Research Article Liver Tumor Segmentation from MR Images...

Research ArticleLiver Tumor Segmentation from MR Images Using3D Fast Marching Algorithm and Single Hidden LayerFeedforward Neural Network

Trong-Ngoc Le12 Pham The Bao3 and Hieu Trung Huynh1

1Faculty of Information Technology Industrial University of Ho Chi Minh City 12 Nguyen Van Bao Go Vap DistrictHo Chi Minh City Vietnam2Faculty of Information Technology University of Science 227 Nguyen Van Cu District 5 Ho Chi Minh City Vietnam3Faculty of Mathematics and Computer Science University of Science 227 Nguyen Van Cu District 5 Ho Chi Minh City Vietnam

Correspondence should be addressed to Hieu Trung Huynh hthieuieeeorg

Received 9 May 2016 Revised 14 July 2016 Accepted 19 July 2016

Academic Editor Yudong Cai

Copyright copy 2016 Trong-Ngoc Le et alThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Objective Our objective is to develop a computerized scheme for liver tumor segmentation in MR imagesMaterials and MethodsOur proposed scheme consists of four main stages Firstly the region of interest (ROI) image which contains the liver tumor regionin the T1-weightedMR image series was extracted by using seed pointsThe noise in this ROI imagewas reduced and the boundarieswere enhanced A 3D fast marching algorithm was applied to generate the initial labeled regions which are considered as teacherregions A single hidden layer feedforward neural network (SLFN) which was trained by a noniterative algorithm was employedto classify the unlabeled voxels Finally the postprocessing stage was applied to extract and refine the liver tumor boundaries Theliver tumors determined by our scheme were compared with those manually traced by a radiologist used as the ldquoground truthrdquoResultsThe study was evaluated on two datasets of 25 tumors from 16 patientsThe proposed scheme obtained themean volumetricoverlap error of 2743 and the mean percentage volume error of 1573 The mean of the average surface distance the root meansquare surface distance and the maximal surface distance were 058mm 120mm and 629mm respectively

1 Introduction

Thehepatic cell carcinoma (HCC) is one of themost commoncancers and rapidly growing worldwide It was estimatedthat up to 90 of patients with liver cancer would be deadwithin five years of diagnosis [1] The early detection andthe treatment response evaluation of patients with the livercancer are very important to improve the survival rateThe basic criterion to evaluate the tumor response to thetreatment is the tumor sizeThere are some criteria for tumorsize including one- bi- or tridimensional measurement [2ndash5] However clinical researches have shown that the volumemeasurement (3D) can give the best reflection of the tumorresponse [3ndash5]The liver tumor volumetry requires the tumorsegmentation Traditionally this task can be performed bymanually tracing the tumor regions on slices It is tediousand time-consuming In addition the volume of manualdelineations is subjective it was estimated that the intra- and

interobserver variability are about 8 for liver tumors [6]Hence it is crucial to investigate in the computerized schemefor liver tumor segmentation

Some computerized schemes have been developed forthe liver tumor segmentation on CT images These methodsinclude the watershed like the paintbrush algorithm [7]deformable models [8] and level-set and active contourtechniques Multiple thresholding and adaptive techniquesfor liver tumor segmentation from CT images also wereproposed by researchers [9ndash11] Wong et al [12] proposed aregion growing technique based on knowledge-constraintsfor liver tumor segmentation in each slice This techniquedemonstrated an average volume difference of 242 intumor segmentation Ben-Dan and Shenhav [13] developedan approach for liver tumor segmentation based on theactive contour using a weighted function of the probabilityof each pixel An approach for liver tumor segmentationbased on Bayesian classification and the active contour

Hindawi Publishing CorporationBioMed Research InternationalVolume 2016 Article ID 3219068 8 pageshttpdxdoiorg10115520163219068

2 BioMed Research International

method was proposed by Taieb et al [14] This approachcan obtain an average volume difference of 442 Smeetset al [15] proposed an approach based on the combinationof the level-set method and the spiral-scanning techniqueIn this method users place a point approximately in themiddle of the liver tumor and specify a maximal radiusby placing other points surrounding the tumor Next thespiral-scanning technique was applied to generate the initialsurface for the level-set segmentation A method for theliver tumor segmentation by applying the minimal surfacesand Markov random fields was developed by Stawiaski et al[16] Some techniques using machine learning methods alsowere developed [17ndash21] Li et al [17] developed a techniquefor liver tumor segmentation based on machine learningusing the intensity profiles of the liver tumors This methodis less accurate for segmenting the tumors with irregularboundaries due to biasing to the blob-like ones Zhou et al[18] developed a scheme in which the tumor region from asingle slice was extracted by using support vector machine(SVM) This region was propagated and applied to classifyvoxels in other slices which contain the tumor A scheme for3D liver tumor segmentation based on the random walkerhas been proposed by Jolly and Grady [19] The user-definedseed point is required and the additional seed points aregenerated from 2D fuzzy-connectedness segmentation of aslice containing the seed point Freiman et al [20] developedan approach for liver tumor segmentation from CTA usingSVM in which the features include the mean the standarddeviation and the minimum and the maximum intensityvalues in a predefined window surrounding the voxel Huanget al [21] demonstrated an approach for liver tumor segmen-tation based on extreme learning machine The features aregenerated from themean variance intensity intensity powerentropy intensity cooccurrence Lawrsquos texture and sum anddifference histogram

In comparison with CT images the number of ap-proaches for liver tumor segmentation on MR images islimited while the increasing use of liverMRI as a single examfor liver disease leads to imperative demands for investigatingresearches in automatic MRI liver tumor volumetry

Developing the computerized scheme for liver tumorsegmentation inMR images is a challenging taskThe imageshave a low gradient response The liver tumors generallyhave different shapes The gray values of tumor dependon several factors including the tumor type the imageacquisition and the contrast injection In this study we focuson a very challenging task of the liver tumor segmentationin abdominal MR images The proposed method utilizesthe local information in the liver it combines the 3D fastmarching algorithm and the neural network The regionof interest (ROI) surrounding the tumor is established toimprove the performance of generating the initial regions aswell as reduce the computational requirements

2 Materials and Methods

Our proposed scheme for liver tumor segmentation isdescribed in Figure 1 It consists of four main stages prepro-cessing generating the initial labeled regions classifying the

Extracting the region of interest (ROI) based on the seed points

Classifying the unlabeled voxels by using the single hidden layer feedforward neural network

(i) Combining the labeled regions from the fast marching algorithm and the neural network(ii) Applying the connected component and relabeling operations(iii) Extracting the tumor region(iv) Applying the morphological operations

Generating the initial labeled regions (teacher regions) by using a fast marching algorithm

3D abdominal MR image

Determining the liver tumor volume

Enhancing liver tumor boundary by a gradient magnitude filter

Reducing noise by using an anisotropic diffusion filter

Figure 1 Overview of the proposed scheme for liver tumorsegmentation

unlabeled voxels using the single hidden layer feedforwardneural network (SLFN) and postprocessing

21 Preprocessing Stage Firstly users can choose the seedpoints inside and outside the tumorThe 3D region of interest(ROI) 1198680

119877(119909 119910 119911) involving the tumor is determined based

on the seed points This ROI is passed to reduce noise andenhance the liver tumor structures by using an anisotropicdiffusion algorithmThis algorithm is handled by a modifiedcurvature diffusion equation given by

120597119868119877

120597119905=

1003816100381610038161003816nabla119868119877

1003816100381610038161003816 nabla sdot 119888 (1003816100381610038161003816nabla119868119877

1003816100381610038161003816)nabla119868119877

1003816100381610038161003816nabla119868119877

1003816100381610038161003816

(1)

where 119868119877(sdot 119905) is the image function with the initial image

at 119905 = 0 given by 119868119877(sdot 119905 = 0) = 119868

0

119877(sdot) and 119888(sdot) is the

diffusion coefficient which controls the sensitivity of the edgecontrast This algorithm smooths noise in the image whilepreserving the major structures including the boundariesThe noise-reduced image is then passed through a Gaussiangradient magnitude filter to generate 119868

119866which enhances the

boundaries The gradient magnitude image is determined by

119868119872

= radic(120597119868119866

120597119909)

2

+ (120597119868119866

120597119910)

2

+ (120597119868119866

120597119911)

2

(2)

BioMed Research International 3

The gradient magnitude image is then employed to producethe edge potential image by using a sigmoid function givenby

119868119875=

1

1 + 119890minus(119868119872minus120573)120572 (3)

where 120572 and 120573 are parameters specifying the range andcenter respectively of the intensity to be enhanced In thisscheme 120572 and 120573 were minus35 and 80 respectively The edgepotential image is used as a speed function for a fastmarchingalgorithm to generate the initial labeled regions

22 Generating the Initial Labeled Regions The main goal inthis stage is to generate the labeled regions or teacher regionsby using a fast marching algorithm This algorithm is basedon the numerical solution of the Eikonal equation given by

|nabla119879| 119865 = 1 (4)

where 119879 is an arrival time function and 119865 is a speed functionwhich is obtained from the potential image 119868

119875The algorithm

requires the initial seed points which correspond to the initiallocation (119879 = 0) These seed points (1-2 points inside thetumor and 1ndash3 points outside the tumor) in our scheme areplaced by a radiologist In the first step of the algorithmthe grid points from the entire image are categorized intothree classesThe initial seed points are classed asKnownTheneighbors of the Known points are classed as Trial and theirarrival time is computed by using the first-order scheme (see(4)) Other points are classified as Far with the infinity arrivaltime The algorithm propagates the information in one wayfrom the smaller values to the larger values of 119879 The point qwith the smallest arrival time in the Trial list is chosen andmoved to the Known one The neighbors of q are moved tothe Trial list and the arrival time is recomputed by using thefirst-order scheme This iterative process is terminated whenthe maximum number of iterations is met The salient pointof this algorithm is that the heap data structure is used tospeed up locating the Trial point with the smallest 119879 valueThe output of the fast marching algorithm is a time-crossingmap image which indicates the time travelling to each pointThe labeled regions are generated by

119868119878(119909 119910 119911) =

2 if 119868TM (119909 119910 119911) le 119872119865and the voxel belongs to regions containing a seed point outside tumor

1 if 119868TM (119909 119910 119911) le 119872119865and the voxel belongs to regions containing a seed point inside tumor

0 otherwise

(5)

where 119872119865is the maximum number of iterations for the

fast marching algorithm and 119868TM is the time-crossing mapimage The points with the values of 1 or 2 in the image119868119878are considered to be labeled otherwise considered to

be unlabeled because they were not treated from the fastmarching algorithm Due to variations of intensity in tumorsthe labeled regions could not form the rough shape of the livertumor However they can be considered as teacher regionsfor training the neural network to classify the unlabeledpoints

23 Classifying the Unlabeled Voxels The labeled points arecategorized into two classes tumor or nontumor corre-sponding to the labels of 1 or 2 respectively Now we haveto classify the unlabeled points which are not treated bythe fast marching algorithm The neural network is oneof the powerful tools in biology and medical data analysisapplications [22 23] Hence in this study we utilized theneural network to categorize the unlabeled points

Several network architectures have been developed how-ever a single hidden layer feedforward neural network(SLFN) can form the decision boundaries with arbitraryshapes if the activation function is chosen properly A typicalarchitecture of SLFN consists of an input layer a hidden layerwith119870 units and an output layer with119872 units In this studythe SLFN handles the input gray levels directly The inputs tothe SLFN comprise the normalized voxel in ROI image (119868

119877)

and its spatially adjacent normalized voxels within the local

window 119871119882 An input pattern corresponding to 119868

119877(119909 119910 119911) is

determined by

p119909119910119911

= minus1 +2 (119868119877(119909 minus 119894 119910 minus 119895 119911 minus 119896) minus 119868min)

(119868max minus 119868min)| 119894 119895 119896

isin 119871119882

(6)

where 119868min 119868max are the minimum and maximum voxelvalues respectively in 119868

119877 The output of the SLFN which

corresponds to the center voxel in the local window is thevalue indicating either tumor or nontumor represented by

o119909119910119911

=

119870

sum

119896=1

a119896120593 (p119909119910119911

sdot w119896+ 119887119896) (7)

wherew119896is the input weights connecting from the input layer

to the 119896th hidden unit 119887119896is its bias a

119896= [1198861198961 1198861198962 119886

119896119872]119879

is the output weights connecting from the 119896th hidden unitto the output layer and 120593(sdot) is the activation function of thehidden layer which was a sigmoidal function in this studyThe activation function of the output layer was an identifyfunction

One of the important issues in the SLFN is to train thenetwork to determine the network weights w 119887 and a Tra-ditionally training the neural network could be performedby employing the backpropagation algorithmThis algorithm


has some limitations such as slow convergence local minimaoverfitting or improper learning rate Although there areseveral improvements to overcome the problems of back-propagation algorithm up to now the training algorithmsbased on the gradient descent approach are still slow due tomany iterative steps In this study we employed an effectivetraining algorithm for SLFN called extreme learningmachine(ELM) and its improvements [24ndash27]

The desired output corresponding to the input patternp119909119910119911

is obtained from the teacher regions in 119868119878 represented

by

t119909119910119911

=

(0 1)119879

if 119868119878(119909 119910 119911) = 1

(1 0)119879

if 119868119878(119909 119910 119911) = 2

(8)

Assume that there are 119899 labeled points in the teacher regionsThe training set is given by 119878 = (p

119894 t119894) 119894 = 1 2 119899 where

the input patterns p119894are obtained from 119868

119877by using (6) and its

corresponding desired output t119894 is obtained from the teacher

regions in 119868119878by using (8)The error function to beminimized

by training process is defined by

119864 =

119899

sum

119894=1

1003817100381710038171003817o119894 minus t119894

1003817100381710038171003817 (9)

where o119894is the actual output corresponding to the input

pattern p119894(by using (7)) In the extreme learning machine

the training process can be modeled by finding the solutionof a linear model given by HA = T where H is the hiddenlayer output matrix of SLFN defined by

H =

[[[[

[

120593 (w1sdot p1+ 1198871) sdot sdot sdot 120593 (w

119870sdot p1+ 119887119870)

d


119870sdot p119899+ 119887119870)

]]]]

]

T = [t1 t2

sdot sdot sdot t119899]119879

A = [a1 a2

sdot sdot sdot a119870]119879

(10)

In the extreme learningmachine the biases and inputweightsare assigned by the random values and the output weights arefound by using the Moore-Penrose generalized inverse

1006704A = HdaggerT (11)

whereHdagger is the pseudoinverse of H Several improvements ofELM have been developed by researchers [26 27] In order tomake the system more stable the regularization approacheshave been proposed in which a coefficient 120582 is added and thesolution of A is given by

1006704A = (H119879H + 120582I)minus1

H119879T (12)

These training algorithms are fast and can offer a goodperformance in many applications including medical dataanalysis

24 Postprocessing The classifying results from the SLFN forunlabeled regions are combined with the ones from the fastmarching algorithm A voxel belongs to tumor region if itis classified as tumor by either the fast marching algorithmor the SLFN The combined results are then passed to theconnected component and the relabeling filters The regioncontaining the tumor seed points is extracted and passedto remove small isolated artifacts and small holes by usingthe morphological operations The liver tumor volume iscalculated from the segmented tumor regions

3 Experimental Results

31 Datasets This study was approved by the institutionalreview board (IRB) of theMedicMedical CenterThe datasetsconsist of 25 liver tumors which were obtained from 16patients Fifteen tumors were obtained from 10 patients byusing the 15 T magnetic resonance imaging (MRI) scanners(Avanto Siemens) at the Medic Medical Center which isone of the largest diagnostic imaging centers in VietnamInformed consent was obtained from all patients Postcon-trast MR images were obtained by using the T1-weightvolumetric interpolated breath-hold examination (VIBE)sequence A flip angle of 10 degrees was used with the TE andTR of 238 and 474 respectively The scanning parametersincluded collimation and reconstruction intervals rangingfrom 35 to 4mm Each MRI slice had a matrix size of 230times 320 pixels with an in-plane pixel size ranging from 118to 14mm The number of slices in each case ranged from44 to 56 Ten other tumors were obtained from 6 patientcaseswhichwere extracted fromTheCancer ImagingArchive(TCIA) [28]

A board-certified abdominal radiologist carefully manu-ally traced the tumor contours on each slice which containsthe liver tumor The liver tumor volume was calculated bymultiplying the areas of the manually traced regions in eachslice by the reconstruction interval The total tumor livervolume in each case was determined by the summation of thevolumes in all of the slicesThe times required to complete themanual contour tracing were recorded

32 EvaluationCriteria The liver tumor volumes obtained byour computerized approach were compared to the ldquoground-truthrdquo manual volumes which were determined by the radi-ologist The true positive (TP) false positive (FP) and falsenegative were calculated for detail analysis The percentagevolume error (119864) for each computerized volume (119881

119888) and the

gold standard manual volume (119881119898) is determined by

119864 =

100381610038161003816100381610038161003816100381610038161003816

(119881119888minus 119881119898)

119881119898

100381610038161003816100381610038161003816100381610038161003816

(13)

The volumetric overlap error representing the fraction of theoverlapping volume and the volume of two segmentationmethods is given by

VO = 1 minus|TP|

|TP| + |FP| + |FN| (14)


Tumor to be extracted

(a) (b) (c)

(d) (e) (f) (g)

Figure 2 The intermediate results of the proposed scheme (a) A slice of the original 3D image (b) A slice of the 3D region of interestcontaining the liver tumor which was extracted from the original 3D MR image (c) The edge image generated by applying the gradientmagnitude filter (d)The labeled regions generated by the fast marching algorithm and thresholding filter (e) Unlabeled voxels were classifiedby using the SLFN (f) The segmented liver tumor (g) A comparison between the computerized liver tumor segmentation (black contour)and the ldquoground-truthrdquo manual liver tumor segmentation (white contour)

Three surface distancemetrics including the average symmet-ric surface distance (ASSD) the root mean square symmetricsurface distance (RMSSD) and the maximal symmetricsurface distance (MSSD) also were used to evaluate theglobal and the local disagreement between two segmentationmethods

33 Results and Discussions The intermediate results of theproposed scheme for an example casewere shown in Figure 2An original 3D MR image was shown in Figure 2(a) The 3Dregion of interest (ROI)whichwas extracted from the original3DMR image was shown in Figure 2(b) This ROI image waspassed to reduce the noise by using an anisotropic diffusionalgorithm A gradient magnitude filter was applied to thereduced-noise image to enhance the boundaries and create anedge image as shown in Figure 2(c)The edge potential imagewas generated and then the fast marching algorithm andthresholding filter were applied to generate the imagewith thelabeled regions as shown in Figure 2(d) The labeled regionswere employed to train the SLFN this trained network wasthen used to classify voxels in the unlabeled regions asshown in Figure 2(e)The classification result from SLFNwascombined with the one from the fast marching algorithmThe connected component and relabeling operations wereapplied and the region which contains the seed points insidetumor was filtered out as shown in Figure 2(f) A compar-ison between the computerized liver tumor segmentation

Table 1 Comparison of liver tumor volume between the computer-ized method and the manual method

Dataset Volume Manual method(cc)

Computerizedmethod (cc)

Medic Medical Center

Average 348 316SD 400 397Min 023 014Max 1610 1618

TCIA


Both


(black contour) and the ldquoground-truthrdquo manual liver tumorsegmentation (white contour) was illustrated in Figure 2(g)The liver tumor volume was computed from the segmentedregions

The average liver tumor volumes measured from twomethods for two datasets Medic Medical Center and TCIAwere shown in Table 1 For both datasets the average tumorvolume of the reference standardmanual method was 1536plusmn3677 cm3 (range 019ndash16289 cm3) whereas the average


Table 2 Summary of the comparison results

Dataset Evaluation measure Mean SD Min Max


Volumetric overlap error () 2666 706 1570 3947Percentage volume error () 1668 1251 017 3947Average surface distance (mm) 044 047 021 212RMS surface distance (mm) 097 097 053 440

Maximal surface distance (mm) 484 489 168 2118

TCIA



Both



tumor volume of the proposed computerized scheme was1388 plusmn 3321 cm3 (range 010ndash14795 cm3) A comparisonof the liver tumor segmentation on MR image between twomethods was shown in Table 2 For both datasets the overallmean of volumetric overlap error and the mean percentagevolume error were 2743 and 1574 respectivelyThemeanof average surface distance root mean square surface dis-tance and maximal surface distance were 058mm 120mmand 629mm respectively

It is not easy to directly compare the proposed methodwith existing methods in literature because of using differentdatabases and quality measurements Freiman et al [20]evaluated the performance of their proposed scheme on CTAimages and it could obtain a volume overlap error anda volume difference of 338 and 226 respectively Themean of average surface distance the RMS surface distanceand the maximal surface distance were 176mm 262mmand 1373mm respectively Smeets et al [15] obtained theoverlap error and the volume difference of 3264 and 1791respectivelyThe average the RMS and themaximum surfacedistance were 197mm 264mm and 1013mm respectivelyThe method proposed by Huang et al [21] could achievethe volume overlap error the volume difference the averagesurface distance the RMS surface distance and the maximalsurface distance of 6715 1416 227mm 247mm and846mm respectively the performance of this scheme wasevaluated on CT images

There are several parameters to be adjusted in our pro-posed scheme They were determined by empirical analysisThe number of hidden nodes in the SLFN was given by 119870 =

200 and120582 = 001 the local window size for extracting features(119871119882) was set to 3 times 3 times 3 The number of iterations for the

fastmarching algorithmwas set to 15 All the parameters werefixed for all patients

Some advanced schemes have applied the statistical mod-els used the intensity profiles of liver tumors or used filtersfor the tumor segmentation However the actual tumors

often differ from the simple models There are tumors ofvarious shapes and inhomogeneous tumors In this studythe artificial neural network (ANN) was developed foraccommodating the task of distinguishing the tumor voxelsfrom the nontumor voxels The network was trained by thesubregions of the 3DMR image andoperates on the voxel datadirectly which is similar to the MTANN (massive trainingartificial neural network) [29] However unlike the originalMTANN the ANN proposed in this study was trained bythe data from the 3D local information of each tumor andthe network was trained by a noniterative training algorithmwhich can overcome problems of the traditional trainingalgorithms such as local minima learning rate and epochsThis training algorithm can offer a good performance withhigh learning speed in many different applications

A limitation in this study is the number of patients Ourproposed scheme was evaluated in 16 patients with 25 tumorswhile other studies evaluated 5 patients with 10 tumors [20]7 patients with 10 tumors [15] and 10 patients with 10 lesions[30] In general a small number of tumors and patientsmay limit the variations among them In the future we willneed to increase the number of patients and tumors used forevaluation

4 Conclusion

The increasing use of liver MRI as a single exam leadsto imperative demands for investigating researches in thecomputerized MRI liver tumor segmentation However fewstudies have been reported for this challenging task Inthis study we developed a computerized scheme for livertumor segmentation on MR images by employing the fastmarching algorithm and the neural network in which theneural network was trained by an effectively noniterativealgorithm The performance was evaluated in 25 tumors of16 patients Our experimental results have shown that theproposed method is accurate and efficient when compared


to the manual ground-truth segmentation Its accuracy alsois comparable or better than the existing semiautomaticmethods With our proposed scheme the time required forliver tumor segmentation is reduced significantly Hence itcan be useful for radiologists for the liver tumor analysis onMR images

Ethical Approval

All procedures performed in studies involving human partic-ipants were in accordance with the ethical standards of theinstitutional andor national research committee and withthe 1964 Helsinki declaration and its later amendments orcomparable ethical standards This paper does not containany studies with animals performed by any of the authors

Consent

Informed consent was obtained from all individual partici-pants included in the study

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

This research is funded by the Vietnam National Foundationfor Science and Technology Development (NAFOSTED)under Grant no 10201-201347 The authors are grateful toPhan Thanh Hai MD and Nguyen Thanh Dang MD forpreparing images and for their clinical advice

References

[1] Primary liver cancer Celsion corporation 2016 httpcelsioncomdocspipeline liver cancer

[2] P Therasse S G Arbuck E A Eisenhauer et al ldquoNew guide-lines to evaluate the response to treatment in solid tumorsrdquoJournal of the National Cancer Institute vol 92 no 3 pp 205ndash216 2000

[3] L N Tran M S Brown J G Goldin et al ldquoComparisonof treatment response classifications between unidimensionalbidimensional and volumetric measurements of metastaticlung lesions on chest computed tomographyrdquo Academic Radi-ology vol 11 no 12 pp 1355ndash1360 2004

[4] R S Tuma ldquoSometimes size doesnrsquot matter reevaluatingRECIST and tumor response rate endpointsrdquo Journal of theNational Cancer Institute vol 98 no 18 pp 1272ndash1274 2006

[5] S R Prasad K S Jhaveri S Saini P F Hahn E F Halpern andJ E Sumner ldquoCT tumor measurement for therapeutic responseassessment comparison of unidimensional bidimensional andvolumetric techniquesmdashInitial observationsrdquo Radiology vol225 no 2 pp 416ndash419 2002

[6] E Bellon M Feron F Maes et al ldquoEvaluation of manual vssemi-automated delineation of liver lesions on CT imagesrdquoEuropean Radiology vol 7 no 3 pp 432ndash438 1997

[7] F Maes D Vandermeulen P Suetens and G Marchal ldquoAuto-matic image partitioning for generic object segmentation in

medical imagesrdquo Computational Imaging and Vision vol 3 pp215ndash226 1995

[8] P J Yim and D J Foran ldquoVolumetry of hepatic metastases incomputed tomography using the watershed and active contouralgorithmsrdquo in Proceedings of the 16th IEEE Symposium onComputer Based Medical Systems pp 329ndash335 New York NYUSA June 2003

[9] J Moltz L Bornemann V Dicken and H Peitgen ldquoSegmenta-tion of liver metastases in CT scans by adaptive thresholdingand morphological processingrdquo The MIDAS JournalmdashGrandChallenge Liver Tumor Segmentation pp 1ndash10 2008 Proceed-ings of the 2008 MICCAI Workshop

[10] N H Abdel-Massieh M M Hadhoud and K M Amin ldquoFullyautomatic liver tumor segmentation from abdominal CT scansrdquoin Proceedings of the 6th International Conference on ComputerEngineering and Systems (ICCES rsquo10) pp 197ndash202 IEEE CairoEgypt December 2010

[11] A Choudhary N Moretto F P Ferrarese and G A Zam-boni ldquoAn entropy based multi-thresholding method for semi-automatic segmentation of liver tumorsrdquo The MIDAS Journal-Grand Challenge Liver Tumor Segmentation (2008 MICCAIWorkshop) vol 41 no 43 pp 43ndash49 2008

[12] D Wong J Liu F Yin et al ldquoA semi-automated methodfor liver tumor segmentation based on 2D region growingwith knowledge-based constraintsrdquoTheMIDAS Journal-GrandChallenge Liver Tumor Segmentation (2008MICCAIWorkshop)vol 41 no 43 p 159 2008

[13] I Ben-Dan and E Shenhav ldquoLiver tumor segmentation inCT images using probabilisticrdquo The MIDAS JournalmdashGrandChallenge Liver Tumor Segmentation pp 1ndash6 2008 Proceedingsof the MICCAI Workshop 2008

[14] Y Taieb O Eliassaf M Freiman L Joskowicz and J Sosna ldquoAniterative Bayesian approach for liver analysis tumor validationstudyrdquo The MIDAS JournalmdashGrand Challenge Liver TumorSegmentation (2008 MICCAI Workshop) pp 1ndash6 2008

[15] D Smeets D Loeckx B Stijnen B De Dobbelaer D Vander-meulen and P Suetens ldquoSemi-automatic level set segmentationof liver tumors combining a spiral-scanning technique withsupervised fuzzy pixel classificationrdquo Medical Image Analysisvol 14 no 1 pp 13ndash20 2010

[16] J Stawiaski E Decenciere and F Bidault ldquoInteractive livertumor segmentation using graph-cuts and watershedrdquo TheMIDAS JournalmdashGrand Challenge Liver Tumor Segmentation(MICCAI Workshop) pp 1ndash6 2008

[17] Y Li S Hara and K Shimura ldquoA machine learning approachfor locating boundaries of liver tumors in CT imagesrdquo inProceedings of the 18th International Conference on PatternRecognition (ICPR rsquo06) pp 400ndash403 Hong Kong August 2006

[18] J ZhouW Xiong Q Tian et al ldquoSemi-automatic segmentationof 3D liver tumors from CT scans using voxel classification andpropagational learningrdquoTheMIDAS JournalmdashGrand ChallengeLiver Tumor Segmentation (MICCAIWorkshop) pp 1ndash10 2008

[19] M-P Jolly and L Grady ldquo3D general lesion segmentation inCTrdquo in Proceedings of the 5th IEEE International Symposium onBiomedical Imaging From Nano to Macro (ISBI rsquo08) pp 796ndash799 IEEE Paris France May 2008

[20] M Freiman O Cooper D Lischinski and L JoskowiczldquoLiver tumors segmentation from CTA images using voxelsclassification and affinity constraint propagationrdquo InternationalJournal of Computer Assisted Radiology and Surgery vol 6 no2 pp 247ndash255 2010


[21] W Huang N Li Z Lin et al ldquoLiver tumor detection andsegmentation using kernel-based extreme learningmachinerdquo inProceedings of the 35th Annual International Conference of theIEEE Engineering in Medicine and Biology Society (EMBC rsquo13)pp 3662ndash3665 Osaka Japan July 2013

[22] H T Huynh J-J Kim and Y Won ldquoPerformance comparisonof SLFN training algorithms forDNAmicroarray classificationrdquoAdvances in Experimental Medicine and Biology vol 696 pp135ndash143 2011

[23] H T Huynh Y Won and J Kim ldquoHematocrit estimationusing online sequential extreme learningmachinerdquoBio-MedicalMaterials and Engineering vol 26 supplement 1 pp S2025ndashS2032 2015

[24] G-B Huang Q-Y Zhu and C-K Siew ldquoExtreme learningmachine theory and applicationsrdquoNeurocomputing vol 70 no1ndash3 pp 489ndash501 2006

[25] N-Y Liang G-B Huang P Saratchandran and N Sundarara-jan ldquoA fast and accurate online sequential learning algorithmfor feedforward networksrdquo IEEE Transactions on Neural Net-works vol 17 no 6 pp 1411ndash1423 2006

[26] H T Huynh Y Won and J-J Kim ldquoAn improvement ofextreme learning machine for compact single-hidden-layerfeedforward neural networksrdquo International Journal of NeuralSystems vol 18 no 5 pp 433ndash441 2008

[27] H T Huynh and Y Won ldquoSmall number of hidden unitsfor ELM with two-stage linear modelrdquo IEICE Transactions onInformation and Systems vol E91-D no 4 pp 1042ndash1049 2008

[28] National Cancer Institute ldquoCancer Imaging Program FrederickNational Laboratory for Cancer Research the Cancer ImagingArchive (TCIA)rdquo 2013 httpcancerimagingarchivenet

[29] K Suzuki and K Doi ldquoHow can a massive training artificialneural network (MTANN) be trained with a small number ofcases in the distinction between nodules and vessels in thoracicCTrdquo Academic Radiology vol 12 no 10 pp 1333ndash1341 2005

[30] S S Kumar R S Moni and J Rajeesh ldquoAutomatic liverand lesion segmentation a primary step in diagnosis of liverdiseasesrdquo Signal Image and Video Processing vol 7 no 1 pp163ndash172 2013

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


method was proposed by Taieb et al [14] This approachcan obtain an average volume difference of 442 Smeetset al [15] proposed an approach based on the combinationof the level-set method and the spiral-scanning techniqueIn this method users place a point approximately in themiddle of the liver tumor and specify a maximal radiusby placing other points surrounding the tumor Next thespiral-scanning technique was applied to generate the initialsurface for the level-set segmentation A method for theliver tumor segmentation by applying the minimal surfacesand Markov random fields was developed by Stawiaski et al[16] Some techniques using machine learning methods alsowere developed [17ndash21] Li et al [17] developed a techniquefor liver tumor segmentation based on machine learningusing the intensity profiles of the liver tumors This methodis less accurate for segmenting the tumors with irregularboundaries due to biasing to the blob-like ones Zhou et al[18] developed a scheme in which the tumor region from asingle slice was extracted by using support vector machine(SVM) This region was propagated and applied to classifyvoxels in other slices which contain the tumor A scheme for3D liver tumor segmentation based on the random walkerhas been proposed by Jolly and Grady [19] The user-definedseed point is required and the additional seed points aregenerated from 2D fuzzy-connectedness segmentation of aslice containing the seed point Freiman et al [20] developedan approach for liver tumor segmentation from CTA usingSVM in which the features include the mean the standarddeviation and the minimum and the maximum intensityvalues in a predefined window surrounding the voxel Huanget al [21] demonstrated an approach for liver tumor segmen-tation based on extreme learning machine The features aregenerated from themean variance intensity intensity powerentropy intensity cooccurrence Lawrsquos texture and sum anddifference histogram

In comparison with CT images the number of ap-proaches for liver tumor segmentation on MR images islimited while the increasing use of liverMRI as a single examfor liver disease leads to imperative demands for investigatingresearches in automatic MRI liver tumor volumetry

Developing the computerized scheme for liver tumorsegmentation inMR images is a challenging taskThe imageshave a low gradient response The liver tumors generallyhave different shapes The gray values of tumor dependon several factors including the tumor type the imageacquisition and the contrast injection In this study we focuson a very challenging task of the liver tumor segmentationin abdominal MR images The proposed method utilizesthe local information in the liver it combines the 3D fastmarching algorithm and the neural network The regionof interest (ROI) surrounding the tumor is established toimprove the performance of generating the initial regions aswell as reduce the computational requirements

2 Materials and Methods

Our proposed scheme for liver tumor segmentation isdescribed in Figure 1 It consists of four main stages prepro-cessing generating the initial labeled regions classifying the

Extracting the region of interest (ROI) based on the seed points

Classifying the unlabeled voxels by using the single hidden layer feedforward neural network

(i) Combining the labeled regions from the fast marching algorithm and the neural network(ii) Applying the connected component and relabeling operations(iii) Extracting the tumor region(iv) Applying the morphological operations

Generating the initial labeled regions (teacher regions) by using a fast marching algorithm

3D abdominal MR image

Determining the liver tumor volume

Enhancing liver tumor boundary by a gradient magnitude filter

Reducing noise by using an anisotropic diffusion filter

Figure 1 Overview of the proposed scheme for liver tumorsegmentation

unlabeled voxels using the single hidden layer feedforwardneural network (SLFN) and postprocessing

21 Preprocessing Stage Firstly users can choose the seedpoints inside and outside the tumorThe 3D region of interest(ROI) 1198680

119877(119909 119910 119911) involving the tumor is determined based

on the seed points This ROI is passed to reduce noise andenhance the liver tumor structures by using an anisotropicdiffusion algorithmThis algorithm is handled by a modifiedcurvature diffusion equation given by

120597119868119877

120597119905=

1003816100381610038161003816nabla119868119877

1003816100381610038161003816 nabla sdot 119888 (1003816100381610038161003816nabla119868119877

1003816100381610038161003816)nabla119868119877

1003816100381610038161003816nabla119868119877

1003816100381610038161003816

(1)

where 119868119877(sdot 119905) is the image function with the initial image

at 119905 = 0 given by 119868119877(sdot 119905 = 0) = 119868

0

119877(sdot) and 119888(sdot) is the

diffusion coefficient which controls the sensitivity of the edgecontrast This algorithm smooths noise in the image whilepreserving the major structures including the boundariesThe noise-reduced image is then passed through a Gaussiangradient magnitude filter to generate 119868

119866which enhances the

boundaries The gradient magnitude image is determined by

119868119872

= radic(120597119868119866

120597119909)

2

+ (120597119868119866

120597119910)

2

+ (120597119868119866

120597119911)

2

(2)



119868119875=

1

1 + 119890minus(119868119872minus120573)120572 (3)



|nabla119879| 119865 = 1 (4)


119875The algorithm


119868119878(119909 119910 119911) =



0 otherwise

(5)






119877)



119877(119909 119910 119911) is

determined by

p119909119910119911


(119868max minus 119868min)| 119894 119895 119896

isin 119871119882

(6)




o119909119910119911

=

119870

sum

119896=1

a119896120593 (p119909119910119911

sdot w119896+ 119887119896) (7)



119896= [1198861198961 1198861198962 119886

119896119872]119879







by

t119909119910119911

=

(0 1)119879

if 119868119878(119909 119910 119911) = 1

(1 0)119879

if 119868119878(119909 119910 119911) = 2

(8)


119894 t119894) 119894 = 1 2 119899 where






119864 =

119899

sum

119894=1

1003817100381710038171003817o119894 minus t119894

1003817100381710038171003817 (9)




H =

[[[[

[


119870sdot p1+ 119887119870)

d


119870sdot p119899+ 119887119870)

]]]]

]

T = [t1 t2


A = [a1 a2


(10)


1006704A = HdaggerT (11)


1006704A = (H119879H + 120582I)minus1

H119879T (12)







119888) and the


119864 =

100381610038161003816100381610038161003816100381610038161003816

(119881119888minus 119881119898)

119881119898

100381610038161003816100381610038161003816100381610038161003816

(13)


VO = 1 minus|TP|

|TP| + |FP| + |FN| (14)



(a) (b) (c)

(d) (e) (f) (g)









TCIA


Both










TCIA



Both











4 Conclusion




Ethical Approval


Consent


Competing Interests


Acknowledgments


References






































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















119868119875=

1

1 + 119890minus(119868119872minus120573)120572 (3)



|nabla119879| 119865 = 1 (4)


119875The algorithm


119868119878(119909 119910 119911) =



0 otherwise

(5)






119877)



119877(119909 119910 119911) is

determined by

p119909119910119911


(119868max minus 119868min)| 119894 119895 119896

isin 119871119882

(6)




o119909119910119911

=

119870

sum

119896=1

a119896120593 (p119909119910119911

sdot w119896+ 119887119896) (7)



119896= [1198861198961 1198861198962 119886

119896119872]119879







by

t119909119910119911

=

(0 1)119879

if 119868119878(119909 119910 119911) = 1

(1 0)119879

if 119868119878(119909 119910 119911) = 2

(8)


119894 t119894) 119894 = 1 2 119899 where






119864 =

119899

sum

119894=1

1003817100381710038171003817o119894 minus t119894

1003817100381710038171003817 (9)




H =

[[[[

[


119870sdot p1+ 119887119870)

d


119870sdot p119899+ 119887119870)

]]]]

]

T = [t1 t2


A = [a1 a2


(10)


1006704A = HdaggerT (11)


1006704A = (H119879H + 120582I)minus1

H119879T (12)







119888) and the


119864 =

100381610038161003816100381610038161003816100381610038161003816

(119881119888minus 119881119898)

119881119898

100381610038161003816100381610038161003816100381610038161003816

(13)


VO = 1 minus|TP|

|TP| + |FP| + |FN| (14)



(a) (b) (c)

(d) (e) (f) (g)









TCIA


Both










TCIA



Both











4 Conclusion




Ethical Approval


Consent


Competing Interests


Acknowledgments


References






































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















by

t119909119910119911

=

(0 1)119879

if 119868119878(119909 119910 119911) = 1

(1 0)119879

if 119868119878(119909 119910 119911) = 2

(8)


119894 t119894) 119894 = 1 2 119899 where






119864 =

119899

sum

119894=1

1003817100381710038171003817o119894 minus t119894

1003817100381710038171003817 (9)




H =

[[[[

[


119870sdot p1+ 119887119870)

d


119870sdot p119899+ 119887119870)

]]]]

]

T = [t1 t2


A = [a1 a2


(10)


1006704A = HdaggerT (11)


1006704A = (H119879H + 120582I)minus1

H119879T (12)







119888) and the


119864 =

100381610038161003816100381610038161003816100381610038161003816

(119881119888minus 119881119898)

119881119898

100381610038161003816100381610038161003816100381610038161003816

(13)


VO = 1 minus|TP|

|TP| + |FP| + |FN| (14)



(a) (b) (c)

(d) (e) (f) (g)









TCIA


Both










TCIA



Both











4 Conclusion




Ethical Approval


Consent


Competing Interests


Acknowledgments


References






































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















(a) (b) (c)

(d) (e) (f) (g)









TCIA


Both










TCIA



Both











4 Conclusion




Ethical Approval


Consent


Competing Interests


Acknowledgments


References






































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of






















TCIA



Both











4 Conclusion




Ethical Approval


Consent


Competing Interests


Acknowledgments


References






































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















Ethical Approval


Consent


Competing Interests


Acknowledgments


References






































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















Research Article Liver Tumor Segmentation from MR Images...

Documents

Transcript of Research Article Liver Tumor Segmentation from MR Images...