Technology and Health Care 22 (2014) 775–784 DOI 10.3233/THC-140845 IOS Press
775
Skull removal in MR images using a modified artificial bee colony optimization algorithm Mohammad Taherdangkoo Department of Artificial Intelligence, Tehran Business School, Tehran, Iran E-mail: [email protected] Received 13 June 2014 Accepted 6 July 2014 Abstract. Removal of the skull from brain Magnetic Resonance (MR) images is an important preprocessing step required for other image analysis techniques such as brain tissue segmentation. In this paper, we propose a new algorithm based on the Artificial Bee Colony (ABC) optimization algorithm to remove the skull region from brain MR images. We modify the ABC algorithm using a different strategy for initializing the coordinates of scout bees and their direction of search. Moreover, we impose an additional constraint to the ABC algorithm to avoid the creation of discontinuous regions. We found that our algorithm successfully removed all bony skull from a sample of de-identified MR brain images acquired from different model scanners. The obtained results of the proposed algorithm compared with those of previously introduced well known optimization algorithms such as Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO) demonstrate the superior results and computational performance of our algorithm, suggesting its potential for clinical applications. Keywords: Skull bone region, particle swarm optimization (PSO), ant colony optimization (ACO), artificial bee colony (ABC), MRI segmentation
1. Introduction The segmentation of brain magnetic resonance (MR) images is an essential preprocessing step which greatly affects subsequent analysis, the results of which can be used to formulate the next medical actions in patient management such as diagnosis and treatment. For decades, many algorithms have been introduced for segmenting MR images but often they can only be applied to data sets generated from a specific vendor’s MR system for which the algorithm was developed. The main limitations for generalizing such algorithms to all or most systems include the lack of a unique standard MR acquisition protocol and more importantly the difficulty introduced by the existence of the skull in the images. Indeed, before segmenting the brain itself it is necessary to identify and remove the structures such as skull that are not required for the tissue classification task so they can be excluded from the new input data set used for subsequent processing. This is a difficult preprocessing task that significantly affects the success of segmentation algorithms of brain tissue [15]. In this paper, we propose a semi-automated algorithm for removing the skull and bony structures from MR images that is generalizable to multiple MR scanners and acquisition protocols. In some previous work, the skull removal step is done manually which is very time consuming and is not recommended c 2014 – IOS Press and the authors. All rights reserved 0928-7329/14/$27.50
776
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
when the ultimate goal is the objective measurement of brain volume [16,17]. Other approaches consist of applying proper filters especially designed to remove the skull such as edge or boundary detection filters, thresholding and region growing algorithms [1]. However, these methods require adjustment of filter parameters for each specific acquisition protocol and each vendor device. For instance, a region growing algorithm can be used to segment the brain image into skull and non-skull structures and then to remove the skull [2,5,6]. However, in this algorithm a seed point belonging to the skull region must be manually selected by an operator or specialist, and errors in seed selection can be propagated through the algorithm resulting in overall poor performance. In this paper, we developed and implemented an optimization algorithm based on the artificial bee colony (ABC) algorithm to identify and remove bony skull structures in MR images of the head. The ABC technique has been recently proposed [7] based on real bee behavior for finding foods. Optimization algorithms are typically used to solve complicated mathematical problems in a simple way. One of the commonly used optimization algorithms is Particle Swarm Optimization (PSO) which was first introduced by Kennedy and Eberhart [8]. It was inspired by social behavior and movement dynamics of insects, birds and fish. It is a global gradient-less stochastic search method which is well suited for continuous variable problems. Due to its high speed, low computational cost, easy encoding and memory aspect of particles, PSO is used widely for different optimization problems and applications such as edge detection [9]. Despite its simple implementation, few required input parameters and global search ability, PSO does not perform well for local search tasks, something which is especially important for image processing of local features as required in segmentation algorithms. Another important optimization algorithm was proposed by Dorigo et al. [10] called Ant Colony Optimization (ACO). It is based on ants’ behavior in finding foods in the real world. It was shown that ants use colony intelligent action which is random and independent to connect with each other in indirect ways. This algorithm has many applications such as edge detection [11] and image segmentation [12]. The ACO algorithm is inherently parallel, very efficient to adapt to changes in problem constraints and powerful for global search. However, it results in a theoretical solution with uncertain (although guaranteed) time to convergence. Optimization algorithms have been successfully used for multimodality medical image registration [13,14], but in general, for medical image segmentation, few methods have been proposed based on optimization algorithms. One of the main challenges in using optimization algorithms is the design of the cost function and the way to minimize this function. In this paper, we propose a new simple criterion in the ABC algorithm especially developed to segment out the skull region in MR images and to remove it. Our algorithm does not depend on scanner protocols or specific devices, and can be extended to segment other organs. 2. Artificial Bee Colony (ABC) algorithm The Artificial Bee Colony (ABC) algorithm is a relatively new optimization algorithm based on insect’s behavior in detecting, tracking and finding food. The algorithm defines three types of bees; employed bees, onlooker bees and scout bees. Bees in a colony cleave into two teams of bees (employed and onlooker bees), where one employed bee is assigned to gather information about an available food source in a predefined search space, such that, all food sources and the total search space of a problem are covered by the group of employed bees. Then the employed bees share their information about food sources gathered from their search spaces with the onlooker bees. The onlooker bees sort the food sources from the richest nectar to poorest nectar. Then the onlooker bees select the employed bee having
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
777
the information about the richest nectar as follows. The chance of an employed bee being selected by onlooker bees can be computed from following equation. F (θi ) , Pi = S k=1 F (θi )
(1)
Where Pi is the chance of selecting the employed bee, F (θi ) is amount of nectar in food source i, θi is the food source i, Sk=1 F (θk ) is the total amount of nectars available in all food sources in the search space, k represents the employed bee and their specific search space, with the maximum probability occurring for (the employed bee with information about) the richest nectar. After being selected, the employed bee becomes a scout bee and starts gathering recruitments to move toward the target, i.e. the richest food source. Their goal is to vacate the target nectar and to search the other food sources in local nearest neighborhoods. Upon return to the hive, the scout bee shares the information obtained from the neighborhoods by the onlooker bees. Then the sort of food sources from richest to poorest will be repeated. This process continues until the optimal solution is found. In brief, the steps of the proposed method are as follows: Step 1. The movement of employed bees toward finding food sources and determining their content. Step 2. Sharing the information gathered by the employed bees with the onlooker bees and sorting the food sources based on their value (e.g., nectar). Step 3. Selecting a scout bee and moving toward the richest selected food source, searching its neighborhoods and vacating it. Step 4. Stopping algorithm when an optimal solution is achieved, otherwise going to step 2 and repeat. 3. Methods The success of brain tissue segmentation is highly dependent on removal of bony structures, (e.g., skull) from the image. First the region containing the skull must be determined and then removed from the MR images. To accomplish this, we modified the ABC algorithm and adapted it for medical image segmentation. The new algorithm is effective because the brain tissue is surrounded by cerebrospinal fluid (CSF) which is a region of gray values with little variation. To begin, we consider all pixels of a given MR image as the search space, and we assign a probability to each pixel in proportion to its intensity value. In this way, white pixels and black pixels have maximum and minimum probability respectively. All other pixels with their respective gray values, including pixels in the target region have probabilities between these minimum and maximum probabilities. After assigning the probability to each pixel, we define an initial interval of probabilities as food sources for scout bees. To do that, a mean probability μ and a variance σ 2 are initialized based on the probability of pixels belonging to CSF region in different MR images. In the conventional ABC algorithm, the direction and coordinates of scout bees are initially selected randomly. However, in our proposed algorithm we impose a direction of search. Namely, we select four initial scout bees (pixels) in the extreme right, left, top and bottom of a brain MR image and require their direction of movement top toward the center of the image (see Fig. 1). Each scout bee searches the space of search in his proscribed direction. Every time a scout bee finds a food source, he shares the information obtained from this food source with others. The probability of the new pixel being considered as a part of CSF region is based on other pixels previously joined to the CSF region. Consequently the mean μ and σ 2 of CSF will be updated every time a pixel becomes a part of
778
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
Fig. 1. Initialization of coordinates and the search direction of scout bees. (Colours are visible in the online version of the article; http://dx.doi.org/10.3233/THC-140845)
CSF region. In this way, an associative memory is created and shared among scout bees. However, it is probable that some discontinuous regions can be formed. In order to avoid the creation of such regions, we impose another criterion for considering a pixel as a part of final region. This criterion is the distance between the intensity values of two adjacent pixels which is defined as: DSij = Ii2 − Ij2 , (2) Where DSij is the distance between two pixels i and j , Ii and Ij are their intensity value respectively. If the distance between two adjacent pixels is near to zero, it means that the two pixels are more likely a part of the same region. This constraint assists in preserving the continuity of regions. The pseudo code of the proposed algorithm is as follows: Input: intensity pixels of a brain MR image (I(i, j)). Output: identifying and labeling pixels belonging to skull region (ISkull (i, j)). Algorithm: Step 1. (Initialization): initialing E (the number of employed bees), O (the number of onlooker bees) such that E = O and S = 4 (the number of scout bees = orientation) where should it be the following with an M (rows) × N (columns) matrix (i1 = 1 to M/2, j1 = 1 to N means search down), (i2 = 1 to M , j2 = 1 to N/2 means search right) (i3 = M to M/2, j3 = 1 to N means search up) (i4 = 1 to M , j4 = N to N/2 means search left) are initial coordinates and searching directions of scout bees for an image of size M × N . Step 2. (Searching process): Set the food source intensity (FSI) equal to the intensity of CSF around the brain and begin the search process. Step 3. (Growing CSF region): Save last intensity to employed bees’ memory (Memployedbees = ISkull ). Step 4. (Stopping criterion): resume movement while criterion (I(i, j) < FSI) for all four search areas is satisfied, checking the distance criterion for neighboring pixels by using Eq. (2).
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
779
Fig. 2. A. Example images from eight de-identified MRI brain examinations. B. Segmented brain results of proposed method.
780
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
Fig. 2. continued.
∗ ∗ ∗ Fitness Criterion ∗ ∗∗ − − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
Move from up to down n = 0 : M/2 − 1
if
I(i1 + n, j1 ) < FSI then
Iskull = I(i1 + n, j1 )
− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
Move from left to right n = 0 : N/2 − 1
if
I(i2 , j2 + n) < FSI then
Iskull = I(i2 , j2 + n)
− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
Move from right to left n = N : N/2
if
I(i1 , j1 − n) < FSI then
Iskull = I(i1 , j1 − n)
− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
The specifications of the computer on which the algorithms were run are as follows: Machintosh OS, Two 2.93 GHz 6-Core Intel processors, 64 GB RAM, HD 5870 1 GB HD, and ATI Radeon Graphics Card.
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
781
Fig. 3. From left to right Original image, results of PSO, results of ACO, results of ABC, results of our proposed method.
4. Experimental results We used several test brain images using MR scanners of 1, 1.5 and 3 Tesla (GE Medical, Waukesha, WI). All test images have size 256 × 256 pixels with gray scales of 16 bit depth and were acquired at Faghihi Hospital in the Department of Radiology in Shiraz, Iran. Figure 2 shows the results of proposed method for several example images from eight de-identified cases of MRI brain examinations. We also compared the results of proposed method with PSO, ACO and conventional ABC algorithms. Figure 3 shows the comparative results. As can be seen, other methods fail to remove skull region
782
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm Table 1 Parameters used in the tested methods PSO Number of swarm C1 C2 ωmax ωmin Number of iterations ACO Number of ants Probability threshold for maximum trail Local search probability Evaporation rate Number of iterations ABC Number of bees (employer and onlooker) (Conventional ABC and Proposed ABC) Number of sites selected for neighborhood search Number of bee recruited for best sites Number of iterations
100 1.9 2.1 1 0.5 500 50 0.95 0.01 0.01 1000 20 10 5 500
Table 2 Cost (CPU time) obtained for each different methods Row of images in Fig. 3 (From top to bottom) 1
2
3
4
5
6
Method PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC
Real time in second 0.0009 0.0026 0.0005 0.00000102 0.00232 0.00764 0.00187 0.0000627 0.0378 0.0674 0.0139 0.0123 0.0196 0.0362 0.0128 0.0016 0.0096 0.0361 0.0082 0.00014 0.0067 0.0092 0.0052 0.00082
completely, while proposed method achieves significantly superior performance in completely removing the skull region. The results also show that the proposed algorithm is independent of scanner modality and it suggests a promising clinical application which could be used by radiologists and physicians. Table 1 shows the parameters used in each method in our experiments. The cost or CPU time of each method for each of the test images in Fig. 3 were computed and added in Table 2.
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
783
5. Discussion Medical image segmentation can be considered as dividing an image into homogenous regions that are often a part of desired objects in the medical image. Many criteria can be used to measure the homogeneity of a region such as pixel intensity. The medical image segmentation is a pre-step that is relatively complex in medical image processing. It can be also considered as a clustering process in which pixel with similar features can be assigned to a cluster. So, clustering methods have widely used in medical image segmentation for separating different organs in medical images [1–3,8–10]. In this paper, it was tried to select the test images in which the pixel intensities of neighboring organs are so close that the extraction of CSF (Cerebrospinal Fluid), WM (White Matter), and GM (Gray Matter) and also removing skull are a difficult task to do. We should state that we have implemented and evaluated different meta-heuristic methods including PSO, ACO and ABC and also some existing tools for skull removal based on region growing and Watersheld. The best results obtained from these previously used methods are brought in experimental results section of the paper their performance have been evaluated and compared with the proposed method. In addition, the execution time of each method has been computed and brought in Table 2. As can be seen from Table 2, the proposed algorithm has achieved the final response in much shorter time than other methods. It is worth to mention that PSO and ACO take more execution time related to other methods because they have many loops and use many If-Then conditions, and so they cannot be applicable in real time applications. Although the existing tools can obtain the final response within an acceptable time, the resulted segmentations are not so good and thus the next steps using these results may face on problems. So they are not suggested for clinical applications. From the images in the third column of Fig. 3 we can obviously see that PSO like ACO did not produce the acceptable results and they both should be reconsidered for being used in brain image segmentation applications. By focusing on fifth column of Fig. 3, the proposed method produce the acceptable results. 6. Conclusion A new algorithm for skull removal from MR Brain images has been proposed based on a modified ABC optimization algorithm. The ABC algorithm is a new powerful optimization algorithm which models the behavior of bees in finding rich sources of foods. Experiments on example images from eight deidentified MRI brain examinations show that our ABC algorithm achieves superior results as compared to other optimization algorithms such as PSO and ACO, successfully removing all bony skull from the images, and is computationally faster. However, its application to image segmentation has not been fully exploited. In this paper, we use a modified ABC algorithm for segmenting the brain MR images into skull and non-skull regions and then remove the skull region. Comparative results have demonstrated the superior performance of the algorithm related to other optimization based algorithms. In addition, our proposed algorithm performs very well across different MR scanners, which shows its potential for use in a clinical framework. References [1]
Ballard DH, Brown DM. Computer vision. Prentice Hall, Englewood Cliffs, NJ; 1982.
784 [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17]
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm Haralick RM, Shapiro LG. Image segmentation techniques. Computer Vision, Graphics, and Image Processing. 1985; 29(1), 100–132. Saha S, Maulik U. A new line symmetry distance based automatic clustering technique: Application to image segmentation. International Journal of Imaging System and Technology. 2011; 21(1), 86–100. Odille F, Steeden JA, Muthurangu V, Atkinson D. Automatic segmentation propagation of the aorta in real-time phase contrast MRI using nonrigid registration. Journal of Magnetic Resonance Imaging. 2011; 23(1), 232–238. Gholipour A, Estroff JA, Barnewolt CE, Connolly SA, Warfield SK. Fetal brain volumetry through MRI volumetric reconstruction and segmentation. International Journal of Computer Assisted Radiology and Surgery. 2011; 6(3). 329– 339. Wright R, Kyriakopoulou V, Ledig C. Rutherford MA, Hajnal JV, Rueckert D, Aljabar P. Automatic quantification of normal cortical folding patterns from fetal brain MRI. NeuroImage. 2014; 91, 21–32. Kemal Apalak M, Karaboga D, Akay B. The artificial bee colony algorithm in layer optimization for the maximum fundamental frequency of symmetrical laminated composite plates. Engineering Optimization. 2014; 46(3), 420–437. Clerc M, Kennedy J. The particle swarm – explosion, stability, and convergence in a multidimensional complex space. IEEE Transactions on Evolutionary Computation. 2002; 6(1), 58–73. Setayesh M, Zhang M, Johnston M. A novel particle swarm optimization approach to detecting continuos, thin and smooth edges in noisy images. Information Sciences. 2013; 246, 28–51. Dorigo M, Thomas S. Ant Colony Optimization. MIT Press. 2004. Cerello P, Cheran SC, Bagnasco S, Bellotti R, Bolanos L, Catanzariti E, Nunzio GD, Fantacci ME, Fiorina E, Gargano G, Gemme G, Torres EL, Masala GL, Peroni C, Santoro M. 3-D object segmentation using ant colonies. Pattern Recognition. 2010; 43(4), 1476–1490. Taherdangkoo M, Bagheri MH, Yazdi M, Andriole KP. An effective method for segmentation of MR brain images using the ant colony optimization algorithm. Journal of Digital Imaging. 2013; 26(6), 1116–1123. Valsecchi A, Damas S, Santamaria J, Marrakchi-Kacem. Intensity-based image registration using scatter search. Artificial Intelligence in Medicine. 2014; 60(3), 151–163. Manikandan S, Ramar K, Willjuice Iruthayarajan M, Srinivasagan KG, Multilevel thresholding for segmentation of medical brain images using real coded genetic algorithm. Measurment. 2014; 47, 558–568. Jeroen de Bresser, Portegies MP, Leemans A, Biessels GJ, Kappelle LJ, Viergever MA. A comparison of MR based segmentation methods for measuring brain atrophy progression, NeuroImage. 2011; 54(2), 760–768. Battaglini M, Smith SM, Brogi S, Nicola De Stefano. Enhanced brain extraction improves the accuracy of brain atrophy estimation, NeuroImage. 2008; 40(2), 583–589. Segonne F, Dale AM, Busa E, Glessner E, Salat D, Hahn HK, Fischi B. A hybrid approach to the skull stripping problem in MRI. NeuroImage. 2004; 22(3), 1060–1075.
Copyright of Technology & Health Care is the property of IOS Press and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
775
Skull removal in MR images using a modified artificial bee colony optimization algorithm Mohammad Taherdangkoo Department of Artificial Intelligence, Tehran Business School, Tehran, Iran E-mail: [email protected] Received 13 June 2014 Accepted 6 July 2014 Abstract. Removal of the skull from brain Magnetic Resonance (MR) images is an important preprocessing step required for other image analysis techniques such as brain tissue segmentation. In this paper, we propose a new algorithm based on the Artificial Bee Colony (ABC) optimization algorithm to remove the skull region from brain MR images. We modify the ABC algorithm using a different strategy for initializing the coordinates of scout bees and their direction of search. Moreover, we impose an additional constraint to the ABC algorithm to avoid the creation of discontinuous regions. We found that our algorithm successfully removed all bony skull from a sample of de-identified MR brain images acquired from different model scanners. The obtained results of the proposed algorithm compared with those of previously introduced well known optimization algorithms such as Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO) demonstrate the superior results and computational performance of our algorithm, suggesting its potential for clinical applications. Keywords: Skull bone region, particle swarm optimization (PSO), ant colony optimization (ACO), artificial bee colony (ABC), MRI segmentation
1. Introduction The segmentation of brain magnetic resonance (MR) images is an essential preprocessing step which greatly affects subsequent analysis, the results of which can be used to formulate the next medical actions in patient management such as diagnosis and treatment. For decades, many algorithms have been introduced for segmenting MR images but often they can only be applied to data sets generated from a specific vendor’s MR system for which the algorithm was developed. The main limitations for generalizing such algorithms to all or most systems include the lack of a unique standard MR acquisition protocol and more importantly the difficulty introduced by the existence of the skull in the images. Indeed, before segmenting the brain itself it is necessary to identify and remove the structures such as skull that are not required for the tissue classification task so they can be excluded from the new input data set used for subsequent processing. This is a difficult preprocessing task that significantly affects the success of segmentation algorithms of brain tissue [15]. In this paper, we propose a semi-automated algorithm for removing the skull and bony structures from MR images that is generalizable to multiple MR scanners and acquisition protocols. In some previous work, the skull removal step is done manually which is very time consuming and is not recommended c 2014 – IOS Press and the authors. All rights reserved 0928-7329/14/$27.50
776
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
when the ultimate goal is the objective measurement of brain volume [16,17]. Other approaches consist of applying proper filters especially designed to remove the skull such as edge or boundary detection filters, thresholding and region growing algorithms [1]. However, these methods require adjustment of filter parameters for each specific acquisition protocol and each vendor device. For instance, a region growing algorithm can be used to segment the brain image into skull and non-skull structures and then to remove the skull [2,5,6]. However, in this algorithm a seed point belonging to the skull region must be manually selected by an operator or specialist, and errors in seed selection can be propagated through the algorithm resulting in overall poor performance. In this paper, we developed and implemented an optimization algorithm based on the artificial bee colony (ABC) algorithm to identify and remove bony skull structures in MR images of the head. The ABC technique has been recently proposed [7] based on real bee behavior for finding foods. Optimization algorithms are typically used to solve complicated mathematical problems in a simple way. One of the commonly used optimization algorithms is Particle Swarm Optimization (PSO) which was first introduced by Kennedy and Eberhart [8]. It was inspired by social behavior and movement dynamics of insects, birds and fish. It is a global gradient-less stochastic search method which is well suited for continuous variable problems. Due to its high speed, low computational cost, easy encoding and memory aspect of particles, PSO is used widely for different optimization problems and applications such as edge detection [9]. Despite its simple implementation, few required input parameters and global search ability, PSO does not perform well for local search tasks, something which is especially important for image processing of local features as required in segmentation algorithms. Another important optimization algorithm was proposed by Dorigo et al. [10] called Ant Colony Optimization (ACO). It is based on ants’ behavior in finding foods in the real world. It was shown that ants use colony intelligent action which is random and independent to connect with each other in indirect ways. This algorithm has many applications such as edge detection [11] and image segmentation [12]. The ACO algorithm is inherently parallel, very efficient to adapt to changes in problem constraints and powerful for global search. However, it results in a theoretical solution with uncertain (although guaranteed) time to convergence. Optimization algorithms have been successfully used for multimodality medical image registration [13,14], but in general, for medical image segmentation, few methods have been proposed based on optimization algorithms. One of the main challenges in using optimization algorithms is the design of the cost function and the way to minimize this function. In this paper, we propose a new simple criterion in the ABC algorithm especially developed to segment out the skull region in MR images and to remove it. Our algorithm does not depend on scanner protocols or specific devices, and can be extended to segment other organs. 2. Artificial Bee Colony (ABC) algorithm The Artificial Bee Colony (ABC) algorithm is a relatively new optimization algorithm based on insect’s behavior in detecting, tracking and finding food. The algorithm defines three types of bees; employed bees, onlooker bees and scout bees. Bees in a colony cleave into two teams of bees (employed and onlooker bees), where one employed bee is assigned to gather information about an available food source in a predefined search space, such that, all food sources and the total search space of a problem are covered by the group of employed bees. Then the employed bees share their information about food sources gathered from their search spaces with the onlooker bees. The onlooker bees sort the food sources from the richest nectar to poorest nectar. Then the onlooker bees select the employed bee having
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
777
the information about the richest nectar as follows. The chance of an employed bee being selected by onlooker bees can be computed from following equation. F (θi ) , Pi = S k=1 F (θi )
(1)
Where Pi is the chance of selecting the employed bee, F (θi ) is amount of nectar in food source i, θi is the food source i, Sk=1 F (θk ) is the total amount of nectars available in all food sources in the search space, k represents the employed bee and their specific search space, with the maximum probability occurring for (the employed bee with information about) the richest nectar. After being selected, the employed bee becomes a scout bee and starts gathering recruitments to move toward the target, i.e. the richest food source. Their goal is to vacate the target nectar and to search the other food sources in local nearest neighborhoods. Upon return to the hive, the scout bee shares the information obtained from the neighborhoods by the onlooker bees. Then the sort of food sources from richest to poorest will be repeated. This process continues until the optimal solution is found. In brief, the steps of the proposed method are as follows: Step 1. The movement of employed bees toward finding food sources and determining their content. Step 2. Sharing the information gathered by the employed bees with the onlooker bees and sorting the food sources based on their value (e.g., nectar). Step 3. Selecting a scout bee and moving toward the richest selected food source, searching its neighborhoods and vacating it. Step 4. Stopping algorithm when an optimal solution is achieved, otherwise going to step 2 and repeat. 3. Methods The success of brain tissue segmentation is highly dependent on removal of bony structures, (e.g., skull) from the image. First the region containing the skull must be determined and then removed from the MR images. To accomplish this, we modified the ABC algorithm and adapted it for medical image segmentation. The new algorithm is effective because the brain tissue is surrounded by cerebrospinal fluid (CSF) which is a region of gray values with little variation. To begin, we consider all pixels of a given MR image as the search space, and we assign a probability to each pixel in proportion to its intensity value. In this way, white pixels and black pixels have maximum and minimum probability respectively. All other pixels with their respective gray values, including pixels in the target region have probabilities between these minimum and maximum probabilities. After assigning the probability to each pixel, we define an initial interval of probabilities as food sources for scout bees. To do that, a mean probability μ and a variance σ 2 are initialized based on the probability of pixels belonging to CSF region in different MR images. In the conventional ABC algorithm, the direction and coordinates of scout bees are initially selected randomly. However, in our proposed algorithm we impose a direction of search. Namely, we select four initial scout bees (pixels) in the extreme right, left, top and bottom of a brain MR image and require their direction of movement top toward the center of the image (see Fig. 1). Each scout bee searches the space of search in his proscribed direction. Every time a scout bee finds a food source, he shares the information obtained from this food source with others. The probability of the new pixel being considered as a part of CSF region is based on other pixels previously joined to the CSF region. Consequently the mean μ and σ 2 of CSF will be updated every time a pixel becomes a part of
778
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
Fig. 1. Initialization of coordinates and the search direction of scout bees. (Colours are visible in the online version of the article; http://dx.doi.org/10.3233/THC-140845)
CSF region. In this way, an associative memory is created and shared among scout bees. However, it is probable that some discontinuous regions can be formed. In order to avoid the creation of such regions, we impose another criterion for considering a pixel as a part of final region. This criterion is the distance between the intensity values of two adjacent pixels which is defined as: DSij = Ii2 − Ij2 , (2) Where DSij is the distance between two pixels i and j , Ii and Ij are their intensity value respectively. If the distance between two adjacent pixels is near to zero, it means that the two pixels are more likely a part of the same region. This constraint assists in preserving the continuity of regions. The pseudo code of the proposed algorithm is as follows: Input: intensity pixels of a brain MR image (I(i, j)). Output: identifying and labeling pixels belonging to skull region (ISkull (i, j)). Algorithm: Step 1. (Initialization): initialing E (the number of employed bees), O (the number of onlooker bees) such that E = O and S = 4 (the number of scout bees = orientation) where should it be the following with an M (rows) × N (columns) matrix (i1 = 1 to M/2, j1 = 1 to N means search down), (i2 = 1 to M , j2 = 1 to N/2 means search right) (i3 = M to M/2, j3 = 1 to N means search up) (i4 = 1 to M , j4 = N to N/2 means search left) are initial coordinates and searching directions of scout bees for an image of size M × N . Step 2. (Searching process): Set the food source intensity (FSI) equal to the intensity of CSF around the brain and begin the search process. Step 3. (Growing CSF region): Save last intensity to employed bees’ memory (Memployedbees = ISkull ). Step 4. (Stopping criterion): resume movement while criterion (I(i, j) < FSI) for all four search areas is satisfied, checking the distance criterion for neighboring pixels by using Eq. (2).
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
779
Fig. 2. A. Example images from eight de-identified MRI brain examinations. B. Segmented brain results of proposed method.
780
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
Fig. 2. continued.
∗ ∗ ∗ Fitness Criterion ∗ ∗∗ − − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
Move from up to down n = 0 : M/2 − 1
if
I(i1 + n, j1 ) < FSI then
Iskull = I(i1 + n, j1 )
− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
Move from left to right n = 0 : N/2 − 1
if
I(i2 , j2 + n) < FSI then
Iskull = I(i2 , j2 + n)
− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
Move from right to left n = N : N/2
if
I(i1 , j1 − n) < FSI then
Iskull = I(i1 , j1 − n)
− − − − − − − − − − − − − − − − − − − − − − − − − − − − − −−
The specifications of the computer on which the algorithms were run are as follows: Machintosh OS, Two 2.93 GHz 6-Core Intel processors, 64 GB RAM, HD 5870 1 GB HD, and ATI Radeon Graphics Card.
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
781
Fig. 3. From left to right Original image, results of PSO, results of ACO, results of ABC, results of our proposed method.
4. Experimental results We used several test brain images using MR scanners of 1, 1.5 and 3 Tesla (GE Medical, Waukesha, WI). All test images have size 256 × 256 pixels with gray scales of 16 bit depth and were acquired at Faghihi Hospital in the Department of Radiology in Shiraz, Iran. Figure 2 shows the results of proposed method for several example images from eight de-identified cases of MRI brain examinations. We also compared the results of proposed method with PSO, ACO and conventional ABC algorithms. Figure 3 shows the comparative results. As can be seen, other methods fail to remove skull region
782
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm Table 1 Parameters used in the tested methods PSO Number of swarm C1 C2 ωmax ωmin Number of iterations ACO Number of ants Probability threshold for maximum trail Local search probability Evaporation rate Number of iterations ABC Number of bees (employer and onlooker) (Conventional ABC and Proposed ABC) Number of sites selected for neighborhood search Number of bee recruited for best sites Number of iterations
100 1.9 2.1 1 0.5 500 50 0.95 0.01 0.01 1000 20 10 5 500
Table 2 Cost (CPU time) obtained for each different methods Row of images in Fig. 3 (From top to bottom) 1
2
3
4
5
6
Method PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC PSO ACO ABC Proposed ABC
Real time in second 0.0009 0.0026 0.0005 0.00000102 0.00232 0.00764 0.00187 0.0000627 0.0378 0.0674 0.0139 0.0123 0.0196 0.0362 0.0128 0.0016 0.0096 0.0361 0.0082 0.00014 0.0067 0.0092 0.0052 0.00082
completely, while proposed method achieves significantly superior performance in completely removing the skull region. The results also show that the proposed algorithm is independent of scanner modality and it suggests a promising clinical application which could be used by radiologists and physicians. Table 1 shows the parameters used in each method in our experiments. The cost or CPU time of each method for each of the test images in Fig. 3 were computed and added in Table 2.
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm
783
5. Discussion Medical image segmentation can be considered as dividing an image into homogenous regions that are often a part of desired objects in the medical image. Many criteria can be used to measure the homogeneity of a region such as pixel intensity. The medical image segmentation is a pre-step that is relatively complex in medical image processing. It can be also considered as a clustering process in which pixel with similar features can be assigned to a cluster. So, clustering methods have widely used in medical image segmentation for separating different organs in medical images [1–3,8–10]. In this paper, it was tried to select the test images in which the pixel intensities of neighboring organs are so close that the extraction of CSF (Cerebrospinal Fluid), WM (White Matter), and GM (Gray Matter) and also removing skull are a difficult task to do. We should state that we have implemented and evaluated different meta-heuristic methods including PSO, ACO and ABC and also some existing tools for skull removal based on region growing and Watersheld. The best results obtained from these previously used methods are brought in experimental results section of the paper their performance have been evaluated and compared with the proposed method. In addition, the execution time of each method has been computed and brought in Table 2. As can be seen from Table 2, the proposed algorithm has achieved the final response in much shorter time than other methods. It is worth to mention that PSO and ACO take more execution time related to other methods because they have many loops and use many If-Then conditions, and so they cannot be applicable in real time applications. Although the existing tools can obtain the final response within an acceptable time, the resulted segmentations are not so good and thus the next steps using these results may face on problems. So they are not suggested for clinical applications. From the images in the third column of Fig. 3 we can obviously see that PSO like ACO did not produce the acceptable results and they both should be reconsidered for being used in brain image segmentation applications. By focusing on fifth column of Fig. 3, the proposed method produce the acceptable results. 6. Conclusion A new algorithm for skull removal from MR Brain images has been proposed based on a modified ABC optimization algorithm. The ABC algorithm is a new powerful optimization algorithm which models the behavior of bees in finding rich sources of foods. Experiments on example images from eight deidentified MRI brain examinations show that our ABC algorithm achieves superior results as compared to other optimization algorithms such as PSO and ACO, successfully removing all bony skull from the images, and is computationally faster. However, its application to image segmentation has not been fully exploited. In this paper, we use a modified ABC algorithm for segmenting the brain MR images into skull and non-skull regions and then remove the skull region. Comparative results have demonstrated the superior performance of the algorithm related to other optimization based algorithms. In addition, our proposed algorithm performs very well across different MR scanners, which shows its potential for use in a clinical framework. References [1]
Ballard DH, Brown DM. Computer vision. Prentice Hall, Englewood Cliffs, NJ; 1982.
784 [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17]
M. Taherdangkoo / Skull removal in MR images using a modified artificial bee colony optimization algorithm Haralick RM, Shapiro LG. Image segmentation techniques. Computer Vision, Graphics, and Image Processing. 1985; 29(1), 100–132. Saha S, Maulik U. A new line symmetry distance based automatic clustering technique: Application to image segmentation. International Journal of Imaging System and Technology. 2011; 21(1), 86–100. Odille F, Steeden JA, Muthurangu V, Atkinson D. Automatic segmentation propagation of the aorta in real-time phase contrast MRI using nonrigid registration. Journal of Magnetic Resonance Imaging. 2011; 23(1), 232–238. Gholipour A, Estroff JA, Barnewolt CE, Connolly SA, Warfield SK. Fetal brain volumetry through MRI volumetric reconstruction and segmentation. International Journal of Computer Assisted Radiology and Surgery. 2011; 6(3). 329– 339. Wright R, Kyriakopoulou V, Ledig C. Rutherford MA, Hajnal JV, Rueckert D, Aljabar P. Automatic quantification of normal cortical folding patterns from fetal brain MRI. NeuroImage. 2014; 91, 21–32. Kemal Apalak M, Karaboga D, Akay B. The artificial bee colony algorithm in layer optimization for the maximum fundamental frequency of symmetrical laminated composite plates. Engineering Optimization. 2014; 46(3), 420–437. Clerc M, Kennedy J. The particle swarm – explosion, stability, and convergence in a multidimensional complex space. IEEE Transactions on Evolutionary Computation. 2002; 6(1), 58–73. Setayesh M, Zhang M, Johnston M. A novel particle swarm optimization approach to detecting continuos, thin and smooth edges in noisy images. Information Sciences. 2013; 246, 28–51. Dorigo M, Thomas S. Ant Colony Optimization. MIT Press. 2004. Cerello P, Cheran SC, Bagnasco S, Bellotti R, Bolanos L, Catanzariti E, Nunzio GD, Fantacci ME, Fiorina E, Gargano G, Gemme G, Torres EL, Masala GL, Peroni C, Santoro M. 3-D object segmentation using ant colonies. Pattern Recognition. 2010; 43(4), 1476–1490. Taherdangkoo M, Bagheri MH, Yazdi M, Andriole KP. An effective method for segmentation of MR brain images using the ant colony optimization algorithm. Journal of Digital Imaging. 2013; 26(6), 1116–1123. Valsecchi A, Damas S, Santamaria J, Marrakchi-Kacem. Intensity-based image registration using scatter search. Artificial Intelligence in Medicine. 2014; 60(3), 151–163. Manikandan S, Ramar K, Willjuice Iruthayarajan M, Srinivasagan KG, Multilevel thresholding for segmentation of medical brain images using real coded genetic algorithm. Measurment. 2014; 47, 558–568. Jeroen de Bresser, Portegies MP, Leemans A, Biessels GJ, Kappelle LJ, Viergever MA. A comparison of MR based segmentation methods for measuring brain atrophy progression, NeuroImage. 2011; 54(2), 760–768. Battaglini M, Smith SM, Brogi S, Nicola De Stefano. Enhanced brain extraction improves the accuracy of brain atrophy estimation, NeuroImage. 2008; 40(2), 583–589. Segonne F, Dale AM, Busa E, Glessner E, Salat D, Hahn HK, Fischi B. A hybrid approach to the skull stripping problem in MRI. NeuroImage. 2004; 22(3), 1060–1075.
Copyright of Technology & Health Care is the property of IOS Press and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
