A 3-D Liver Segmentation Method with Parallel Computing for ...

62

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

A 3-D Liver Segmentation Method with Parallel Computing for Selective Internal Radiation Therapy Mohammed Goryawala, Magno R. Guillen, Mercedes Cabrerizo, Armando Barreto, Seza Gulec, Tushar C. Barot, Rekha R. Suthar, Ruchir N. Bhatt, Anthony Mcgoron, and Malek Adjouadi

Abstract—This study describes a new 3-D liver segmentation method in support of the selective internal radiation treatment as a treatment for liver tumors. This 3-D segmentation is based on coupling a modified k-means segmentation method with a special localized contouring algorithm. In the segmentation process, five separate regions are identified on the computerized tomography image frames. The merit of the proposed method lays in its potential to provide fast and accurate liver segmentation and 3-D rendering as well as in delineating tumor region(s), all with minimal user interaction. Leveraging of multicore platforms is shown to speed up the processing of medical images considerably, making this method more suitable in clinical settings. Experiments were performed to assess the effect of parallelization using up to 442 slices. Empirical results, using a single workstation, show a reduction in processing time from 4.5 h to almost 1 h for a 78% gain. Most important is the accuracy achieved in estimating the volumes of the liver and tumor region(s), yielding an average error of less than 2% in volume estimation over volumes generated on the basis of the current manually guided segmentation processes. Results were assessed using the analysis of variance statistical analysis. Index Terms—3-D reconstruction, image segmentation, liver segmentation, MATLAB, k-means algorithm, parallel computing.

I. INTRODUCTION HE American Cancer Society estimates that nearly 18 000 people are diagnosed with liver cancer annually. This unfortunately makes 1 out of 183 individuals born today susceptible to liver cancer at some point during their lifespan [1]. Selective internal radiation therapy (SIRT) with Yttrium-90 (Y90) microspheres is emerging as an effective liver-directed therapy [2]. The treatment involves the accurate calculation of functional tumor volumes and anatomical volumes of liver for the

T

Manuscript received April 15, 2011; revised August 22, 2011; accepted September 25, 2011. Date of publication October 10, 2011; date of current version February 3, 2012. This work was supported by the National Science Foundation under Grant CNS-0959985, Grant HRD-0833093, and Grant CNS1042341. M. Goryawala, R. N. Bhatt, and A. Mcgoron are with the Department of Biomedical Engineering, Florida International University, Miami, FL 33199 USA (e-mail: [email protected]; [email protected]; [email protected]). M. R. Guillen is with Radiology Department, Miami Children Hospital, Miami, FL 33155 USA (e-mail: [email protected]). M. Cabrerizo, A. Barreto, and M. Adjouadi are with the Department of Electrical and Computer Engineering, Florida International University, Miami, FL 33199 USA (e-mail: [email protected]; [email protected]). S. Gulec, T. C. Barot, and R. R. Suthar are with the Herbert Wertheim College of Medicine, Florida International University, Miami, FL 33199 USA (e-mail: [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TITB.2011.2171191

determination of the tumor to normal liver ratio and consequently for the calculation of the dose of Y-90 microspheres [3]. Present techniques for the determination of the anatomic volume of the liver involve tedious manual segmentation of the liver from the computerized tomography (CT) scans. The number of slices in a CT dataset varies from 200 to 400, which makes the task of computing the volume of the liver manually excessively tedious and time consuming. Also, the task is greatly dependent on the skill and proficiency of the technician/doctor, which could contribute to human error and skew the results. Current automatic and semiautomatic procedures for liver segmentation are based on techniques that rely on: 1) shapeconstrained segmentation using heuristic approaches [4], local shape models [5], atlas-based techniques [6], [7] or nonlinear models [8], [9]; 2) rule-based segmentation [10]; 3) gradient vector flow [11]; and 4) 2 or 3-D region growing [12]. Although these techniques offer highly accurate results, the algorithms need to accommodate varying protocols, data from different sources, artifacts, and the presence of pathological structures such as tumors [13]. In order to overcome inherently high computational requirements that tasks such as segmentation and volume calculation demand, a parallel computing process within a single desktop computer is employed. Such tasks are hence distributed across the cores/processors of the computer. MATLAB provides effective different alternatives to address computationally intensive tasks based on issues of parallelism and high-performance computing (HPC). Krishnamurthy et al. presented an exhaustive description of all the HPC techniques offered by MATLAB [14]. The most popular techniques are MATLAB-MPI [15], [16], pMATLAB [17], Star-P [18], [19], the MATLAB Distributed Computing Toolbox (DCT) [20], and the Parallel Computing Toolbox (PCT) [21], [22]. In this study, PCT and DCT are selected to set up the parallel computing network. This computational process is structured in a way that its deployment can extend to clusters and grids [23], [24], if the need arises. II. RESEARCH AIM The research aim is to design a new algorithm for accurate and fast liver volume calculation using little as possible user intervention and yet maintaining the accuracy of the obtained results. This algorithm semiautomatically segments the liver region from 3-D CT scans. The algorithm extracts the liver region and renders the segmented liver for 3-D viewing. The algorithm is also structured as a parallel-aware process so that a computationally taxing task can be distributed over various computing nodes.

1089-7771/$26.00 © 2011 IEEE

GORYAWALA et al.: 3-D LIVER SEGMENTATION METHOD WITH PARALLEL COMPUTING FOR SIRT

Fig. 1.

63

General algorithm for 3-D liver segmentation and rendering.

III. METHODOLOGY Liver segmentation based on CT images is a challenging task due to the presence of similar intensity objects in the abdomen with no clear delineation between these objects and the liver. These objects include the spleen, stomach, wall of the abdomen, and kidneys. A new method for liver segmentation has, thus, been developed based on a combination of a modified k-means clustering process and the active contours algorithm. The implementation strategy is illustrated in Fig. 1. Although reduced manual interaction is desired, the variations in the liver datasets seen in patients having tumor along with the need to obtain accurate volume determination for the SIRT treatment justify the use of the proposed semiautomatic method. The proposed method used to select the slices requiring human interaction is based on tracking the changes observed in the liver region across the slices. The CT scans of the patients are deidentified on-site. This process of deidentification is performed to conform to the Internal Review Board on human subjects (IRB approval 051810-01) and to the standards and guidelines established by the Department of Health and Human Services-HIPAA. A. Design Structure of the Algorithm At this juncture, the algorithm as designed requires the user to manually pick five points of varied intensity in the scan for the k-means segmentation as well as a rough outline of the liver in widely spaced slices throughout the given CT dataset. From this initialization process, the following automated steps are then considered. 1) k-means Algorithm: The k-means clustering is a traditional clustering technique that tries to partition the given dataset points into various clusters whose means are similar [25], [26]. In the case of liver segmentation, the CT slices are portioned into five regions whose mean intensity levels can be given either by the user or preset. The five regions identified are liver, surrounding organs, peripheral muscles, ribs/spinal cord, and the outside of the body. It was observed that using five regions yielded the best results. Fewer regions resulted in poor separation between the liver and surrounding organs whereas using a larger number

Fig. 2. Five reference regions, M1 through M5 , needed to apply the k-means based segmentation method on a liver CT dataset.

of regions distributed the liver region over various masks and affected the accuracy of the segmentation. For the modified k-means approach adopted in the procedure segments the different regions of the CT slice around the user selected points. The user selected points act as the seeds for each of the five masks. The selection of the seeds, rather than a random selection or uniform selection of points in the range, yielded much better segmentation results. To obtain results free of misclassifications, a logical intersection of three independent segmentation runs are employed. Also the “online update mode” that calculates the sum of the distances with the movement of every pixel is used to obtain higher accuracies at the cost of processing time. The segmentation results using the k-means yield five masks namely M1 through M5 corresponding to the aforementioned five regions are depicted in Fig. 2. The masks liver (M1 ) and surrounding organs (M2 ) are logically ORed together to obtain the final mask (Mfinal ). Based on empirical results, it was determined that the optimal mask would require a combination of the first two identified regions since in some cases the entire liver is not seen in mask M1 due to the inhomogeneous intensity distribution across the entire liver region in the CT scans. 2) Initialization of the Contouring-Based Segmentation: One of the major issues that concern contour-based segmentation processes is the initialization of the contour. This contourbased segmentation algorithm takes a user-defined mask as input for the initialization algorithm. The user marks the approximate boundaries of the liver using a mouse pointer in slices that are widely apart from one another. However, the selection of the slices in which the user picks the initial contour depends on the change in information (presence of liver) from slice to slice. This is achieved by first dividing the entire image obtained after the modified k-means segmentation into blocks of 16 × 16 pixels. The selection of blocks of 16 × 16 pixels was determined to be the best solution. Inside each block, the pixels that fall in the abdominal CT window of −40 HU to 180 HU were counted. The choice of the CT window is based on the fact that liver and other soft tissue

64

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

function HΦ(x) as ⎧ 1, Φ(x) < −ε ⎪ ⎪ ⎪ ⎨ 0, Φ(x) > ε HΦ(x) =   ⎪ ⎪1 1 πΦ(x) Φ ⎪ ⎩ 1 + + sin , otherwise 2 ε π ε (2) where Φ(x) is a smoothened initial contour and [−ε, ε] represents the boundary of the Heaviside function. In reference to (2), the exterior of closed contour C can then be expressed as {1 − HΦ(x)}. In order to calculate only the local energies, localization radius α is defined in terms of the size of the image and is calculated using (3) Fig. 3. Generation of dataset profile curve depicting the slices corresponding to the marked points on the dataset profile curve with black bars showing slices marked for manual initialization.

are predominantly found in this window [12], [27]. Also, the CT window has been extended slightly to incorporate the variations in the dataset utilized for the study. A block was marked as being part of the liver region if at least more than half of the pixels were found in the abdominal CT window as described in (1). Thus  block marked =

1 0

n ≥ 0.5 N otherwise

if

(1)

where n is the number of pixels in the abdominal window, and N is the total number of pixel elements in the block. Fig. 3 shows a typical profile curve for dataset 1 that displays the number of blocks belonging to the liver window as a function of the slice number in a particular dataset and the slices used to calculate the value of the curve at some particular instant of time. The first slice of the CT dataset was always selected for manual contour initialization. In this algorithm, in order to minimize the number of slices considered for contour initialization, a subsequent slice would need to show a 5% change in the number of blocks as marked in (1) with respect to its predecessor. The intermediate slices that are not selected for manual initialization used the same initial contour as the slice closest to it in the dataset. This approach yielded around 4–5% of the entire dataset to be initialized by the user, rather than the time consuming slice by slice contouring of the entire dataset that is seen in clinical practice. The process yielded segmentation results that were not affected and the volume measurements that were highly accurate. As an example, the slices chosen for manual initialization are marked by the black bars in Fig. 3 for a particular dataset. 3) Localized-Region-Based Active Contouring Algorithm: Once the initialization for all the slices of the dataset is obtained interactively, a localized-region-based active contouring algorithm is employed. Localized region growing algorithms are more robust than contouring algorithms not based on global energies for segmenting heterogeneous objects like the liver [28]. The interior of the contour is given by the smoothened Heaviside

1 (Pr + Pc ) (3) 4 where Pr is the number of rows in the image and Pc is the number of columns in the image. The mask generated by using the aforementioned value of α is given by (4)  1 ||x − y|| < α . (4) Mα = 0 otherwise α=

The algorithm incorporates the well-known Chan–Vese energy paradigm to model the interior and the exterior of the contour for segmentation [29]. The localized version of the Chan–Vese energy function is given by (5) ECV = HΦ(y)(S1 (y) − μint )2 + (1 − HΦ(y)(S1 (y) − μext )2 dy

(5)

where μint and μext are the mean intensities of the interior and exterior regions of the contour, respectively. The localized versions of the mean intensities are obtained by restricting the field of view only to the masked region given by masking the image with Mα . B. Need for Parallel Processing Parallel computing is employed to reduce the computational times by distributing the task between more than one core/processor available on a single workstation, where a number of slices of the dataset will be processed simultaneously (8 in this case). The expected improvement in the computational time Tc is referred here as the “speed up” of the process, and is computed using the following ratio: Tc =

C1 Cn

(6)

where C1 and Cn denote the computational time required to complete the task using a single worker and n workers, respectively. All the experiments were conducted on a system running Windows Vista based on an HP DC6700 computer with 2 Intel Core 2 Quad Processor operating at 2.67 GHz with 8 GB RAM. The use of the software and hardware configurations was determined by benchmarking experiments using various hardware and operating system configuration [30].

GORYAWALA et al.: 3-D LIVER SEGMENTATION METHOD WITH PARALLEL COMPUTING FOR SIRT

65

TABLE I CHARACTERISTICS OF THE FIVE DATASETS

Fig. 4.

Five slices showing detailed segmentation results for dataset 5.

C. 3-D Image Rendering The 3-D datasets are rendered using cost-effective third party software called ScanIP developed by Simpleware Ltd., based in the United Kingdom. The software renders the segmented dataset in 3-D space and offers the possibility for the physician to view/edit/correct the rendered liver if necessary. The software also calculates the volume of the liver by determining the number of voxels that are marked as being within the liver region by the segmentation algorithm. The only inputs fed to the software are the segmented dataset and the original resolution of the CT datasets. The calculated volumes are in milliliters (mL). D. Experimental Evaluation Five datasets are used for testing the developed algorithm. The characteristics of the datasets used are given in Table I. In order to validate the segmentation results obtained by the algorithm, the volumes of the extracted livers were compared to manually calculated volumes. The manual calculation was done by an expert and is treated as the gold standard for the comparison that is routinely the suggested method [13]. Also, statistical analysis is provided to show the consistency of the results obtained in various runs and to determine the dependence of the algorithm on the user utilizing it. One-way analysis of variance (ANOVA) is used to compare the means of two or more samples and to determine the effect of the variable (runs or users) on the calculated volumes. Furthermore, improvements in processing speed were achieved for the liver segmentation process using parallel computing by deploying the algorithm on 1, 2, 4, and 8 processing workers, successively. IV. RESULTS A. Segmentation Results Fig. 4 shows the segmentation results (shown in red) obtained for a particular dataset (Dataset 5). Fig. 4 shows five slices across the database. The slices are displayed in the [−40 180] window for the CT dataset. Fig. 5 displays the results for other datasets used in the validation of the algorithm. The point to be noted is that there is a great variation in the intensity, structure, and

Fig. 5. Example of segmentation results for other evaluated datasets. Each column shows two slices. (a) Dataset 1, (b) Dataset 2, and (c) Dataset 4. TABLE II COMPARISON OF VOLUMES CALCULATED ON FIVE DATASETS

position of the liver from dataset to dataset. Moreover, some of the datasets show the presence of tumors. Comparative results are shown in Table II. Results show that the volumes of the liver calculated by the algorithm are in agreement with that of the medical experts. Across the five evaluated datasets we see a mean accuracy of as high as 98.27%. Results of the 3-D rendering process obtained from the ScanIP software are shown in Fig. 6 for the different datasets (a) through (e). The 3-D rendering displayed in Fig. 6(a) is for the same dataset for which the slices were shown in Fig. 5. The renderings shown here have solid surfaces and a mesh finish. However, translucent surfaces can be generated with varying opacities and colors if needed. The need to generate translucent surfaces would be useful to demonstrate the presence of a tumor inside the liver as an example.

66

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

TABLE IV PROFILE OF USERS SELECTED FOR THE EXPERIMENTS

TABLE V (a) ABSOLUTE ERROR OBTAINED USING DIFFERENT USERS FOR INITIALIZATION (b) ANOVA ANALYSIS FOR VARIOUS USERS

Fig. 6. (a–e) Display 3-D renderings of livers in dataset 1–5 using ScanIP software, respectively. TABLE III (a) ABSOLUTE ERROR OBTAINED DURING THE VARIOUS RUNS (b) ANOVA ANALYSIS OF VARIOUS RUNS

TABLE VI COMPARISON OF SOME OF THE CURRENT TECHNIQUES AND THE PROPOSED TECHNIQUE

A statistical analysis is also carried out to determine if the changes seen in the calculation of the volumes between different trials are significant. For the analysis, an ANOVA is carried out on the absolute error obtained during the various runs. Table III(a) shows the errors in the various runs, and Table III(b) shows the results of the ANOVA. The results of the ANOVA show that the trials are not a significant factor (p > 0.05). It hence shows that the algorithm does provide consistent results. In order to determine if the user initialization had impact on the results obtained by the algorithm, three users with different levels of knowledge in the field of medicine and engineering were asked to initialize the algorithm. The users were asked to rate themselves on their knowledge of liver anatomy and algorithms/programs on a scale of 1 to 5. Also, all users were given a basic understanding of the liver anatomy to help them identify the liver in the CT scans for the initialization purpose. Table IV shows the profiles of the different users selected for the purpose of testing the algorithm. An ANOVA was again carried out on the errors obtained after the segmentation process initialized by the three users.

The errors calculated as the absolute difference between the algorithm calculated volumes and the manual gold standard volumes are shown in Table V(a) with the ANOVA results shown in TABLE V(b). The results of the aforementioned ANOVA show that the error in the volume calculation is not dependent on the user who initializes the dataset since the factor “Users” is not significant with a p-value of 0.84 (p > 0.05). Table VI shows a comparison between some of the current techniques and the proposed technique in terms of the accuracy and the computational times required to segment the liver.

GORYAWALA et al.: 3-D LIVER SEGMENTATION METHOD WITH PARALLEL COMPUTING FOR SIRT

Fig. 7.

Computational results of the parallel processing approach.

For the comparison of accuracy, the average volume difference between the calculated and the manual volumes is presented along with the standard deviation for the particular study. For comparing the computational times, Table VI provides the time taken per slice to segment the entire dataset. Such a comparison is essential since different algorithms use datasets with different number of slices for the analysis that determines the time needed for processing the entire dataset. B. Parallel Computing Results Fig. 7 displays the computational time for the different datasets as a function of the number of workers employed for the task. We see that parallel processing reduces the time for the segmentation for dataset 5 from around 270 min to as low as 56 min. This is the highest processing time experienced among all the datasets since dataset 5 is the largest. The capability to complete the whole segmentation process in less than 1 h as compared to 4.5 h is a great improvement in context of hospital’s efficiency. The reduction in computational time results in an average speed up of around five for eight workers employed for the parallel processing task. V. DISCUSSION Fig. 3 displays a typical profile curve for a liver dataset with the black bars marking the position of the slices in which human interaction is needed. It can be observed from the figure that the number of slices picked for human initialization is directly related to how fast the liver volume is changing across the dataset. Hence, regions on the profile curve that have a higher slope or greater changes in the liver structure have more slices selected for human initialization. Regions that show little or no changes in the structure use the same initialization and hence a reduction in the human interaction can be achieved without compromising the accuracy. Fig. 4 displays visually that the segmentation process is capable of extracting the liver region faithfully. The liver region is, thus, delineated from the heart, spleen, and other organs. Very little leakage to the adjoining organs is seen due to the use of local properties for the contouring algorithm. Also, Fig. 5 showed that the segmentation is independent of the structure, size, position, and intensity distribution of the liver region. It can be observed that the intensity distribution is not similar across the three liver datasets displayed in Fig. 6(a–c) and the algorithm

67

faithfully identifies and segments the liver region. Also, the size of the liver is vastly different in Fig. 5(a), where the patient has undergone an earlier resection. This makes the algorithm capable of use in posttreatment follow-up also. Such a high accuracy is very advantageous for treatment planning in SIRT. SIRT relies on the calculation of the tumor to liver volume ratio for the calculation of the radioactive dose to the patient. Accurate volume calculation specific to each patient’s anatomy would help in the calculation of the absolute precise dose to deliver to that patient. This would reduce the risk for excess dosing that may damage healthy surrounding liver tissue or on the other end of the estimation reduce the risk for under dosing and the potential relapsing of the tumor. As seen in Table VI, the proposed technique is highly accurate in estimating the liver volume. An important feature to note is the low standard deviation in the volume calculation that suggests high consistency in the results. The algorithm fairs well in respect with the computational times required to segment the liver. It should be noted that since the algorithm is parallel aware and capable of being deployed on larger clusters, the computational times can be brought down considerably. In retrospect, the contribution of this study is in developing a highly accurate segmentation algorithm for 3-D treatment planning in SIRT-based therapy. The new techniques require minimal human intervention for the initialization process of segmentation, by incorporating information-based techniques to estimate the slices to be initialized by the user automatically. Also, the implementation of the localized contouring algorithm using kmeans segmentation is a novel idea that was integrated in this approach. This study also highlights the advantage of developing algorithms aware of parallel computing in medical imaging, before investing in a large scale distributed system or an expensive parallel machine, providing opportunity for accelerating the image processing tasks using a single desktop computer, especially in the advent of future processors with six or more cores on a single chip. This feature allows the algorithm to be easily portable and adaptable in hospital settings with minimal space requirements. More importantly, beyond the computational gains that are made under the proposed approach, the hybrid algorithm introduced in this study displays a high accuracy in calculating the volumes of different livers with a minimal number of slices to be initialized by the user. VI. CONCLUSION This study demonstrated the development of a robust and accurate method for the segmentation of the liver from CT images for the purpose of volume calculations. The algorithm is independent of the dataset-properties namely structure, size, position, and intensity distribution of the liver region due mainly to the novel hybrid segmentation method that coupled the modified k-means algorithm with a newly established contouring method that relied on radio density of the CT images. The experiment results reported in this study display very high accuracies of 98.27% in agreement with manual expert analysis. Also, the development of this initialization method, which aims at reducing

68

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

human/user interaction, achieved highly accurate results with only 4% to 5% of the dataset initialized by the user, minimizing greatly human intervention as a consequence. The proposed algorithm is also structured to lend itself to parallel computing, which is important in medical imaging applications that necessitate heavy computational requirements. The results of the parallelization of the liver image segmentation yield a significant speed up in the performance, achieving more than 500% increase by using eight MATLAB workers on a dual quad-core machine. Also, SIRT relies on the calculation of the tumor to liver volume ratio for the calculation of the radioactive dose to the patient. Accurate volume calculation specific to each patient’s anatomy would help in the calculation of the absolute precise dose to deliver to that patient. This would reduce the risk for excess dosing that may damage healthy surrounding liver tissue or reduce the risk of under dosing and the potential relapsing of the tumor. ACKNOWLEDGMENT The authors appreciate the support provided by the National Science Foundation under grants: CNS-0959985, HRD0833093 and CNS-1042341. The authors give special thanks to the Department of Oncology at the North Jackson Miami Hospital and to the Rinker Family Foundation for their clinical support. REFERENCES [1] A. Jemal, F. Bray, M. M. Center, J. Ferlay, E. Ward, and D. Forman, “Global cancer statistics,” Ca-a Cancer J. Clinicians, vol. 61, pp. 69–90, Mar./Apr. 2011. [2] W. Y. Lau, S. Ho, T. W. Leung, M. Chan, R. Ho, P. J. Johnson, and A. K. Li, “Selective internal radiation therapy for nonresectable hepatocellular carcinoma with intraarterial infusion of 90yttrium microspheres,” Int. J. Radiat. Oncol. Biol. Phys., vol. 40, pp. 583–592, 1998. [3] R. Murthy, R. Nunez, J. Szklaruk, W. Erwin, D. C. Madoff, S. Gupta, K. Ahrar, M. J. Wallace, A. Cohen, D. M. Coldwell, A. S. Kennedy, and M. E. Hicks, “Yttrium-90 microsphere therapy for hepatic malignancy: Devices, indications, technical considerations, and potential complications,” Radiographics, vol. 25 (suppl. 1), pp. S41–S55, Oct. 2005. [4] D. Kainm¨uller, “Shape constrained automatic segmentation of the liver based on a heuristic intensity model,” presented at the Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007. [5] D. Seghers, P. Slagmolen, Y. Lambelin, J. Hermans, D. Loeckx, F. Maes, and P. Suetens, “Landmark based liver segmentation using local shape and local intensity models,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 135–142. [6] D. Furukawa, A. Shimizu, and K. H., “Automatic liver segmentation based on maximum a posterior probability estimation and level set method,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 117. [7] E. Rikxoort, Y. Arzhaeva, and B. Ginneken, “Automatic segmentation of the liver in computed tomography scans with voxel classification and atlas matching,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 101–108. [8] T. Heimann, H. Meinzer, and I. Wolf, “A statistical deformable model for the segmentation of liver CT volumes,” in Proc. MICCAI Workshop 3-D Segmentation in the Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 161. [9] M. Saddi, K. A. Rousson, C. Chefd’hotel, and F. Cheriet, “Global-to-local shape matching for liver segmentation in CT imaging,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 207–214.

[10] G. Schmidt, M. A. Athelogou, R. Sch¨onmeyer, R. Korn, and G. Binnig, “Cognition network technology for a fully automated 3-D segmentation of liver,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 125–133. [11] Y. Chi, P. M. M. Cashman, F. Bello, and R. I. Kitney, “A discussion on the evaluation of a new automatic liver volume segmentation method for specified CT image datasets,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007. [12] L. Rusk´o, G. Bekes, G. Nemeth, and M. Fidrich, “Fully automatic liver segmentation for contrast-enhanced CT images,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 143–150. [13] T. Heimann, B. van Ginneken, M. A. Styner, Y. Arzhaeva, V. Aurich, C. Bauer, A. Beck, C. Becker, R. Beichel, G. Bekes, F. Bello, G. Binnig, H. Bischof, A. Bornik, P. M. M. Cashman, Y. Chi, A. Cordova, B. M. Dawant, M. Fidrich, J. D. Furst, D. Furukawa, L. Grenacher, J. Hornegger, D. Kainmuller, R. I. Kitney, H. Kobatake, H. Lamecker, T. Lange, J. Lee, B. Lennon, R. Li, S. Li, H. P. Meinzer, G. Nemeth, D. S. Raicu, A. M. Rau, E. M. van Rikxoort, M. Rousson, L. Rusko, K. A. Saddi, G. Schmidt, D. Seghers, A. Shimizu, P. Slagmolen, E. Sorantin, G. Soza, R. Susomboon, J. M. Waite, A. Wimmer, and I. Wolf, “Comparison and evaluation of methods for liver segmentation from CT datasets,” IEEE Trans. Med. Imag., vol. 28, no. 8, pp. 1251–1265, Aug. 2009. [14] A. Krishnamurthy, J. Nehrbass, J. C. Chaves, and S. Samsi, “Survey of parallel MATLAB techniques and applications to signal and image processing,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process., Honolulu, Hawaii, U.S.A., 2007. [15] D. E. Hudak, N. Ludban, V. Gadepally, S. Samsi, J. C. Chaves, and A. Krishnamurthy, “Computational Science IDE for HPCMP Systems,” in Proc. DoD High Performance Computing Modernization Program Users Group Conf., Pittsburgh, PA, 2007, pp. 317–321. [16] R. Choy and A. Edelman, “Parallel MATLAB: Doing it right,” in Proc. IEEE, vol. 93, no. 2, pp. 331–341, Feb. 2005. [17] N. T. Bliss and J. Kepner, “pMATLAB parallel MATLAB library,” Int. J. High Perform. Comput. Appl., vol. 21, pp. 336–359, Fal. 2007. [18] R. Choy, A. Edelman, J. R. Gilbert, V. Shah, and D. Cheng, “StarP: High productivity parallel computing,” in Proc. 8th Annu. Workshop on High-Perform. Embedded Comput. (HPEC 04), Lexington, MA, 2004. [19] I. Mirman, “Going parallel the new way,” Desktop Eng., vol. 11, pp. 24– 25, 2006. [20] J. Krupa, A. Pavelka, O. Vˇysata, and A. Proch´azka, “Distributed signal processing,” in Proc. Conf. Technical Computing, Prague, Humusoft, Praha, Czech Republic, 2007. [21] P. Luszczek, “Parallel programming in MATLAB,” Int. J. High Perform. Comput. Appl., vol. 23, pp. 277–283, Aug. 2009. [22] H. Pohlheim, S. Pawletta, and A. Westphal, “Parallel Evolutionary Optimization under Matlab on standard computing networks,” in GECCO’99 - Proc. Genetic Evolutionary Comput. Conf. - Workshop Program, W. Banzhaf Ed. San Francisco, CA: Morgan Kaufmann, 1999, pp. 174–176. [23] S. J. Caton, O. F. Rana, and B. G. Batchelor, “Distributed image processing over an adaptive campus grid,” Concurrency Comput.-Pract. Experience, vol. 21, pp. 321–336, Mar. 10, 2009. [24] MatLab Parallel Computing ToolboxTM 5 User’s Guide, Natick, MA: The MathWorks Inc., 2010. [25] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, and A. Y. Wu, “An efficient k-means clustering algorithm: Analysis and implementation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 881–892, Jul. 2002. [26] T. N. Pappas et al., “An adaptive clustering-algorithm for image segmentation,” IEEE Trans. Signal Process., vol. 40, no. 4, pp. 901–914, Apr. 1992. [27] L. Gao, D. G. Heath, B. S. Kuszyk, and E. K. Fishman, “Automatic liver segmentation technique for three-dimensional visualization of CT data,” Radiology, vol. 201, pp. 359–364, Nov. 1996. [28] S. Lankton and A. Tannenbaum, “Localizing region-based active contours,” IEEE Trans. Image Process., vol. 17, no. 11, pp. 2029–2039, Nov. 2008. [29] T. E. Chan, B. Y. Sandberg, and L. A. Vese, “Active contours without edges for vector-valued images,” J. Visual Commun. Image Represent., vol. 11, pp. 130–141, Jun. 2000. [30] M. Goryawala, M. Guillen, R. Bhatt, A. Mcgoron, and M. Adjouadi, “A comparative study on the performance of the parallel and distributing computing operation in MATLAB,” in Proc. IEEE 24th Int. Conf. Adv. Inform. Networking Appl., Perth, Australia, 2010.

GORYAWALA et al.: 3-D LIVER SEGMENTATION METHOD WITH PARALLEL COMPUTING FOR SIRT

[31] A. Beck and V. Aurich, “HepaTux-a semiautomatic liver segmentation system,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 225–233. [32] R. Beichel, C. Bauer, A. Bornik, E. Sorantin, and H. Bischof, “Liver segmentation in CT data: A segmentation refinement approach,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 235–245. [33] B. M. Dawant, R. Li, B. Lennon, and S. Li, “Semi-automatic segmentation of the liver and its evaluation on the MICCAI grand challenge data set,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 215–221. [34] J. Lee, N. Kim, H. Lee, J. B. Seo, H. J. Won, Y. M. Shin, and Y. G. Shin, “Efficient liver segmentation exploiting level-set speed images with 2.5D shape propagation,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 189–196. [35] L. Massoptier and S. Casciaro, “A new fully automatic and robust algorithm for fast segmentation of liver tissue and tumors from CT scans,” Eur. Radiol., vol. 18, pp. 1658–1665, 2008. [36] R. Susomboon, D. S. Raicu, and J. Furst, “A hybrid approach for liver segmentation,” in Proc. MICCAI Workshop on 3-D Segmentat. Clinic: A Grand Challenge, Brisbane, Australia, 2007, pp. 151–160.

Mohammed Goryawala received the B.S. degree in electronics and communication engineering from South Gujarat University, Surat, India, in 2006, and the M.S. degree in biomedical engineeing from Florida International University, Miami, in 2007. He is currently a Ph.D. student with the Department of Biomedical Engineering, Florida International University, Miami. His primary research interests include developing image and signal processing application for colorectal cancer liver metastases, neuroimaging, neurosignal processing, facial recognition, and high-performance computing.

Magno R. Guillen received the Ph.D. degree in electrical engineering from Florida International University, Miami, in 2008. He is currently a Neuroimaging Data Analyst for Radiology Department, Miami Children Hospital, Miami. His research interests include computer networks, neuroimaging: DTI–FMRI integration, data mining.

Mercedes Cabrerizo received the M.S. degree in computer engineering in 2003 and the Ph.D. degree in electrical engineering from Florida International University (FIU), Miami, in 2006. She is a Ware Foundation Research Scientist with the neuroengineering program (FIU-MCH) with the Department of Electrical and Computer Engineering, FIU. Her research interests include in signal processing and neuroscience.

Armando Barreto received the Ph.D. degree in electrical engineering from the University of Florida, Gainesville, in 1993. He is a Professor in the Department of Electrical and Computer Engineering, Florida International University (FIU), Miami. He is the Director of the FIU Digital Signal Processing Laboratory.

69

Seza Gulec received the M.D. degree from Ankara University School of Medicine, Ankara, Turkey, in 1984, and the F.A.C.S. degree from Louisiana State University, New Orleans, in 2002. He is a Professor of surgery and radiology/nuclear medicine at the Herbert Wertheim College of Medicine, Florida International University, Miami. He is the Chief of Surgical Oncology, and the Director of Research in the Department of Surgery, College of Medicine. His research interests include treatment of thyroid and parathyroid diseases, melanoma, breast cancer, and primary and metastatic liver cancers.

Tushar C. Barot received the MBBS degree from the Gujarat University, Ahmedabad, India, in 2007 and the MPH degree from Florida International University, Miami, in 2011. He is a Research Fellow in the Department of Surgical Oncology, Herbert Wertheim College of Medicine, Florida International University, Miami. His research interest includes radiomicrosphere therapy for colorectal cancer liver metastases.

Rekha R. Suthar received the MBBS degree from the Gujarat University, Ahmedabad, India, in 2007 and the MPH degree from Florida International University, Miami, in 2011. She is a Research Fellow in the Department of Radiology/Nuclear Medicine, Herbert Wertheim College of Medicine, Florida International University, Miami. Her primary research interests include radiomicrosphere therapy for colorectal cancer liver metastases and immunoradio therapy of metastatic pancreatic cancer.

Ruchir N. Bhatt received the B.S. degree in biomedical engineering from the University of Mumbai, India in 2006. He is currently working toward the Ph.D. degree in biomedical engineering at the Florida International University. He is with the Department of Biomedical Engineering, Florida International University, Miami. His research interests include developing algorithms for segmentation of tumors and partial volume effect correction in positron emission tomography images.

Anthony McGoron received the Ph.D. degree in biomedical engineering from Louisiana Tech University, Ruston, LA, in 1991. He is currently an Associate Professor in the Department of Biomedical Engineering, College of Engineering, Florida International University, Miami. His research interest includes image-guided therapy developing new drugs and methods that simultaneously image and treat cancer.

Malek Adjouadi received the M.S. and Ph.D. degrees in electrical engineering from the University of Florida, Gainesville, in 1981 and 1985, respectively. He is a Professor in the Department of Electrical and Computer Engineering,Florida International University, Miami. Since 1993, he has been the Founding Director of the Center for Advanced Technology and Education funded by the National Science Foundation.. His research interests include image and signal processing, assistive technology research, and neuroscience.