An effective method for computerized prediction and segmentation of multiple sclerosis lesions in brain MRI

Accepted Manuscript

An Effective Method for Computerized Prediction and Segmentation of Multiple Sclerosis Lesions in Brain MRI Sudipta Roy , Debnath Bhattacharyya , Samir Kumar Bandyopadhyay , Tai-Hoon Kim PII: DOI: Reference:

S0169-2607(16)31450-X 10.1016/j.cmpb.2017.01.003 COMM 4332

To appear in:

Computer Methods and Programs in Biomedicine

Received date: Accepted date:

25 December 2016 6 January 2017

Please cite this article as: Sudipta Roy , Debnath Bhattacharyya , Samir Kumar Bandyopadhyay , Tai-Hoon Kim , An Effective Method for Computerized Prediction and Segmentation of Multiple Sclerosis Lesions in Brain MRI, Computer Methods and Programs in Biomedicine (2017), doi: 10.1016/j.cmpb.2017.01.003

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Highlights



AC

CE

PT

ED

M

AN US

  

Accurate detection and segmentation of Multiple sclerosis (MS) diseases with lesions positions identification Adaptive background generation and binarization using global threshold are the key steps for MS lesions detection Evaluates performance with other recent method Proposed method produce good results visually as well as metrically Proposed method reduced the under segmentation, over segmentation, and spurious lesions generation

CR IP T



ACCEPTED MANUSCRIPT

An Effective Method for Computerized Prediction and Segmentation of Multiple Sclerosis Lesions in Brain MRI

CR IP T

Sudipta Roy 1, Debnath Bhattacharyya2, Samir Kumar Bandyopadhyay 3, Tai-Hoon Kim 4 Department of Computer Science and Engineering, 4 Department of Convergence Security 1 Institute of Computer Technology (UVPCE), Ganpat University, Sola, Ahmadabad-380060, Gujarat, India 2 Vignan‟s Institute of Information Technology, Visakhapatnam-530049, AP, India 3 Calcutta University Technology Campus, JD-2, Sector-III, Salt Lake, Kolkata-98, India 4 Sungshin Women's University, 249-1, Dongseon-dong 3-ga, Seoul, 136-742, Korea 1,2, 3

[email protected] 1, [email protected] 2, [email protected] 3, [email protected] 4

AN US

Abstract

AC

CE

PT

ED

M

Background and Objectives: Multiple sclerosis is one of the major diseases and the progressive MS lesion formation often leads to cognitive decline and physical disability. A quick and perfect method for estimating the number and size of MS lesions in the brain is very important in estimating the progress of the disease and effectiveness of treatments. But, the accurate identification, characterization and quantification of MS lesions in brain magnetic resonance imaging (MRI) is extremely difficult due to the frequent change in location, size, morphology variation, intensity similarity with normal brain tissues, and inter-subject anatomical variation of brain images. Methods: This paper presents a method where adaptive background generation and binarization using global threshold are the key steps for MS lesions detection and segmentation. After performing three phase level set, we add third phase segmented region with contour of brain to connect the normal tissues near the boundary. Then remove all lesions except maximum connected area and corpus callosum of the brain to generate adaptive background. The binarization method is used to select threshold based on entropy and standard deviation preceded by non-gamut image enhancement. The background image is then subtracted from binarized image to find out segmented MS lesions. Results: The step of subtraction of background from binarized image does not generate spurious lesions. Binarization steps correctly identify the MS lesions and reduce over or under segmentation. The average Kappa index is 94.88%%, Jacard index is 90.43%, correct detection ration is 92.60284 %, false detection ratio is 2.55% and relative area error is 5.97% for proposed method. Existing recent methods does not have such accuracy and low value of error rate both mathematically as well as visually due to many spurious lesions generation and over segmentation problems. Conclusions: Proposed method accurately identifies the size and number of lesions as well as location of lesions detection as a radiologist performs. The adaptability of the proposed method creates a number of potential opportunities for use in clinical practice for the detection of MS lesions in MRI. Proposed method gives an improved accuracy and low error compare to existing recent methods.

Keywords: Binarization, Brain MRI, Level Set, Multiple Sclerosis, Lesion Segmentation, Normal Tissues, Performance Evaluation

1.

Introduction

Multiple sclerosis (MS) is a central nervous system (CNS) disease that damages to the insulating myelin sheaths around the axons in the brain. MS causes the immune system to attack these nerve fibers. The rate of progress of MS varies from person to person and can have periodic remission and relapse. The healthy brain contains White Matter (WM), Grey Matter (GM), and Cerebrospinal Fluid (CSF) [1-2]. There are three main types of segmentation approaches: manual, semi-automatic, and automatic. Manual segmentation is the base method for

ACCEPTED MANUSCRIPT

2.

Review Work

ED

M

AN US

CR IP T

lesion segmentation. An expert physician examines different modalities to select the lesion pixels. Unfortunately, the manual process is time consuming and somewhat subjective. Still, manual segmentations are considered the best results available and serve as the baseline for evaluating other methods. The expert segmentations can be considered a "silver standard" since they are not perfect representations of the ground truths, but provide the best in-vivo estimates available. Computer-aided methods do provide some benefit to MS lesion segmentation, where experts can have difficulty combining information from multiple MRI modalities and from multiple adjacent slices; welldesigned algorithms can efficiently blend this data. As a result, it is important to focus on the development of semiautomated and automated lesion segmentation methods. Automatic segmentation offers an attractive alternative to manual segmentation which remains a time-consuming task and suffers from intra-expert and inter-expert variability. However, the progression of the MS lesions shows considerable variability and MS lesions present temporal changes in shape, location, and area between patients and even for the same patient. The objective of research work included in this paper is to provide a robust technique for automatic segmentation of multiple sclerosis lesions from brain MR images. Most of segmentation techniques in the literature suffer from high false positives due to the similarity between MS lesions and the normal tissue and also due to basing the learning on pixels while the lesions form regions. Comprehensive study of false positive and negative in MS segmentation is needed with proposed segmentation technique or any other technique to provide more accurate and clinical friendly results. The proposed technique is designed mainly for MS lesions detection and the different tissues of the brain are not segmented. Proposed method has generates the background for each image. Three phase level set is the key idea to generate backgrounds for each image. Contour detection performed after artifact and skull removal image. Then we add the level set image and the contour image to generate adaptive background. To detect MS lesion a global threshold selection methodology has been done by the combination of entropy and standard deviation. Then the binary threshold generated image is subtracted by background image and finally the MS lesion is obtained. This approach captures the neighboring lesion properties and produces encouraging results, with a general improvement in the detection rate of lesions. The rest of the paper is organized as follows. Literature review on some existing method has been described in section 2. The proposed methodology has been described in section 3. Detailed results and comparison with some recent superior methods has been described in section 4. Finally, we conclude our method in section 5.

AC

CE

PT

Today more than 2,300,000 people around the world are affected in MS [3] disease. Symptoms vary widely including blurred vision, weak limbs, tingling sensations, unsteadiness and fatigue. For some people, MS is characterized by periods of relapse and remission while for others it has a progressive pattern. There exist some general categories for automated segmentation of MS lesions in MRI scans of the brain. The methods can be broadly categorized on the approach and grouped based on their implementation [4]. Conventional methods are limited by lack of pathological specificity and lack of sensitivity to grey matter lesions and to microscopic damage in normal appearing white matter, which can also be associated with other chronic inflammatory diseases of the CNS [5]. Focal cortical lesions are typically not seen on some conventional method as they are smaller in size and have poor contrast with the surrounding normal grey matter, in addition to partial volume effects from the CSF [6]. Lorenzo et. al. (2009) [7] proposed an automate multimodal graph cuts in order to automatically segment MS lesions in MRI which replace the manual interaction in order to discriminate between MS lesions and the normal appearing brain tissues. Evaluation is performed in synthetic and real images showing good agreement between the automatic segmentation and the target segmentation. It is observed with an example of semi-automatic edition of our automatic segmentation. When a lesion is missed, a user can add a seed, in this case a source seed, and the graph cuts are recomputed in few seconds. Derraz et. al. (2010) [8] proposed a semi-automatic segmentation based active contour model and statistic prior knowledge of MS Lesions that can find in regions of interest within brain MRI. The authors [8] showed a significant improvement of the proposed model but it suffers from over segmentation problem. But it is useful for interactive segmentation due to its high performance and the facility to add or remove training prototypes to improve the results. Forbes et. al. (2010) [9] proposed an augmented multi-sequence hidden Markov model that includes additional weight variables to account for the relative importance and control the impact of each sequence. The

ACCEPTED MANUSCRIPT

AC

CE

PT

ED

M

AN US

CR IP T

augmented framework has the advantage of allowing 1) the incorporation of expert knowledge on the a priori relevant information content of each sequence and 2) a weighting scheme which is modified adaptively according to the data and the segmentation task under consideration. The model, applied to the detection of multiple sclerosis and stroke lesions shows promising results. But investigating other settings, particularly in relation to targeting specific lesion types is limitation of that method [9]. Derraz et al. (2014)[10] proposed to develop and evaluate an automated lesion segmentation method based on Active Contours (AC) model incorporating tissue knowledge issued from T1-weighted and tissues distribution on FLAIR image. The GM and WM as well as CSF tissue classes issued from T1-weighted and the tissues intensities issued is used in order to determine an automatic outlier of each tissue class is used in order to detect outliers. Neither training nor thresholding is needed to performed the fully automatic segmentation based AC and outlier. Another advantage of our approach lies in the fact that it involves no thresholding step. Without explicit modeling, either soft or hard rejection, a predetermined threshold has to be used to decide the separation between the normal tissue and the outlier pixels. Since the thresholds are often data-dependent, manually chosen values tend to not work consistently across different data sets. Cabezas et al. (2013) [11] proposed a pipeline method for MS lesion segmentation that combines prior knowledge and contextual information into a boosting classifier. The prior knowledge was introduced in terms of atlas distribution of the main brain tissues. The contextual information was based on a large set of features describing the spatial context in the lesion neighborhoods. The experimental results obtained with two datasets from two different hospitals were shown a better segmentation was obtained using the extended outlier map as a feature in conjunction with the classical and contextual features. Still some improvement is needed in terms of specificity as a common issue. Geremia et al. (2013) [12] used segmentation problem that formalized as a binary classification of pixels samples into either background or lesions. It takes the advantages of context-aware features in the classification task to detect the differences in appearance of MS lesions with respect to healthy brain tissue. Subsequently, Geremia et al. (2013) [12] used the classification forest technique for MS lesion segmentation. But over segmentation occurred when MS lesions like similar intensity appears. Biediger et al. (2014) [13] used a strategy to improve the segmentation results of an automated system on whole-brain tissue classification and lesion detection. The first strategy leverages the current processing system at a granularity finer than the whole brain to detect lesions at a local level. This reflects the way that a physician considers only a part of the brain at a time. Then it combines the series of local results to produce whole-brain segmentation. This approach better captures the local lesion properties and produces encouraging results, with a general improvement in the detection rate of lesions. The second method dives deeper and looks at the individual pixels level. But spurious lesion generation is the one of the disadvantage of this method. Mechrez et al. (2016) [14] presents an automatic lesion segmentation method based on similarities between multichannel patches. A patch database is built using training images for which the label maps are known. For each patch in the testing image, similar patches are retrieved from the database. The matching labels for these patches are then combined to produce an initial segmentation map for the test case. Finally, an iterative patch-based label refinement process based on the initial segmentation map is performed to ensure the spatial consistency of the detected lesions. Till improvement is possible by using better database characterization and advanced metric learning. Authors [18] are still working on a system for screening of highly noisy patches, or patches which are extracted from fundamentally different brains. That segmentation method has strong capture capability of estimation, achieves great flexibility and broad applicability but does not perform well on contrast images with intensity homogeneity. Gao et al. (2014) [15] used a method to segment MS lesions, while normal tissues are also segmented and bias field is estimated to handle intensity in-homogeneities in the images. Another contribution of this paper is the introduction of a nonlocal means technique to achieve spatially regularized segmentation, which overcomes the influence of noise. Experimental results of that method [15] have demonstrated the effectiveness and advantages of the proposed algorithm but do not perform well on intensity homogeneity. Jain et al (2015) [16] proposed a MSmetrix based segmentation for lesion detection without requiring any training data. In MSmetrix, a probabilistic model is used to detect WM lesions as an outlier to normal brain while segmenting the brain tissue into GM, WM and CSF. Through its robustness and automation, MSmetrix could bring

ACCEPTED MANUSCRIPT

Proposed Methodology

AN US

3.

CR IP T

an added value (possibility to measure lesion volumes) for the clinical routine evaluation of MS patients. Hill et al. (2015) [17] proposed an automated segmentation of MS lesions in multi-contrast MRI using improved jump clustering method. Khayati et al.(2008)[18] proposed a fully automatic segmentation of multiple sclerosis lesions in brain MR images using adaptive mixtures and markov random field model(AMMR) . AMMR approach is based on a Bayesian classifier, utilizes the adaptive mixtures method and Markov random field model to obtain and upgrade the class conditional probability density function and the a priori probability of each class. To compare the performance of the proposed approach with those of previous approaches including manual segmentation, the similarity criteria of different slices related to 20 MS patients were calculated. A number of automated methods seek to lessen this burden by providing fast, accurate and repeatable segmentation results. The current system represents a significant investment of time and produces good results till it suffers from over, under segmentation and spurious lesion generation problem. Automatic MS lesion segmentation in MRI is a still challenging task due to the small size of the lesions, its heterogeneous characterizes and distribution, overlapping tissue intensity distributions with normal tissues, and the inherent artifacts of MRI. Thus the goal is to improve on the automated results to bring the results closer to those of an expert radiologist and perform better results than existing methods.

MS lesions do not appear within skull portion of human head. So, we use a skull removal methodology [19] to improve the MS lesion segmentation on input image IN(x, y) and store it into I(x, y). Then decompose our method in two key steps as background generation, and binarization. Final step for MS lesions detection and segmentation is used after background and binarization. A flowchart of proposed methodology has been shown is Figure 1. Input brain MRI

M

Start

Skull removed MRI of brain Outside contour of brain

ED

Select 3rd region of level set

PT

Connect the segmented tissues nearer to skull Sort the connected component (co.) in descending order

CE

Yes

Keep 1st component only and remove others

AC

No 30 0.75 Yes

Keep 1st, 2nd, and 3rd components only and remove others

Background Generation

End

No Keep 1st and 2nd components only and remove others

Binarization to identify MS

Select MS lesions by keeping only those regions which are present in Binarization but not present in background image

ACCEPTED MANUSCRIPT

Figure 1: Flowchart of proposed methodology

CR IP T

Total workflow of proposed methodology has been shown in above Figure 1 step by step. The details of segmentation, contour generation and selection of connected component as intermediate steps to generate background are described in first stage of methodology. In background generation step the output will contain some normal tissue but it does not contain any MS lesions. The details methodology of binarization to correctly identify the MS lesion is described in second stage. Binarization using global threshold gives us the MS lesions with few other normal tissues as output. Finally, MS lesions are identified by selecting only those regions which are present in binarization but not present in background image for accurate detection and segmentation. First stage: The combination of three phase level set and contour of brain is the key idea for adaptive background generation. We define the segmentation boundary as part of a surface where the contour level is 0, i.e., the zero level set. Zero level set [20] is used to represent the closed curve C, defined on the image domain by setting ( ) . Inside and outside C is represented signed minimum distance d from level zero curve to the (x, y) position. Level set in a closed curve C defined on the image on time t is represented by )

,

( ( (

) ) )

AN US

(

CE

PT

ED

M

(1) The curve C modified by where F is derived from generic energy function of the speed of evolution, and N designated as the outside unit normal vector to the increasing curve C. The progression of the level set function then represented by . We used regularization of ( ) using ( ) functions for calculating the related Euler-Lagrange equation, as projected in [21], using the heaviside function and dirac delta function . For the three region segmentation we consider two level set functions of region . We use two level set functions to define ( ) ( ) ( ), ( ) ( )( ( )), and ) enables ( ) ( ) to give a three region segmentation. Defining the level set function ( representation of an unknown curve C, regions Cin and Cout in terms of the energy functional ( ) [21]. Cin denotes the image region bounded by C, and Cout the image region outside C. In the proposed model, we minimize the following energy functional ( ) ( ) ( ) (2) Where are fixed parameters. The first two terms are regularization terms that impose local constraints on the curve, while Fext denotes the external force on the curve C. We define F ext via a Bayesian approach of minimizing the estimated thrashing incurred due to pixel misclassification. We define a defeat matrix L, where each element Lij corresponds to the loss incurred if a pixel belonging to the i th class is misclassified as that belonging to the jth class. We consider a two-class categorization problem and the loss in 2D matrix L is represented as row major order L11, L12, L21, L22. Let us denote by the related regularized energy functional, defined by

AC

( )

∫

( )

∫

( )

∬

∑

((

)|

∬

∑

.(

)|

) (

/ (

)(

( ))

)(

( ))

(3) wi; i = 1……M, represent mixture weights or prior probabilities of each factor satisfying summation of all weight equals to one for M pieces Gaussian densities. Minimizing the regularized energy functional with respect to gives

ACCEPTED MANUSCRIPT

the related Euler-Lagrange equation. Parameterization by an artificial time t≥0 gives the following update equation of ( ) in the descent direction: (

)

( )*

(

)

∑

∑

)|

.(

)|

((

(

/

)

(

)

)+

CR IP T

(4) Where n is calculated using a suitable finite difference scheme. The choice of a scheme is discussed in below. Returning to our original problem in two dimensions, we can calculate for every grid point as ( ) ( )( ) (5) ( ) ( ) , ( ) ( ) (6) We calculate in a similar fashion:

Where ( (

AN US

√

) )

( (

To prevent the appearance of instabilities, the timestamp t and , default expansion, we implementation Vn and s as ( )

M

(7)

) )

,

, and

is implemented as [22] . For

(8)

(

)

ED

(9)

CE

PT

(10) Where IG is the input image blurred with a Gaussian filter, is a constant for controlling how fast Vn approaches 0 when an edge is encountered in the image and is a constant for controlling the strength of s. edges), Vn approaches 1 and s approaches 0. However, when nearing an edge, Vn will approach zero, while s will point towards the edge. A larger will let Vn approach 0 at a faster rate. Note for default contraction, V n is multiplied by -1. We calculate these values using central difference: ( ) (11)

AC

( ) (12) √( )

( ) (13)

( )

(

) (14)

(

) (15)

ACCEPTED MANUSCRIPT

(

) (16) ((

)

(

) ) (17)

Each Gaussian module can be represented as: ( |

∑ )

(

)

⁄

*

⁄

(

)∑ (

)+

(

)

( )

CR IP T

(18) Consider M3 region IM3(x, y) is the final outcome of three phase level set. First-order derivative of I(x, y) is the difference on pixels along x and y direction. We deals with binary form of I(x, y) to detect the contour thus first derivative has been applied for contour detection. Binary of I(x, y) stores in IB(x, y), it contains brain portion as one value and rest background as zero. Horizontal contour is defined by

(19)

Vertical contour is defined by )

( )

AN US

(

(20) Horizontal and vertical contour detection does not produce continuous line which is not desirable for contour detection thus we combine both horizontal and vertical contour by (21)

M

Now we combine contour Cc(x,y) with the IM3(x, y) by using the following way ( ) ( ) ( )

AC

CE

PT

ED

(22) The above steps will connect all the normal tissues which are nearer to the skull or connected to the skull. If the ratio of IM3(x, y) segmented area and brain area of IB(x, y) is greater than 0.5 then need a subtraction. This subtraction is defined as ( ) ( ) ( ) (23) Then sort connected areas of IM3(x, y) in descending order in an array. If ratio of 1st and 2nd connected component is greater than 30 then keep 1st component only. Thus ratio of 1 st and 2nd connected component is greater than 30 means corpus callosum and segmented normal tissues are connected. Basically the ratio between segmented tissue and corpus callosum never reaches 30. But the ratio between the combination of segmented tissue and corpus callosum to the second highest connected component is high (e.g. high value 30). Giving 30 as high value we ensure corpus callosum is connected or not. Maximum variation between disconnected visible corpus callosum is 4/3 times (higher area to smaller area between corpus callosum) for MS lesions image in T2 images. Thus, we select 3 rd and 2nd connected component area (areas is in descending order) where ratio greater than 0.75 that means corpus callosum is disconnected. That empirical value 0.75 tested all images provided by “whole brain atlas”. Ratio of 3rd and 2nd connected component area is less than 0.75 signifies that corpus callosum is connected and we consider only 2nd component. Thus IM3(x, y) does not contains any MS lesions as MS lesions sizes are smaller than corpus callosum. Second stage: we use a binarized method that select threshold based on entropy and standard deviation preceded by non-gamut image enhancement [23]. To perform the binarization, total threshold calculated by ( ) difference between threshold by using entropy and threshold calculated from standard deviation i.e. ( ) (24)

ACCEPTED MANUSCRIPT

The values of has been chosen on the basis of observations. We use the Tf intensity to find out the MS lesion because MS lesion has almost similar intensity as normal tissues. Thus standard deviation ( ) of I(x,y) is calculated by √

( )

∑, (

( )-

)

*( ∑

( ) ( )

( ) ) ( )

(

∑

CR IP T

(25) Where np and m(I) are the total number of pixels and mean intensity of image I(x,y). An optimal threshold value Tfb is given by ( ) ( )

( ) )+ ( )

Results and Discussion

PT

4.

ED

M

AN US

(26) Where p(I) collective probability function and P(T) probability mass function [23]. Now, all the pixel intensity greater than of the gray image is set to 1 otherwise 0 and we get our desire binarized image. ( ) ( ) { ( ) (27) Third stage: Binary image IBN(x, y) consist of MS lesion with other normal tissues and binary image IM3(x, y) consist of some normal tissues without any MS lesions. Now we will select only those pixels (segmented region) which are present in IBN(x, y) but does not present in IM3(x, y). We formulate our method by using the following: ( ) ( ) ( ) (28) IBN(x, y) is the segmented image where MS lesion contains pixel value 1 and other part contains 0 or -1. ( ) ( ) { ( ) (29) IMS(x,y) is the segmented binary MS lesions image for our proposed methodology. We also find the top, bottom, left and right of the MS lesions from the top, bottom, left and right of brain and center of brain to the centroid of the MS lesions by using simple calculation the distance between two points [24].

AC

CE

We evaluated our method using data set of “Whole Brain Atlas” image data base [25], which consists of T1 weighted, T2 weighted, proton density (PD) MRI image. The method is implemented on Intel core2duo@ 2.13GHz processor and 2 GB RAM with windows 7 home basic 32 bit operating system. The support analysis software is used as MATLAB R2009a version. Automated MS lesions detection and segmentation is complex and challenging. The progression of the MS lesions shows considerable variability and it present temporal changes in shape, location, and area between patients and even for the same patient. So correct segmentation with its localization is very important and the results of our method have been shown in Figure 2. Figure 2(b) is the binary output of Figure 2 (a) after skull removal. Skull removal is used because MS lesions does not contains in skull. Skull removal makes post processing easier and we generate Figure 2(d) as contour of Figure 2(b). The level set segmentation of region three has been shown in Figure 2(c). Figure 2(c) contains normal tissues as well as some abnormal tissues. It is also very clear from Figure 2(c) that normal tissues nearer to skull are not connected. But it requires for background generation. To make normal tissues connected nearer to skull tissues we combine it with Figure 2(d) and generate Figure 2(e). Keeping one those maximum areas described in methodology are shown in Figure 2(e). Now Figure 2(e) does not contain any MS lesions. Figure 2(f) is the binarized output generated using entropy-standard deviation method. Figure 2(f) contains MS lesions and some other normal tissues. Figure 2(g) is final segmented MS lesions after some small area removal

ACCEPTED MANUSCRIPT

(e) Background

(b) After skull removal

(c) Combination of normal and abnormal

AN US

(a) Input brain image

CR IP T

(8 pixels) which is generated from subtraction of Figure 2(e) to Figure 2(f). Finally the MS lesions with red color have been shown in Figure 2(h) inputted MRI.

(f) Binarized result

(g) Segmented MS lesions

(d) Contour combination of horizontal and vertical

(h) MS lesion in brain

M

Figure 2. a) input brain image, b) Binary image after skull removal from (a), c) CSF segmentation from without skull image, d) combination of horizontal and vertical contour from (b), e) back ground generation from (c) and (d), f) output after binarization , g) segmented MS lesions from (f) and (e), and h) finally MS lesions are shown as red color in inputted brain image (a).

AC

CE

PT

ED

Now look at Figure 2(g) and Figure 2(h), and we can see three lesions of approximately the same size and one lesion as different size, which appear along with the associated of edema. The 4 foci of inflammitory activity are clearly not in synchrony. Notice the left lower lesion: it is at the base of the marginal sulcus and occupies white matter under the right post-central gyrus. As the acute inflammation grows, notice how the architecture of the post-central sulcus is displaced posterolaterally. MS lesions may change their position at the end of the year, this lesion has almost disappeared, but another has appeared just behind it, if some scan is performed over different months. More clear visions of different lesions with their position have been shown below in Figure 3.

(a) Top right lobe

(b) Near middle left lobe

(c) bottom right lobe

(d) bottom left lobe

Figure 3: Locations of four MS lesions are shown in (a), (b), (c) and (d) with red color

ACCEPTED MANUSCRIPT

Intersections of red line indicate the midpoint. Green dots in the top, left, bottom, and right are the corresponding extreme ends of brain. Blue dots in the top, left, bottom, and right are the highest, left, right and lowest pixels of segmented MS lesion. From those set of positions we measures the distances. Distance between brain top to segmented top position is denoted as „TT‟, distance between brain left to segmented left position is denoted as „LL‟, distance between brain bottom to segmented bottom position is denoted as „BB‟, distance between brain right to segmented right position is denoted as „RR‟ and distance between brain center to segmented center position is denoted as „CC‟. Distance of CC, TT, BB, LL, are shown in Table 1 for the input image shown in Figure 2 and position shown in Figure 3.

Number lesion detected

4

Position of the lesions or CC (x, y) 15.2593, 43.7593 -18.7400, -4.1600 22.1250, -19.2813 -24.4836, -31.3361

CR IP T

Table 1: Position of MS lesions and different distances TT

BB

50.4480 80.0250 105.1190 104.0769

124.1934 71.1758 68.2422 41.0488

LL

RR

77.2787 41.7732 85.3756 42.0595

68.7677 78.4092 41.1096 88.4590

ED

M

AN US

Once we find out the position of the lesions then it will be helpful for diagnosis form MS disease. The shifting of localization of MS lesions in different scan can be determined. Volume calculation is a great importance in diagnosis of MS lesions. We have used a stack based implementation [26] to form 3 dimensional (3D) figure from multiple 2 dimensional (2D) images. We have also measure the inter slice area to minimize the volume error. Volume V is calculated for n number of slices as, V=A1+ (L/2)*(A1+A2)+A2+ (L/2)*(A2+A3)+A3+….+An. Here A1, A2, A3,..,An are the individual area of n slices and L is inter slice distances. (L/2)*(A1+A2), (L/2)*(A2+A3), …., (L/2)*(A(n-1)+An) are the inter slice area to reduce the volume error. Inter slice distance L may vary depending on the thickness of MRI slices, In our case slice thickness is 0.5mm and inter slice distance is 1mm. If thickness is x and inter slice distance is y then we need to calculate ((y/x)-1)/2 time of total areas (eg. A1+A2) for each inter slice distance. The individual and inter slice areas of individual 24 slices of tour 1 in case 5 from „whole brain atlas‟ dataset is given below in Table 2. Table 2: Individual slice area with their inter-slice areas to calculate volume of MS lesions

AC

CE

1 1-2 2 2-3 3 3-4 4 4-5 5 5-6 6 6-7 7 7-8 8 8-9

Area

Input Sequence

Area

Input Sequence

209 194 179 202.5 226 217.5 209 225 241 250.5 260 250 240 193.5 147 127

9 9-10 10 10-11 11 11-12 12 12-13 13 13-14 14 14-15 15 15-16 16 16-17

107 117 127 160.5 194 129.5 65 59 53 54 55 128 201 141 81 72

17 63 17-18 66.5 18 70 18-19 157.5 19 245 19-20 342 20 442 20-21 433 21 424 21-22 286 22 148 22-23 287 23 426 23-24 269 24 112 Volume (total area) =8887.5

PT

Input Sequence

Area

ACCEPTED MANUSCRIPT

CR IP T

Shaded background shown in Table 2 are the inter slice areas and summation of all areas considered as volume of MS lesions. Measurement of volume is considers here 8887.5 number of pixels. It can be converted into measurable units (for eg. mm) that are depending on pixels dimension (for eg. 1 mm2 or 2 mm2 pixel dimensions). The segmented MS lesions for 24 input images have been shown in Appendix section (Figure 11). 3D outputs of stack based implementation from 2D images are shown in Figure 4 below:

AN US

(a) (b) (c) Figure 4: Three different views (a, b, c) of 3D representation of MS lesions form multiple images shown in red color. The above Figure 4 shows the different views of 3D representation for towards the volumetric computation of MS lesions. Thus 3D view, volume, area and positions will help the overall diagnosis system for MS disease from MRI of brain. Accuracy of volume is depending on accuracy of segmentation of MS lesions and segmentation accuracy and error are discussed in the below section. 4.1 Performance Evaluation

CE

PT

ED

M

To evaluate the segmentation techniques, comparison should be performed between the automatic segmentation generated by the proposed approach and the manual segmentation provided by the expert radiologist. To evaluate the effectiveness of any improvements it was necessary to specify a set of quantitative metrics for evaluating the progress of any proposed solutions. A number of metrics exist for comparing the computerized and expert results. The metrics [23,24] of interest for these comparisons are the True Positives (TP), False Positives (FP), False Negatives (FN), Sensitivity, Similarity-Index also called Kappa Index (KI), Jacard Index (JI), Relative Area Error (RAE), Correct Detection Ratio (CDR), and False Detection Ratio (FDR). The pixels that are marked as MS in both segmentation and reference are the true positives (TP), pixels those appear only in segmentation are false positives (FP), and pixels appearing only in reference are false negatives (FN). The sensitivity expresses the number of true positives (TP) identified by the computerized method over the sum of the true positives and false negatives (FN). Sensitivity [23, 24] represents a measure of the underestimation in the method, or the lesion pixels missed by the analysis.

AC

(30) The KI represents the amount of overlap in the identification regions provided by the experts and the method of interest. It is computed as the ratio of twice the area of intersection of the regions to the sum of the areas of the regions. The KI between two areas is calculated by the following equation ( ) ( ) (31) Here, AS and MS are the automated and manual segmented area. This similarity index is sensitive to both differences in size and location. The JI between two areas is represented as follow:

ACCEPTED MANUSCRIPT

(32) This metric is more sensitive to differences since both denominator and numerator change with increasing or decreasing overlap. CDR is defined by the following equation:

(32) FDR is defined by the following equation:

CR IP T

(34) The RAE [23, 24] is depend on difference of segmentation by automated method and manual segmented area can be calculated as ( )

AN US

(35) For Sensitivity, KI, JI, and CDR metrics, values fall between one and zero, with values closer to one representing better results and closer to zero being worse results. These four metrics represents a qualitative measure of the results and were central to the evaluation of improvement gains. As the automated methods are intended to assist the radiologist in identifying lesions, but under-segmentation and over-segmentation both are very important factor. If any over segmentation occurs, radiologist could more easily reject incorrectly identified lesions than under segmentation. The metric FDR, and RE metrics, values fall between one and zero, with values closer to zero representing better results and closer to one being worse results. Segmented area of proposed method, area of reference image, TP, FN, FP and sensitivity of proposed method has been shown in Table 3. Table 3. Segmented areas and sensitivity for proposed method

CE

M

Area REF 474 422 226 224 176 124 263 337 193 356

ED

Area PM 412 422 198 216 170 131 243 306 188 348

PT

Image No. 1 2 3 4 5 6 7 8 9 10

TP 412 422 196 211 168 114 238 295 185 345

FN 62 0 30 13 8 10 25 42 8 11

FP 0 0 2 5 2 17 5 11 3 3

sensitivity 0.869198 1.000000 0.867257 0.941964 0.954545 0.919355 0.904943 0.875371 0.958549 0.969101

AC

From above Table 3, it is clear that maximum images have a sensitivity greater than 0.90, thus it indicates a good result. FN and FP value is also very small. TP values are close to the reference segmented area, it also indicates a good segmentation. The results provided a quantitative way to evaluate the power of the proposed method to correct segmentation of lesions. Some recent methods have also been used to segment MS lesions and they evaluate their performance with respect to different metrics. We select some good methods for comparison with our proposed method and the methods are adaptive mixtures and markov random field model (AMMR) [18], Active Contour (AC) [10] and multimodal Graph Cuts (MGC) [7] based segmentation. The outputs produced in proposed method, AMMR, AC, and MGC based has been shown in Figure 5.

(b1) reference image

(d1) AMMR segmentation [18]

(e1) MGC segmentation [7]

(c1) AC segmentation [10]

AN US

(a1) input brain image

CR IP T

ACCEPTED MANUSCRIPT

PT

ED

M

(f1) proposed method

(b2) reference image

(c2) AC segmentation [10]

AC

CE

(a2) input brain image

(d2) AMMR segmentation [18], (e2) MGC segmentation [7]

(f2) proposed method

(b3) reference image

(d3) AMMR segmentation [18]

(e3) MGC segmentation[7]

(c3) AC segmentation [10]

AN US

(a3) input brain image

CR IP T

ACCEPTED MANUSCRIPT

ED

M

(f3) proposed method

(b4) reference image

(c4) AC segmentation [10]

AC

CE

PT

(a4) input brain image

(d4) AMMR segmentation [18] (e4) MGC segmentation [7]

(f4) proposed method

(d5) AMMR segmentation [18]

(b5) reference image

(c5) AC segmentation [10]

AN US

(a5) input brain image

CR IP T

ACCEPTED MANUSCRIPT

(e5) MGC segmentation [7]

(f5) proposed method

M

Figure 5: Outputs of different methods (AC, AMMR, MGC, and proposed method) on five different input images are shown above with their corresponding reference image for visual comparisons.

AC

CE

PT

ED

Our proposed method produced good results compared to other recent methods shown in Figure 5. Proposed method generates a very similar result as reference image shown in above figure for all images. Proposed method also ensures the reduction of spurious lesion generations than the comparable method. Only one spurious lesion is generated in Figure 5(f4). But MGC and AMMR method generates many spurious lesions compared to AC and proposed method visually. Few spurious lesions have been also generated from AC method compared to reference image. Some lesions are also missed by AC, AMMR and MGC based method that reduces the accuracy of comparable methods. So, from visual inspection it very clear that proposed method gives very good results compare to others. Look at the spot in the right frontal region of Figure 5. This is a relatively new lesion, and we see how it enlarges or spreads very rapidly over the weeks using proposed method. With time, the lesion changes their positions and characteristics and at last there is a white single lesion. We also see that the lesion has nearly disappeared while another lesion appearing. Visually comparisons may be biased so, we need metric based analysis of different method. We analyze with the metrics KI, JI, CDR, FDR and RAE discussed above and their column chart representations have been shown below in Figure 6. Input image is taken from “The Whole Brain Atlas” Case5, mri2 slices for comparison. 10 slices (selection is random but consecutive) numbers 28 to slice 37 T2 images have been shown in Figure 6 below because MS lesions are present on those slices. Proposed method has been tested on set of 100 images to validate our approach.

CR IP T

ACCEPTED MANUSCRIPT

Figure 6: Column chart representation of KI metric

ED

M

AN US

Graphical representation of KI metric (in %) for proposed method, AC, AMMR, and MGC method has been shown in Figure 6. Proposed and AMMR method produce better results than AC and MGC from the above Figure 6. AMMR does not reached above 80 for many images. Proposed method reaches above 80 for all images. Higher KI means higher similarity with reference image. Thus, proposed method produced best result with respect to KI metric which implies over segmentation, under segmentation and spurious lesion generation very small.

PT

Figure 7: Column chart representation of JI metric

AC

CE

Graphical representation of JI metric (in %) for proposed method, AC, AMMR, and MGC method has been shown in Figure 7. Proposed and AMMR method produce better results than AC and MGC from the above Figure 7 because for some images JI value of AC and MGC is less than 60. AMMR reached 60 for many images but does not reached above 80 for most of the images. Proposed method reaches above 80 for all images. Higher JI refers smaller over and under segmentation with respect to reference image. Thus, proposed method produced best result with respect to JI metric.

CR IP T

ACCEPTED MANUSCRIPT

Figure 8: Column chart representation of CDR metric

ED

M

AN US

Graphical representation of CDR metric (in %) for proposed method, AC, AMMR, and MGC method has been shown in Figure 8. Here proposed, AMMR and MGC method produce better results than MGC from the above Figure 8. Among the proposed, AMMR and MGC method, the proposed method performs better. CDR indicates how much lesions correctly detected but it is not concerned with the generation of spurious lesion. Its value does not reflect over segmentation or false segmentation. So, in context of correct segmentation proposed method always reaches above 80 but the other methods fail to reach 80 for few images.

PT

Figure 9: Column chart representation of FDR metric

AC

CE

Graphical representation of FDR metric (in %) for proposed method, AC, AMMR, and MGC method has been shown in Figure 9. FDR is very sensitive on spurious lesions and over segmentation. So, FDR will be higher for over segmentation. AMMR and MGC method have high FDR for maximum images and AC also generates high FDR for few images. But for proposed method only one images it crosses 10. Thus, in context of FDR, proposed method gives better results than the selected recent method.

CR IP T

ACCEPTED MANUSCRIPT

Figure 10: Column chart representation of RAE metric

M

AN US

Graphical representation of RAE metric (in %) for proposed method, AC, AMMR, and MGC method has been shown in Figure 10. RAE does not depend on the true positive value. Area difference from reference image is important factor for RAE. AMMR and MGC have high RAE because it crosses 20 for maximum images. AC also crosses 20 for few images. But proposed method does not reach 20 for any images which is a good indication for good segmentation. The average values of performance evaluation metrics for the different method have been shown in Table 4. Table 4: Average value Method name RAE KI JI CDR FDR Proposed Method 5.973647 94.88385 90.43217 92.60284 2.552548 AMMR Method 31.75755 80.09598 67.56388 89.00945 33.76821 MGC Method 28.88982 74.11517 61.26577 82.75727 39.44054 AC Method 24.94408 77.69067 66.84883 77.89032 20.63257

CE

PT

ED

AMMR based segmentation is good method in context of KI, JI, and CD but it does not produce low RAE and FDR values. Thus, AMMR method has high RAE and FDR due to the generation of many spurious lesions. AC method has low error rate compare to MGC and AMMR method but it suffers from similarity and true positive problem. Under and over segmentation both occurs in MGC based segmentation CDR is very good. KI, JI, and CDR values of proposed method greater than 90% and it indicates high similarity, true positive and correct detection. RAE and FDR value of proposed method less than 10% that indicates a low error rate. Thus, proposed method gives better results than the three other methods in all respect. But for few images proposed method too suffers from under or over segmentation problems. The research objective is to improve MS lesion segmentation based on the results of an existing system and it successfully executed. The goal of better match physician results without introducing excessive false detections is also executed successfully. To validate the approach, the results and analyses for each improvement step are executed for a set of 100 images (10 results has been shown in this paper).

AC

4.2 Complexity Analysis The background generation has time complexity of O((M× (N-1)+(M-1) ×N)+M×N×log(M×N)) where M, N are the numbers of pixels in the image. Time requirement for entropy is O(H×K×M×N), where H denotes the size of the image histogram, time taken by standard deviation O(M×N), and time taken for subtraction of two images is O(M×N) respectively. So, the total time for binarized image (for K=2 and H as constant ) is O((M× (N-1)+(M-1) ×N)+M×N×log(M×N)) + O(H2×M×N)+ O(M×N)+ O(M×N)=O(M×N×log(M×N))=O(MNLogMN). AC for measuring sum of the pixels that were marked by only one region is O(MN CI), where M,N is the image size, C is the number of the set of connected regions that can be labeled and I is number of iterations. In MGC method, the weights for each point are calculated as the average distance, the region weight distance metric must be calculated O(M×N) and O(M×N) times for the object and background weights respectively

ACCEPTED MANUSCRIPT

where M×N is the number of pixels in the image. The region distributions required is O(M×Nlog(M×N)). Thus the total number of comparison is= O(M×N)+ O(M×N)+ O(M×Nlog(M×N))= O(MNlog(MN)) In AMMR method of finding the best region in one image is O(X ×C). To find out shortest distance it requires XlogX for probability density function. Initialize the first point of data is require O(M×N) and minimum squared distance of the new point from all the terms is (XlogX). Here X=M×N number of pixels, thus complexity is =O(M×N×C)+ O(M×N ×log M×N)+ O(M×N) = O(MNlog(MN)). Thus, proposed method is computationally comparable with three selected recent method.

Conclusions

CR IP T

5.

AN US

Proposed method provides a better segmentation when tested with defined data sets both visually as well as metrically. We used two different strategies to improve the lesion detection and segmentation. The use of background subtraction concept deals with higher level of MS lesion segmentation which does not generates any spurious lesions. The use of entropy and standard deviation based binarization increased detection efficiency. Thus, the combination of two concepts gives higher accuracy with lower error rate compared to other recent AMMR, AC, and MGC method. The method accurately identifies the size, number of lesions and location of lesion detections as a radiologist perform. The adaptability of the proposed method creates a number of potential opportunities in clinical practice for the detection of MS lesions in MRI. More efforts are still needed to remove the spurious lesion generation where lateral ventricles outside the symmetric line and selection range. Acknowledgement The authors wish to thanks Dr. Pradip Saha, MD in radiology from NRS medical college for his valuable suggestions and helpful comments. He is currently working as radiologist at M N Roy diagnostic center, Kolkata, India. He has supported in the arrangement of reference image, guided, and technically assistance to enrich our knowledge in the preparation of this manuscript.

M

Conflict of Interests The authors declare that there is no conflict of interests regarding the publication of this paper.

AC

CE

PT

ED

Reference [1] Q. Mahmood, A. Chodorowski, M. Persson, "Automated MRI brain tissue segmentation based on mean shift and fuzzy c-means using a priori tissue probability maps," Elsevier, IRBM, VOL 36 PP. 185–196, 2015 [2] P K Roy, A Bhuiyan, A Janke, P M. Desmond, T Yin Wong, W P. Abhayaratna, E Storey, K Ramamohanarao, "Automatic white matter lesion segmentation using contrast enhanced FLAIR intensity and Markov Random Field," Computerized medical imaging and graphics, Volume 45, Pages 102–111 , October 2015 [3] http://www.msif.org/about-ms/what-is-ms/ [Online], 30-Octobar-15. [4]D. Mortazavi, Z. Abbas and H. Soltanian-Zadeh, "Segmentation of Multiple Sclerosis Lesions in MRI Images: A Review," Neuroradiology, vol. 54, no. 4, pp. 299-330, 2011. [5] M Inglese, “MRI measures of neuroprotection and repair in multiple sclerosis,” Journal Neurological Science, Volume 311, Issue 1, Pages 16–23. 2016 [6] M Filippi, MA Rocca, F Barkhof, W Brück, JT Chen, G Comi, G DeLuca, N D Stefano, BJ Erickson, N Evangelou, F Fazekas, JJ Geurts,C Lucchinetti, DH Miller, D Pelletier, BF Popescu,H Lassmann, “Association between pathological and MRI findings in multiple sclerosis,” Lancet Neurol, Volume 11, Issue 4, pp. 349-60, 2012. [7]D. Garc´ıa-Lorenzo, J. Lecoeur, D.L. Arnold,D.L. Collins, and C. Barillot, “Multiple Sclerosis lesion segmentation using an automatic multimodal Graph Cuts,” 12th International Conference on Medical Image Computing and Computer Assisted Intervention, Londres, UK, 2009. DOI : 10.1007/978-3-642-04271-3 [8]F Derraz, L Peyrodie, A Pinti, A Taleb, A Chikh, “Semi-automatic Segmentation of Multiple Sclerosis Lesion Based Active Contours Model and Variational Dirichlet Process,” Computer Modeling in Engineering and Sciences, Tech Science Press, 67 (2), pp.95-117, 2010. [9] F. Forbes, S. Doyle D. Garcia-Lorenzo, C. Barillot M. Dojat “A Weighted Multi-Sequence Markov Model For Brain Lesion Segmentation” Proc. AISTATS, Chia Laguna Resort, Sardinia, Italy. 2010,

ACCEPTED MANUSCRIPT

AC

CE

PT

ED

M

AN US

CR IP T

[10]F Derraz, Antonio Pinti, Laurent Peyrodie, Miloud Boussahla, Hechmi Toumi and Patrick Hautecoeur, "Multiple Sclerosis lesion segmentation using Active Contours model and adaptive outlier detection method ," Proceedings IWBBIO 2014. Granada 7-9 April, pp. 878-889, 2014 [11] M Cabezas, A Oliver, J Freixenet, X Llado, “A Supervised Approach for Multiple Sclerosis Lesion Segmentation Using Context Features and an Outlier Map”, IbPRIA 2013, LNCS 7887, Springer-Verlag Berlin Heidelberg, pp. 782–789, 2013. [12] E. Geremia, D. Zikic, B. Menze, B. Glocker, E. Konukoglu, J. Shotton, O. M. Thomas, S. J. Price, T. Das, R. Jena, N. Ayache and A. Criminisi , “Classification Forests for Semantic Segmentation of Brain Lesions in MultiChannel MRI,” published in Decision Forests for Computer Vision and Medical Image Analysis 2013, Springer London, pp. 245-260, 2013 [13] D Biediger, C Collet, J-P Armspach, “Multiple sclerosis lesion detection with local multimodal Markovian analysis and cellular automata „GrowCut‟”, Journal of Computational Surgery, Springer, vol 1: no 3, pp.1-15, 2014 [14] R Mechrez, J Goldberger, H Greenspan, “Patch-Based Segmentation with Spatial Consistency: Application to MS Lesions in Brain MRI” Hindawi Publishing Corporation, International Journal of Biomedical Imaging, Volume 2016, Article ID 7952541, pp. 1-13,2016. [15] J Gao, C Li, C Feng, M Xie, Y Yin, C Davatzikos, “Non-locally regularized segmentation of multiple sclerosis lesion from multi-channel MRI data,” Magnetic Resonance Imaging, Elsevier, Volume 32, pp.1058–1066, 2014. [16] S Jain, D M Sima, A Ribbens, M Cambron, A Maertens, W V Hecke, J DeMey, F Barkhof, M D Steenwijk, M Daams, F Maes, S Van Huffel, H Vrenken, D Smeets, "Automatic segmentation and volumetry of multiple sclerosis brain lesions from MR images," NeuroImage: Clinical, Volume 8, pp. 367–375, 2015 [17] J Hill, K Matlock, B Nutter, S Mitra ,“Automated Segmentation of MS Lesions in MR Images Based on an Information Theoretic Clustering and Contrast Transformations,” Technologies, MDPI, Vol 2015, No. 3, pp. 142161, 2015; doi:10.3390/technologies3020142 [18] R Khayati, Mansur Vafadust, Farzad Towhidkhah, S. Massood Nabavi, "Fully automatic segmentation of multiple sclerosis lesions in brain MR FLAIR images using adaptive mixtures method and markov random field model," Computers in Biology and Medicine, vol 38, pp. 379 – 390, 2008 [19] S Roy, S Nag, I K Maitra, S K Bandyopadhyay, "Artefact Removal and Skull Elimination from MRI of Brain Image," International Journal of Scientific and Engineering Research, Volume 4, Issue 6, June-2013, pp 163-170. [20] S Osher, J A Sethian, “Fronts propagating with curvature dependent speed: Algorithms based on HamiltonJacobi formulations,” Journal of Computational Physics, vol. 79, no. 1, pp. 12–49, 1988. [21] C Li, R Huang, Z Ding, J C Gatenby, D N Metaxas, "A Level Set Method for Image Segmentation in the Presence of Intensity Inhomogeneities With Application to MRI," IEEE Transactions On Image Processing, Vol. 20, No. 7, 2011. [22] T F Chen, "Medical Image Segmentation using Level Sets," Technical Report #CS-2008-12, pp.1-8, 2008. [23] S Roy, D Bhattacharyya, S K Bandyopadhyay, T H Kim, “An Improved Brain MR Image Binarization Method as a Preprocessing for Abnormality Detection and Features Extraction,” Frontiers of Computer Science, Springer, pp. 1-12, 2016. DOI: 10.1007/s11704-016-5129-y. [24] S Roy, S Nag, S K Bandyopadhyay, D Bhattacharyya, T H Kim, “Automated Brain Hemorrhage Lesion Segmentation and Classification from MR Image Using an Innovative Composite Method,” Journal Of Theoretical and Applied Information Technology, Volume 78, No. 1, pp. 34-45, August, 2015. [25] http://www.med.harvard.edu/AANLiB/cases/ [online] 15-Janurary-2015 [26] S Roy, S Sadhu, S K Bandyopadhyay “A Useful Approach towards 3D Representation of Brain Abnormality from Its 2D MRI Slices with a Volumetric Exclamation,” Proc. IEEE, C3IT, India, pp.1-6, 7-8 February, 2015 Appendix: Segmented MS lesions from 24 input brain images for 3D view and volume calculation are shown below in Figure 11. Input image taken from “The Whole Brain Atlas” Case5, mri2 and where slice 15 to slice 37 T2 images has been selected.

ACCEPTED MANUSCRIPT

O1

I3

O3

I5

O5

I2

O2

I4

O4

I6

O6

CE

PT

ED

M

AN US

CR IP T

I1

O7

I8

O8

I9

O9

I10

O10

AC

I7

ACCEPTED MANUSCRIPT

O11

I13

O13

I15

O15

I12

O12

I14

O14

I16

O16

CE

PT

ED

M

AN US

CR IP T

I11

O17

I18

O18

I19

O19

I20

O20

AC

I17

ACCEPTED MANUSCRIPT

O21

I22

O22

AN US

CR IP T

I21

AC

CE

PT

ED

M

I23 O23 I24 O24 Figure 11: Segmented output ‘O’ for input images ‘I’ with sequence number have been shown here. Data set selected from ‘whole brain atlas’ case 5, tour 1.