image. The foreground image in this case is classified by comparing it through a threshold value with the background image that classifies it as a foreground image if there is a difference in the intensity values. Additional operations are needed to eliminate noise factor from the image and to acquire more effective results in the process of segmentation [19]. In this case image is first converted into a binary image and then a defined threshold value is used which separates the different regions of image. Some recent work that has been done is this prospect can be analyzed in [20]. The research is basically an overview of image segmentation techniques based on thresholding process. The five techniques of thresholding discussed in the paper include Mean method, P-tile method, Histogram Dependent Technique (HDT), Edge Maximization Technique (EMT) and visual technique. Another thresholding approach

of medical image segmentation is presented in [21]. The approach makes use of watershed segmentation together with the texture based region merging method. The results achieved through this method are 92.2 % as compared to the previous manual segmentation approaches. SVM thresholding of medical images is described in [22]. The method is proposed for the oscillation detection from the chest x-ray image. Results obtained through the technique are feasible. The working is carried out by first filtering the image and segmenting through threshold algorithm. After that a kernel function is set in SVM to segment out the image precisely. Soft thresholding of medical images is proposed in [23]. The method works on the concept of membership function that classifies each image pixel to a different region through the image histogram. The achieved advantages through the method are that the method works automatically and does not require human interaction and all the hard decisions are postponed to final stage for the analysis of spatial operations. Spatial operations make the method more robust and effective. Mammogram segmentation is described in [24]. An overview of threshloding methods is presented and row by row threshold checking method is used to fix an average threshold value for the image. The method shows that by eliminating sharpness of edges within an image excellent results of segmentation can be obtained. 2.2. Region Growing In this process a region of interest is picked up through a predefined condition. The condition defined in this case is based on the information pro result achieved through the intensity or edge details of image. In this method an initial point is defined manually and then all points which are connected to that initial point having the same intensity values as that point are selected [25]. The main application of this method in the medical field is to depict the tumor regions. Region growing method cannot be utilized on its own. Additional operations are required to be performed before application of this method. The main disadvantage of this method is that it requires manual depiction of the initial point because of which there is a need to initialize an initial point for every region that is to be extracted. Now a brief analysis of some recent region growing methods in the medical image segmentation prospect will be

6


presented. A new method for the brain abnormalities segmentation is proposed in [26]. The method works on the concept of seed based region growing. It takes adult male and female MR images of different sizes from brain. Brain tissues and background are divided into different categories and receive different sized MR images as input. The results are promising and light images yield better results as compared to darker abnormalities images. A similar approach is given in [27]. The method is an automatic approach of region growing. Co-occurrence matrix is used that selects starting point for the seed. The advantage acquired from the method is that it reduces time factor for the manual post processing process. MRI brain tumor segmentation is described in [28]. The method focuses on gradients and variations along with the boundaries. Edge information is preserved through the use of anisotropic filter. Next mean variance and mean gradient of the boundary curve is calculated. The method is effective for categorizing timorous regions. Another similar ultrasound segmentation method is proposed in [29]. The method is an automatic approach for masses segmentation from ultrasound images. The method can be said optimal for the segmentation of ultrasound images because they preserve the spatial information and are insensitive to speckle noise. A hybrid approach of ultrasound image segmentation is presented in [30]. The method is a composition of two approaches; region growing and region merging. Effective results are achieved through the method. A byes based medical image segmentation approach is proposed in [31]. The method works by adjusting parameters during the region growing approach. It is a multi stage processing approach. The method yields effective results as it is insensitive to noise and decreases the computational time to a great extent. 2.3. Classifiers Classification of image is done in this method by deriving a feature space from the image. This feature space is then further divided into different regions depending upon the function being defined in the feature space. A feature space can be defined as it covers whole range of a defined function for the classification purpose. For instance image intensities can be an example of a defined function. Classifiers methods work on the basis of pattern recognition process. This method is also known as supervised method. The reason is that the data utilized in this case is trained which is segmented manually and further utilized for the automatic segmentation process [32]. Classifiers methods are further divided into various techniques which can be seen in Figure 6 below. 2.3.1. Maximum Like Hood (ML) / Byes Method ML method involves least risk and is capable of providing best results because of minimum error rate. The method requires that the features range included in the feature space should contain exact information of the probabilities of features. The method is not much applicable because accurate information is not often acquired [33]. 2.3.2. Nearest Neighbor It is a non-parametric method. The method works by training all data for the classification purpose. The main dis-

Masood et al.

Fig. (6). Classifiers Types.

advantage of method is that it involves high computational convolution in the process of classification. The error rate that can occur in a process is twice in this case as compared to the byes method [34]. 2.3.3. KN-Nearest Neighbor It is a non-parametric method. The method works by placing k different points in the feature space characterized by clustered objects. These points basically represent the centroids of initial groups. After placing the points each object is assigned to the cluster that has the contiguous centroid. After assigning the groups the position of each k point is altered again. This step is repeated until the movement of centroids becomes static. This repetition will formulate a metric to be calculated by separating the objects into diverse clusters. The main advantage of the method is that it requires no training summit of assurance in results. On the other hand low accuracy in classification and high required storage is a main disadvantage of the method [35]. 2.3.4. Parzen Window Parzen windows classifier is a method for non-parametric density assessment which is utilized in support of classification process. By means of a given kernel function, the method estimates a certain training set allocation through a linear blend of kernels based on the experimental spots. The probability distribution function in this case is approximated through the weighted common of numerous Gaussians. The performance and complexity of the technique is similar to KNN method [36]. 2.3.5. Neural Networks It is a non-parametric method that involves a multilayer perceptron which is used for learning purpose and is a standard supervised network. There are different types of neural networks; some are used for supervised learning and others for unsupervised learning. The process of training in this case can be time consuming but classification is quick [37]. 2.3.6. Decision Tree Decision Tree Classifier is an easy and commonly used method of classification. The decision tree classifier structures a succession of assessment queries and provisions in a tree formation. In the decision tree, the root and internal


nodes enclose feature assessment circumstances to detach evidences that encompass diverse distinctiveness. The entire terminal node is allocated a class tagged with Yes or No [38]. 2.4. Bayesian Approach Bayesian assessment theory is used for the classification purpose. The method mainly works by considering probability in the image to construct models based on the probability that is further utilized for the class assignment of voxels in the image. These voxels are treated as random variables in the image. There are four main approaches in the Bayesian category of image segmentation. These approaches are shown in Figure 7 below.


7

2.4.4. Expectation Maximization (EM) This is also a statistical approach used to figure out MAP or ML of parameters of a statistical model. This approach works on the basis of iterations. Here steps are performed in alterations; first estimation (E) step is performed followed by a maximization (M) step whose information is then utilized for the next E step and the process goes on [42]. 2.5. Clustering If we compare the functions of clustering and classifiers we can say that both are carrying out the same function with the difference in their way of working. The classifiers make use of training data to classify the image and thus are called supervised methods. Clustering approach contains unsupervised methods as it does not make use of training data. This inability of learning in clustering approach is compensated by iteratively dividing the image through the segmentation process and then illustrating the possessions of every division. In other words we can say that clustering techniques instruct themselves by means of existing statistics [43]. Clustering process is mainly suitable for applications where the intensities distributions of pixels in the image are fit detached. The main application of this method can be observed in the segmentation of MRI. There are two main methods of clustering approach which are commonly used for the segmentation of medical images. These methods are shown in Figure 8 below.

Fig. (7). Bayesian Approaches. K-Mean

2.4.1. Maximum a Posteriori (MAP) In Bayesian figures, MAP approximation is an approach of the posterior allotment. The MAP can be employed to acquire a summit approximation of an unnoticed measure taking place in the foundation of experimental information. It is intimately interrelated to Fisher's technique of maximum likelihood (ML) method although occupies an amplified optimization purpose which integrates a preceding allotment above the measure one desires to approximate. MAP evaluation can be observed as a regularization of ML evaluation [39]. 2.4.2. Markov Random Field (MAP) It basically makes use of undirected graph that determines the markov values of some arbitrary variables contained within a graphical model. MAP is quite similar to the Bayesian approach in view of representation. The only difference is that this approach is undirected whereas Bayesian method is comprised of directed graphs [40]. 2.4.3. Maximum Likehood (ML) This is a phenomenon in statistics that aims at providing estimation of parameters in a given statistical model. It is also regarded as a well-recognized estimation technique. In some scenarios it is also used to maximize the likelihood function when we are given with fixed amount of data together with its statistical model from where values are selected of the parameters that carry out overall job of maximization [41].

Clustering Approaches Fuzzy cMeans

Fig. (8). Clustering approaches.

2.5.1. K-means The clustering process in this case is carried out by iteratively calculating the mean of intensities values of each separated class or cluster of the image. And the segmentation is carried out by categorizing each pixel with the closest obtained mean of the image [44]. 2.5.2. Fuzzy c-means Segmentation through this process is carried out on the basis of fuzzy set premise. This process is also called generalization of k-means process. The difference between the two processes is that the points are categorized in separate classes in k-means process whereas fuzzy c-means permits the points to be connected with more than one class [45]. Now we will have a brief review of some recent work done in this regard. Segmentation of MR images using fuzzy c means approach is presented in [46]. A new version of fuzzy method is proposed that makes it possible to determine automatically required number of clusters for the segmenta-

8


tion process. The method makes use of statistical histogram information to achieve the task. The results showed that the method produces more accurate results as compared to the classic fuzzy mean approach. Another similar approach for MRI segmentation is proposed in [47]. In order to reduce the computational time for the segmentation process a new method is proposed in [48]. The method works based on the method BCFCM by introducing a new factor in it. The results show that it is a quick and optimal approach for the endoscopy of brain. Fuzzy means approach together with the dominated grey levels in the image can be observed in [49]. The method works by converting the image into grey level and reducing the noise by applying wavelet approach. Then image is clustered by taking into consideration central grey levels as base for clustering process. Similar approach is presented in [50]. MRI segmentation using fuzzy c-means and neural networks is presented in [51]. The results acquired through this combination are robust, fast and accurate even in the presence of high noise. An approach to improve the speed of segmentation process based on k-means is presented in [52]. The results showed that this approach is much fast and reliable and provides good quality as well. Volume based medical image segmentation using k-means approach is proposed in [53]. The process works by initially preprocessing the image to speed up rest of the processes. Then different clustering approaches are analyzed and a new approach is presented to accurately segment out 3D images and to speed up the segmentation process. Using trained k-means clustering approach for the MRI segmentation is described in [54]. LM-k-means approach is used to segment white and grey matter from the MR images. High precision is achieved through this method as compared to the classic k-means approach. A new approach for the segmentation of CT images is presented in [55]. The method works in three sections. The sections are grouped on the basis of detecting abnormal regions, cerebrospinal fluid and lastly the brain matters. Results acquired through the process are not outstanding. A most recent work in this regard can be seen in [56]. The method is basically a blood cell segmentation approach using the k-means and median cut approach. Initially best outcome of blood cell segmentation is analyzed through the k-means, fuzzy c-means and mean shift approach and then the median cut approach is applied. The results achieved through the process showed that it is better process for object segmentation when further feature extraction process is needed. A mountain based clustering approach for medical image segmentation is presented in [57]. The results achieved are effective for diagnosing many issues in the medical field. A combination of clustering approaches involving fuzzy cmeans, Bayesian method and user interaction is proposed in [51]. Medical image segmentation using k-means together with the watershed approach is presented in [57]. Medical image segmentation using fuzzy similarity relation is presented in [58]. 2.6. Deformable Methods Deformable method works on the basis of object boundaries. The features considered in view of image boundaries are the shape, smoothness and internal forces together with the external forces on the object under consideration [59]. All these factors influence the effectiveness of obtainable

Masood et al.

results. Closed curves and shapes in the image are utilized to outline the object boundaries. The process of outlining the boundary of an object is a closed curvature or plane that is initially positioned close to the preferred edge and later permitted to experience an iterative reduction progression. In order to keep the segmentation process smooth internal forces are derived within the image. The external forces are derived in order to originate a plane towards the preferred element in the image. The main advantage of these methods is the piece wise continuity. Deformable forces are mainly classified into two different categories as shown in Figure 9 below.

Fig. (9). Deformable Methods.

2.6.1. Parametric Deformable Models (Explicit) In the statistics of deformable models parametric models are the one that can be described using finite number of parameters. These methods are also called active contours and make use of parameters generated curves for the representation of shape model. Parametric models are further divided into two categories which are: •

Edge based methods

•

Region based methods

The methods in edge based category take edges information as image features for the segmentation process and are really responsive to the noise factor as any noise can alter the accurate information of edges [60]. The other category under parametric methods is region based methods that make use of different areas of image to segment out the image. In the model evaluation process the information of regions is not updated which makes it difficult to obtain any changes in the region features. The main drawback of these methods is that it is difficult to handle the topology changes in the anonymous entity segmentation [61]. 2.6.2. Non-Parametric Models (Implicit) These methods are also called geometric active contour methods. These methods are the level set approaches and are based on the concepts of convolution theory. In the process of defining curve for the segmentation a level set function is utilized together with the additional time aspect [62].The evaluation of curve in this case is independent of parametric values. The drawback of parametric models is handled through these models which allow automatic handling of


variations in the topological factors. These methods also make use of edge based and region based methods although their implementation process is different from that of parametric models [63]. Now we will have a short overview of some of the recent techniques proposed in this regard. In [64] we can analyze 3D medical image segmentation through the combination of Conditional Random Fields and Deformable Models. Results acquired through the approach are promising as compared to previously developed methods. Medical image segmentation using minimal path deformable model is presented in [65]. The approach is based on extracting the organ contours. The method was considered as one of the great achievements in the segmentation of various types of medical images. Similar work can be seen in [66]. Geometric deformable model for medical image segmentation is proposed in [67]. Medical image segmentation using the combination of genetic algorithm and non convex approach is presented in [68]. The typical gradient of deformable model in this case is replaced with the genetic algorithm as it is assumed that the genetic algorithm cannot provide optimal solutions but when combined with deformable models satisfactory results are achieved. Implicit shaped deformable model for medical image segmentation is presented in [69]. The method makes use of region based and statistical model based approaches to extract an object. The use of shape and appearance priors in the deformable models for medical image segmentation is presented in [70]. The main advantage acquired through the method is fast processing speed. A fuzzy non-parametric approach for medical image segmentation can be observed in [71]. The method provides more accuracy as compared to the tractography approach in this prospect. Medical image segmentation using local binary fitting approach and deformable models is presented in [72]. It has the main advantage of speeding up the curve evaluation process. A survey on deformable models is given in [73]. Similar work is also presented in [60]. Non parametric mixture model based medical image segmentation is described in [74]. 2.7. Atlas Guided Approaches Medical images segmentation based on Atlas guided approaches is a way of analyzing image through labeling a preferred structure or set of framework commencing images made through modalities of medical imaging. The main purpose of this approach is to lend a hand to radiologists in the discovery and identification of diseases. The working flow of approach is optimized by identifying significant anatomy in the medical images [75]. These approaches, as shown in Figure 10, are also called adaptable templates. The segmentation in this case is carried out by preparing an atlas using compiled information of anatomy. After the generation of atlas it is used as a reference structure for the segmentation of fresh images. These approaches consider registration problem to handle the segmentation process. Atlas wrapping is used for the segmentation process that works by mapping the generated atlas on the objected image [11]. The main application of these approaches is in the images where there is no well-defined relation between image pixels and regions. The other main applications include its use in


9

clinical practice and computer aided diagnosis to analyze shape and morphological differences between image regions.

Fig. (10). Atlas guided approaches.

2.7.1. Atlas as Average Shape Here in this category we have two approaches which are: 2.7.1.1. Active Shape Model Active shape model makes use of Principal Component Analysis (PCA) and a pre constructed shape model for the segmentation of medical images. All shapes are trained and aligned and are used with PCA for the segmentation purpose [76]. The working is carried out by using average mean shape for the scan and then deformation is carried out by means of deformable models. The process is carried out through iterations and in each iteration, previous utilized curve or shape is used to measure the desired target object. The parameters of shape also remain unchanged. By means of this process only the desired deformations are permitted and process terminates when variations are faced in the shape model [77]. 2.7.1.2. Active Appearance Model (AAM) The active appearance model works in the same way as ASM with the difference that together with the shape model it also makes use of intensity model for the segmentation purpose. The intensity model is generated through registration among the training statistics [78]. 2.7.2. Atlas as Individual Image The process works by manual segmentation of anatomical structures originated from a reference image which creates a spatial atlas or map. The registration of reference image is done on the atlas for the purpose of automatic segmentation. The mapping among two should be coordinated. After that intensity correspondence evaluation is carried out using different approaches. The most used methods in this case are cross correlation and mutual information for registration. The smoothness factor is handled through Gaussian or elastic model [79]. Now we will have a brief overview of some recent approaches in this regard. A latest work in this view can be analyzed in [80]. The approach works on the basis of graph cut method by combining it with Active Appearance model (AAM). Three main steps of the approach include model building, object recognition and lastly the delineation. The method was tested for the segmentation of liver, kidney and

10


spleen images. Overall obtained results contain an accuracy of 94.3%. Segmentation of spinal images through AAM can be analyzed in [81]. A different work that is carried out in this approach makes use of combination of ICA and AAM to segment out the spinal images. The results showed that this approach contains more accuracy than the traditional PCAAAM approach. The part of Active Shape model in this prospect can be analyzed in [80]. Medical image segmentation using statistical shape model is presented in [82]. The approach works by combining the point distribution fraction with two dimensional PCA to segment out medical images. The results showed improvement as compared to the traditional approach. Segmentation of pelvic x-ray images through splines and shape modes is proposed in [83]. The deformation process is enhanced in this case by combining the active shape model with cubic spline interpolation approach due to low resolution and blurring effect within the xray images. Results showed improvement even in the presence of fracture. 3D active shape model for MRI image segmentation is carried out in [84]. Another 3D segmentation of MRI using level set approach is presented in [85]. The method includes 3D filtering followed by 3D segmentation of MR images. Graph cut based DT-MRI segmentation can be analyzed in [86]. Similar work is carried out in [87]. Segmentation of liver tumor using graph cuts and watershed approach is presented in [88]. Atlas registration based cerebellum MRI segmentation can be analyzed in [89]. 2.8. Edge Based Approaches

Masood et al.

image are alike such as in disguise images; stronger compassion and minor quantization are necessary. 2.10. Other Techniques There are numerous other techniques proposed and developed in the view of medical image segmentation. Some of these methods include watershed algorithm, model fitting, Partial differential equation-based methods, Split-and-merge methods, Fast Marching methods and Multi-scale segmentation. Some of the techniques presented in this regard can be analyzed here. Medical image segmentation using the watershed transform is presented in [93]. The method works on the basis of morphological operations to extract objects from the medical images. Improved watershed transforming process can be analyzed in [94]. Medical image segmentation using feature based GVF snake is presented in [95]. 3. ANALYSIS AND DISCUSSION After having an overview of basic approaches in the medical image segmentation process, it can be noticed that with advancements in the process of search and development new effective and efficient approaches are coming into existence. The methods classification according to the generation can be analyzed to understand this point. Approaches that exist in this regard are categorized into three generations; 1st, 2nd and 3rdrespectively as shown in Figure 11.

These approaches are the most common way of detecting discontinuity and boundaries of objects with in an image. The method is the most common way of detecting pixels with same intensity level of an object. In this case the two connected pixels have same intensity distribution form the edge and it is not essential that they will form a closed path [90]. The distinction between the pixels is this case is carried out by estimating the intensity gradient. These methods are mainly used as base or central technique for other segmentation approaches [91]. 2.9. Compression Based Approaches Compression centered techniques assume that the best possible segmentation is the one that reduces the excess of every achievable segmentation and the development period of statistics. The association connecting these two conceptions is that segmentation attempts to discover samples in an image and reliability in the image might be utilized to compress it. The technique explains every division by means of its surface and edge outline [92]. On behalf of any specified segmentation of an image, this method gives the amount of bits essential to predetermine that the image is centered on known segmentation. Consequently, surrounded by any achievable segmentation of an image, the aim is to discover segmentation that generates the straight coding span. This might be accomplished through an easy clustering technique. The deformation inside the lossy compression decides the roughness of segmentation and its best assessment could be different for every image. This limitation can be projected commencing the disparity of consistency in an image. For instance, when the textures in an

Fig. (11). Generations of medical segmentation approaches.

Development and progress in the field of medical image segmentation can be analyzed by placing the methods into different groups. This classification will show us the research and development carried out in this regard. First generation group contains methods which require little prior information for processing the image hence involves low level techniques. With the passage of time and advancement in technology some new and more effective methods came into existence; 2nd generation occupies methods based on optimization, image and uncertainty models. 3rd generation techniques are highly dependent on prior information of the image and require experts defined models and rules for classification of an image. Discussion can also be made on the comparison of different methods given in Table 1. This will help us to select the best method for a given situation.



11

Table 1. Comparison of medical image segmentation methods. Serial #

Method

Advantages

Limitations

Applications

Memory Usage

1

Thresholding

These methods are fastest, simplest and easiest to implement.

These methods are responsive to artifacts and piecewise continuity is not assured by them.

They are mainly applicable to structures that have divided intensity allotment.

Fastest

2

Region Growing

These methods assure the piecewise continuity and are less sensitive to noise.

Position of the start point and blurring affects are the main limitations of these methods.

Work well for the structures with high contrast boundaries.

Fast

3

Clustering

These methods are easy to implement and can also be used as starting point for other approaches.

They require a spatial constraint to perform well.

Mainly work well for MR images; not capable to handle CT images.

Medium

4

Classifiers

Most widely used approaches for segmentation process.

They are computationally complex and slow approaches.

Work well for MR and CT images.

Slow

5

Bayesian approach

Helps to combine the prior available information with the given data and offers a suitable situation for an extensive variety of models.

Does not provide proper ways to figure out a prior and makes use of posterior distributions which are highly dependent on priors.

These are mainly applicable to the verification problems.

Fast

6

Deformable Methods

They effectively handle the topological changes and assure the piecewise continuity. These methods are noise insensitive and provide sub-pixel accuracy.

Requires the tuning of parameters and thus can affect speed of the system.

Work best with statistical regional information of the image.

Medium

7

Atlas guided Approaches

These approaches are fact and assure an optimum solution for two class segmentation.

Precise segmentation of difficult composition is itself difficult.

They are mainly applied to MR images.

Slow

8

Edge based methods

They are easy to implement and offer effective computational factor.

Not appropriate to figure out all kinds of problems.

Can be applied in all modalities of medical image segmentation.

Fast

9

Compression based methods

They provide an advantage of less storage consumption.

These approaches are relatively slow.

These approaches have main applications for MR and CT images.

Medium

CONCLUSION This paper provides a brief overview of some methods and techniques available under the umbrella of medical image segmentation. Medical field is comprised of a number of medical modalities and each one of them contains a number of diseases and issues under its heading. So this paper is basically analyzing the techniques proposed and implemented in all of these modalities to help medical field in view of analyzing or figuring out a particular problem. Each method has its own pros and cons. The usage of each method depends on the type of application built together with the resources available. Although much research work has been done in this regard but we can still say that there is huge room available for more efficient and effective techniques.

ACKNOWLEDGEMENTS Declared none. REFERENCES [1]

[2]

[3]

[4]

CONFLICT OF INTEREST The authors confirm that this article content has no conflict of interest.

[5]

Zuva T, Oludayo OO, Ojo SO, Ngwira SM. Image segmentation, available techniques, developments and open issues. Canadian Journal on Image Processing and Computer Vision 2011; 2(3): 209. Sharma N, Aggarwal LM. Automated medical image segmentation techniques, Journal of medical physics/Association of Medical Physicists of India 2010; 35(1): 3. Costin H. A Fuzzy Rules-Based Segmentation Method for Medical Images Analysis, International Journal of Computers, Communications & Control 2013;8(2). Shrimali V, Anand R, Kumar V. Current trends in segmentation of medical ultrasound b-mode images: a review. IETE technical review 2009; 26(1): 8. Birkfellner W. Applied medical image processing: a basic course: Taylor & Francis; 2014.

12 [6]

[7] [8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

Current Medical Imaging Reviews, 2015, Vol. 11, No. 1 Al-amri SS, Kalyankar NV, Khamitkar SD. A Comparative Study of Removal Noise from Remote Sensing Image, International Journal of Computer Science Issues (IJCSI) 2010; 7(1) He BS, Zhu F, Shi YG. Medical Image Segmentation, Advanced Materials Research 2013; 760: 1590-3. Rastgarpour M, Shanbehzadeh J, editors. Application of AI Techniques in Medical Image Segmentation and Novel Categorization of Available Methods and. Tools, Proceedings of the International Multi Conference of Engineers and Computer Scientists 2011 Vol I, IMECS 2011, March 16-18, 2011, Hong Kong; 2011: Citeseer. Olabarriaga SD, Smeulders AW. Interaction in the segmentation of medical images: A survey. Medical image analysis 2001;5(2): 12742. Hu Y, Grossberg M, Mageras G. Survey of recent volumetric medical image segmentation techniques. Biomed Eng. 2009: 32146. Pham DL, Xu C, Prince JL. Current methods in medical image segmentation 1, Annual review of biomedical engineering 2000; 2(1): 315-37. Yasmin M, Sharif M, Masood S, Raza M, Mohsin S. Brain image reconstruction: A short survey. World Applied Sciences Journal 2012; 19(1): 52-62. Yasmin M, Mohsin S, Sharif M, Raza M, Masood S. Brain Image Analysis: A Survey, World Applied Sciences Journal. 2012; 19(10): 1484-94. Yasmin M, Sharif M, Masood S, Raza M, Mohsin S. Brain image enhancement-A survey. World Applied Sciences Journal, 2012; 17(9): 1192-204. Masood S, Sharif M, Yasmin M, Raza M, Mohsin S. Brain Image Compression: A Brief Survey, Research Journal of Applied Sciences. 2013; 5 Xiao-juan C, Dan L, editors. Medical image segmentation based on threshold SVM. Biomedical Engineering and Computer Science (ICBECS), 2010 International Conference on; 2010: IEEE. Aja-Fernandez S, Vegas-Sanchez-Ferrero G, Fernandez M, editors. Soft thresholding for medical image segmentation, Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE; 2010: IEEE. Shahzad A, Sharif M, Raza M, Hussain K. Enhanced watershed image processing segmentation, Journal of Information & Communication Technology. 2008; 2(1): 01-9. Abubaker A, Qahwaji R, Aqel MJ, Saleh MH, editors. Average Row Thresholding Method for Mammogram Segmentation. Engineering in Medicine and Biology Society, 2005 IEEE-EMBS 2005 27th Annual International Conference of the; 2006: IEEE. Li J, Zhu S, Bin H. [Medical image segmentation techniques]. Sheng wu yi xue gong cheng xue za zhi= Journal of biomedical engineering= Shengwu yixue gongchengxue zazhi. 2006; 23(4): 8914. Ng H, Huang S, Ong S, Foong K, Goh P, Nowinski W, editors. Medical image segmentation using watershed segmentation with texture-based region merging, Engineering in Medicine and Biology Society, 2008 EMBS 2008 30th Annual International Conference of the IEEE; 2008: IEEE. Kotropoulos C, Pitas I. 22 Segmentation of ultrasonic images using support vector machines. Pattern Recognition Letters, 2003; 24(4): 715-27. Khare A, Tiwary US. Soft-thresholding for de noising of medical images—a multi resolution approach. International Journal of Wavelets, Multi resolution and Information Processing 2005; 3(04): 477-96. Cascio D, Fauci F, Magro R, Raso G, Bellotti R, De Carlo F, et al. Mammogram segmentation by contour searching and mass lesions classification with neural network. Nuclear Science, IEEE Transactions on 2006; 53(5): 2827-33. Haider W, Sharif M, Raza M. Achieving accuracy in early stage tumor identification systems based on image segmentation and 3D structure analysis. Computer Engineering and Intelligent Systems 2011; 2(6): 96-102. Siddique I, Bajwa IS, Naveed MS, Choudhary MA, editors. Automatic Functional Brain MR Image Segmentation using Region Growing and Seed Pixel, International Conference on Information & Communications Technology, December; 2006. Oghli MG, Fallahi A, Pooyan M, editors. Automatic region growing method using GSmap and spatial information on ultrasound

Masood et al.

[28]

[29]

[30]

[31]

[32]

[33] [34]

[35] [36]

[37]

[38]

[39]

[40] [41] [42]

[43]

[44]

[45]

[46]

[47]

[48]

[49]

images, Electrical Engineering (ICEE), 2010 18th Iranian Conference on; 2010: IEEE. Deng W, Xiao W, Deng H, Liu J, editors. MRI brain tumor segmentation with region growing method based on the gradients and variances along and inside of the boundary curve. Biomedical Engineering and Informatics (BMEI), 2010 3rd International Conference on; 2010: IEEE. Poonguzhali S, Ravindran G, editors. A complete automatic region growing method for segmentation of masses on ultrasound images, Biomedical and Pharmaceutical Engineering, 2006 ICBPE 2006 International Conference on; 2006: IEEE. Guan H, Li D-y, Lin J-l, Wang T-F, editors. Segmentation of Ultrasound Medical Image Using A Hybrid Method. Complex Medical Engineering, 2007 CME 2007 IEEE/ICME International Conference on; 2007: IEEE. Pan Z, Lu J. A Bayes-based region-growing algorithm for medical image segmentation, Computing in science & Engineering 2007; 9(4): 32-8. Bezdek JC, Hall L, Clarke L. Review of MR image segmentation techniques using pattern recognition, MEDICAL PHYSICSLANCASTER PA- 1993; 20: 1033 Le Cam L. Maximum likelihood: an introduction. International Statistical Review 1990; 58(2): 153-71. Arya S, Mount DM, Netanyahu NS, Silverman R, Wu AY. An optimal algorithm for approximate nearest neighbor searching fixed dimensions. Journal of the ACM (JACM) 1998; 45(6): 891-923. Liao Y, Vemuri VR. Use of K-nearest neighbor classifier for intrusion detection, Computers & Security 2002; 21(5): 439-48. Yeung D-Y, Chow C, editors Parzen-window network intrusion detectors, Pattern Recognition, 2002 Proceedings 16th International Conference on; 2002: IEEE. Song T, Jamshidi MM, Lee RR, Huang M. A modified probabilistic neural network for partial volume segmentation in brain MR image Neural Networks, IEEE Transactions on. 2007; 18(5): 142432. Safavian SR, Landgrebe D. A survey of decision tree classifier methodology, IEEE transactions on systems, man, and cybernetics 1991; 21(3): 660-74. Wells III WM, Grimson WEL, Kikinis R, Jolesz FA. Adaptive segmentation of MRI data, Medical Imaging, IEEE Transactions on 1996; 15(4): 429-42. Li SZ, Singh S. Markov random field modeling in image analysis: Springer; 2009. King G. Unifying political methodology: The likelihood theory of statistical inference: University of Michigan Press; 1989. Yan M, Bui AA, Cong J, Vese LA. General convergent expectation maximization (EM)-type algorithms for image reconstruction. Inverse Problems & Imaging 2013; 7(3) Saikumar T, FasiUddin K, Reddy BV, Uddin MA. Image Segmentation Using Variable Kernel Fuzzy C Means (VKFCM) Clustering on Modified Level Set Method. Computer Networks & Communications (NetCom): Springer; 2013. p. 265-73. Lee TH, Fauzi MFA, Komiya R, editors. Segmentation of CT brain images using K-means and EM clustering. Computer Graphics, Imaging and Visualisation, 2008 CGIV'08 Fifth International Conference on; 2008: IEEE. Muda TZT, Salam RA, editors Blood cell image segmentation using hybrid k-means and median-cut algorithms, Control System, Computing and Engineering (ICCSCE), 2011 IEEE International Conference on; 2011: IEEE. Li M, Huang T, Zhu G, editors. Improved fast fuzzy C-Means algorithm for medical MR images segmentation. Genetic and Evolutionary Computing, 2008 WGEC'08 Second International Conference on; 2008: IEEE. Ozyurt O, Dincer A, Ozturk C, editors. Brain MR Image segmentation with fuzzy C-means and using additional shape elements, Biomedical Engineering Meeting, 2009 BIYOMUT 2009 14th National; 2009: IEEE. Szilagyi L, Benyo Z, Szilágyi SM, Adam H, editors. MR brain image segmentation using an enhanced fuzzy c-means algorithm, Engineering in Medicine and Biology Society, 2003 Proceedings of the 25th Annual International Conference of the IEEE; 2003: IEEE. Balafar M, Ramli AR, Iqbal Saripan M, Mahmud R, Mashohor S, editors. Medical image segmentation using fuzzy C-mean (FCM) and dominant grey levels of image, Visual Information Engineering, 2008 VIE 2008 5th International Conference on; 2008: IET.

A Survey on Medical Image Segmentation [50]

[51]

[52]

[53]

[54]

[55]

[56] [57]

[58]

[59]

[60] [61]

[62] [63] [64]

[65]

[66]

[67]

[68]

[69]

[70]

[71]

Birgani PM, Ashtiyani M, Asadi S, editors. MRI segmentation using fuzzy c-means clustering algorithm basis neural network, Information and Communication Technologies: From Theory to Applications, 2008 ICTTA 2008 3rd International Conference on; 2008: IEEE. Balafar MA, Ramli AR, Iqbal Saripan M, Mashohor S, editors. Medical image segmentation using fuzzy C-mean (FCM), Bayesian method and user interaction, Wavelet Analysis and Pattern Recognition, 2008 ICWAPR'08 International Conference on; 2008: IEEE. Nandagopalan S, Dhanalakshmi C, Adiga B, Deepak N, editors A fast k-means algorithm for the segmentation of echocardiographic images using DBMS-SQL. Computer and Automation Engineering (ICCAE), 2010 the 2nd International Conference on; 2010: IEEE. Li X, editor. A volume segmentation algorithm for medical image based on k-means clustering. Intelligent Information Hiding and Multimedia Signal Processing, 2008 IIHMSP'08 International Conference on; 2008: IEEE. Kumbhar AD, Kulkarni A, editors. Magnetic resonant image segmentation using trained k-means clustering. Information and Communication Technologies (WICT), 2011 World Congress on; 2011: IEEE. Rathnayaka K, Sahama T, Schuetz MA, Schmutz B. Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions. Medical engineering & physics 2011; 33(2): 22633. N, Ramakrishnan A, editors. Blood Cell Segmentation Using Sinha EM Algorithm ICVGIP; 2002 Ng H, Ong S, Foong K, Goh P, Nowinski W, editors. Medical image segmentation using K-means clustering and improved watershed algorithm Image Analysis and Interpretation, 2006 IEEE Southwest Symposium on; 2006: IEEE. Tabakov M, editor A fuzzy clustering technique for medical image segmentation. Evolving Fuzzy Systems, 2006 International Symposium on; 2006: IEEE. Metaxas DN. Physics-based deformable models: applications to computer vision, graphics, and medical imaging: Kluwer Academic Publishers; 1996. Xu C, Pham DL, Prince JL. Image segmentation using deformable models, Handbook of medical imaging 2000; 2: 129-74 Hegadi R, Kop A, Hangarge M. A survey on deformable model and its applications to medical imaging, IJCA Special Issue on Recent Trends in Image Processing and Pattern Recognition 2010: 64-75. Sarkar P, Chakrabarti D, Jordan M. Non-parametric Link Prediction. arXiv preprint arXiv: 11091077. 2011. Heinz D. Hyper markov non-parametric processes for mixture modeling and model selection: Carnegie Mellon University; 2010. Tsechpenakis G, Metaxas D. CoCRF deformable model: a geometric model driven by collaborative conditional random fields. Image Processing, IEEE Transactions on 2009; 18(10): 2316-29. Yan P, Kassim AA. Medical image segmentation using minimal path deformable models with implicit shape priors, Information Technology in Biomedicine, IEEE Transactions on 2006;10(4): 677-84. Yan P, Kassim AA, editors. Medical image segmentation with minimal path deformable models Image Processing, 2004 ICIP'04 2004 International Conference on; 2004: IEEE. Lee M, Park S, Cho W, Kim S, Jeong C. Segmentation of medical images using a geometric deformable model and its visualization, Electrical and Computer Engineering, Canadian Journal of. 2008; 33(1): 15-9. McIntosh C, Hamarneh G. Medial-based deformable models in nonconvex shape-spaces for medical image segmentation. Medical Imaging, IEEE Transactions on 2012; 31(1): 33-50. Farzinfar M, Teoh EK, Xue Z, editors. A coupled implicit shapebased deformable model for segmentation of MR images, Control, Automation, Robotics and Vision, 2008 ICARCV 2008 10th International Conference on; 2008: IEEE. El-Baz A, Gimel'farb G, editors. Image segmentation with a parametric deformable model using shape and appearance priors, Computer Vision and Pattern Recognition, 2008 CVPR 2008 IEEE Conference on; 2008: IEEE. Awate SP, Zhang H, Gee JC. A fuzzy, nonparametric segmentation framework for DTI and MRI analysis: with applications to DTItract extraction. Medical Imaging, IEEE Transactions on 2007; 26(11): 1525-36.

Current Medical Imaging Reviews, 2015, Vol. 11, No. 1 [72]

[73]

[74]

[75]

[76]

[77] [78]

[79]

[80]

[81]

[82]

[83]

[84]

[85]

[86]

[87]

[88]

[89]

[90]

[91]

[92]

[93]

13

Nakhjavanlo BB, Ellis T, Raoofi P. Medical Image Segmentation Using Deformable Model and Local Fitting Binary: Thoracic Aorta. Sonka M, Fitzpatrick JM, editors. Handbook of medical imaging (Volume 2, Medical image processing and analysis) 2000, SPIEThe international society for optical engineering Joshi N, Brady M. Non-parametric mixture model based evolution of level sets and application to medical images. International journal of computer vision 2010; 88(1): 52-68 Kasiri K, Kazemi K, Dehghani MJ, Helfroush MS, editors. Atlasbased segmentation of brain MR images using least square support vector machines. Image Processing Theory Tools and Applications (IPTA), 2010 2nd International Conference on; 2010: IEEE. Van Ginneken B, Frangi AF, Staal JJ, ter Haar Romeny BM, Viergever MA. Active shape model segmentation with optimal features, Medical Imaging, IEEE Transactions on 2002; 21(8): 924-33. NUS S. ACTIVE SHAPE MODELS, Computer Vision and Image Understanding. 2012; 16: N1 Cootes TF, Edwards GJ, Taylor CJ. Active appearance models, IEEE Transactions on pattern analysis and machine intelligence. 2001; 23(6): 681-5. Chang P-L, Teng W-G, editors. Exploiting the self-organizing map for medical image segmentation, Computer-Based Medical Systems, 2007 CBMS'07 Twentieth IEEE International Symposium on; 2007: IEEE. Chen X, Udupa JK, Bagci U, Zhuge Y, Yao J. Medical image segmentation by combining graph cuts and oriented active appearance models. Image Processing, IEEE Transactions on 2012; 21(4): 2035-46. Zhan S, Chang H, Jiang J-g, Li H, editors. Spinal images segmentation based on improved active appearance models. Bioinformatics and Biomedical Engineering, 2008 ICBBE 2008 the 2nd International Conference on; 2008: IEEE. Neumann A, Lorenz C. Statistical shape model based segmentation of medical images. Computerized Medical Imaging and Graphics 1998; 22(2): 133-43. Smith R, van Najarian K, editors. Splines and Active Shape Model for segmentation of pelvic x-ray images. Complex Medical Engineering, 2009 CME ICME International Conference on; 2009: IEEE. Rousson M, Paragios N, Deriche R. Implicit active shape models for 3D segmentation in MR imaging. Medical Image Computing and Computer-Assisted Intervention–MICCAI 2004: Springer; 2004. p. 209-16. Baillard C, Hellier P, Barillot C. Segmentation of brain 3D MR images using level sets and dense registration. Medical image analysis 2001; 5(3): 185-94. Weldeselassie YT, Hamarneh G, editors. DT-MRI segmentation using graph cuts. Medical Imaging; 2007: International Society for Optics and Photonics. Stawiaski J, Decenciere E, Bidault F, editors. Interactive liver tumor segmentation using graph-cuts and watershed, Workshop on 3D Segmentation in the Clinic: A Grand Challenge II Liver Tumor Segmentation Challenge MICCAI, New York, USA; 2008. Peng B, Zhang L, Zhang D. A survey of graph theoretical approaches to image segmentation, Pattern Recognition 2013; 46(3): 1020-38. van der Lijn F, de Bruijne M, Hoogendam YY, Klein S, Hameeteman R, Breteler MM, et al., editors. Cerebellum segmentation in MRI using atlas registration and local multi-scale image descriptors, Biomedical Imaging: From Nano to Macro, 2009 ISBI'09 IEEE International Symposium on; 2009: IEEE. Li BN, Chui CK, Chang S, Ong SH. Integrating spatial fuzzy clustering with level set methods for automated medical image segmentation, Computers in Biology and Medicine. 2011; 41(1): 1-10. Senthilkumaran N, Rajesh R. Edge detection techniques for image segmentation–a survey of soft computing approaches, International Journal of Recent Trends in Engineering. 2009; 1(2): 250-4. Bhattacharyya S. A brief survey of color image preprocessing and segmentation techniques, Journal of Pattern Recognition Research 2011; 1(1): 120-9. Jia-xin C, Sen L, editors. A medical image segmentation method based on watershed transform, Computer and Information Technology, 2005 CIT 2005 The Fifth International Conference on; 2005: IEEE.

14 [94]


Masood et al.

Li S, Wu L, Sun Y, editors. Cell image segmentation based on an improved watershed transformation. Computational Aspects of Social Networks (CASoN), 2010 International Conference on; 2010: IEEE.

Received: November 17, 2014

[95]

Harini R, Chandrasekar C, editors. Image segmentation using nearest neighbor classifiers based on kernel formation for medical images. Pattern Recognition, Informatics and Medical Engineering (PRIME), 2012 International Conference on; 2012: IEEE.

Revised: January 21, 2015

Accepted: January 27, 2015