MS Lesions Detection in MRI Using Grouping Artificial Immune Networks

Akmal A. Younis, Ahmed T. Soliman, Mansur R. Kabuka, and Nigel M. JohnDepartment ofElectrical and Computer Engineering, University ofMiami, Miami, FL, USAayounisgmiami.edu, a.solimangumiami.edu, m.kabukagmiami.edu, nigel.johngmiami.edu

Abstract

A dual-channel 3D MRI segmentation techniquebased on Grouping Artificial Immune Networks(GAIN) is introduced to detect MS lesion in MRimages. The technique demonstrates the ability ofartificial immune networks to handle MS lesionsdetection in Ti- and T2-weighted brain MRI TheGAIN-based MRI segmentation technique wasevaluated using simulated MS brain images from theMcConnell Brain Imaging Centre, MontrealNeurological Institute of McGill University. 3Danisotropic filtering is used to handle noise artifacts inthe simulated 3D MRI data sets. Experimental resultsdemonstrated that dual channel MS segmentationapproach exhibited high accuracy in segmenting thesimulated MS brain data and an even higher accuracywhen compared to techniques based on single channel3D MRI data sets in terms of the Dice coefficient, anobjective measure ofoverlap.

1. Introduction

Analysis of Magnetic Resonance (MR) images hasgained increased attention due to their wide spectrumof clinical applications including those addressing thestudy of normal brain structure and functions [1][2],the study of the development and aging of the brain[3]-[5], as well as other applications focusing on thedetection and quantification of abnormal brain lesionssuch as malignant tumors and multiple sclerosis [6]-[10].

MR images have demonstrated their efficacy inthe detection of Multiple Sclerosis (MS) lesions, whichaffect the white matter regions of the brain [11]-[15].Different MR channels, which differ in terms of imageacquisition parameters, can be used to detect andquantify MS lesions due to the variation of theirappearance. For example, MS lesions have lowerintensity in the Ti -weighted MR images than normalwhite matter tissues while Gadolinium enhanced T1-weighted MR images provide higher contrast for the

MS lesions. On the other hand, T2-weighted andProton Density (PD) MR images show MS lesions witha higher intensity than normal white matter regions.However, the issue in T2-weighted and PD- images isthe difficulty in discriminating Cerebro-Spinal Fluid(CSF) of the brain from MS lesions due to theircomparable intensities. Other channels of MRI such asT2 Fluid-Attenuated Inversion-Recovery (T2-FLAIR),Magnetic Transfer Ratio (MTR), and Single EmittedPhoton Computed Tomography (SPECT) have alsobeen used to quantify and detect MS lesions [6] [16].

In order for the analysis of MR images to beeffectively utilized in investigating the progression ofMS lesions, especially when monitoring the effects oftreatment, the segmentation of MS lesions from MRIshould be accurate and robust. Both requirements areespecially important in clinical trials involving the useof MRI studies to detect MS lesions, where a verylarge number of studies are analyzed and segmented,e.g., during phase III clinical trials a large number ofpatients, typically between 1000 and 3000 are studied.The accuracy of the segmentation is important toassess the progression of the MS lesions. As thenumber of MRI studies to be segmented becomeslarge, manual segmentation, which is a time-consuming process, is rendered non-viable. In addition,the subjective evaluation of the location and extent ofthe MS lesions is highly dependent on the specialistsegmenting data. Based on these different factorspertaining to MS lesion segmentation, there is a needfor an automated technique to detect and segment MSlesions in MR images that should be accurate, fast androbust to meet the requirements of clinicalenvironments, especially during clinical trials.

Several approaches have been proposed tosegment MS lesions in MR images. Most of theapproaches use a previously established model for anornal brain with spatial prior probability for eachvoxel with respect to normal brain tissues, i.e., whitematter (WM), gray matter (GM) and Cerebro-SpinalFluid (CSF). Based on the established models,techniques were introduced for MS lesionsegmentation based on different algorithms such as

1-4244-1509-8/07/$25.00 02007 IEEE 1139

artificial neural networks [8], Markov random fields[7], Gaussian mixture models [9], image analysisapproaches [16], and fuzzy connectedness [6][9][17].Generally, MS lesion segmentation techniques can bedivided into two categories: algorithms that aim tomodel the MS lesions and algorithms that detect theMS lesions as outliers with respect to a model for thenormal brain [7]. The first category involves themodeling ofMS lesions through the acquisition of theircharacteristic parameters in MR images that are laterutilized in the detection of regions with similarcharacteristics. Meanwhile, the second categoryinvolves segmenting the voxels based on the normalbrain model and identifying MS lesions as those whichcan be identified as belonging to none of the normaltissues or as belonging to more than one tissue, i.e.,outlier regions. Recently, artificial immune networkshave been widely used for different applications suchas data clustering [18], optimization problems [19] andpattern recognition [20]. In [21], a grouping artificialimmune network system was used to successfullysegment normal brain MR images. GAIN-basedsegmentation relies on the modeling of classifierssimilar to the immune system's T-cells that are capableof binding to, or identifying, self- and non-self cells,with different types of T-cells targeting each type ofpathogen. In the context of brain MR segmentation,different classifiers are utilized for each type of tissueto be identified, e.g., gray matter, white matter andCSF regions in the case of normal brain images.GAIN-based segmentation exhibits severaldistinguishing characteristics including specificity, i.e.,the ability to accurately identify the different classes oftissues, knowledge accumulation, i.e., the ability to addnew tissues to the GAIN network and acquireknowledge that enables the segmentation of the addedtissue, and robustness, i.e., the ability to render similarsegmentation results under the effect of somevariations in external factors, such as noise. However,the most important characteristic of GAIN applicationto MR image segmentation is the low computationalload of the segmentation process that enables highlyreduced segmentation time, which enhances itsviability for use in clinical trials or duringtreatment/surgical planning procedures.

In this paper, a technique for automated MS lesiondetection and quantification is presented based on acascaded pair of grouping artificial immune networks(GAINs) that exploits the characteristics of MSlesions' histograms in MR images in relation to theintensity distributions of normal brain tissues. Thesegmentation is performed in two stages using twochannels: Ti- and T2-weighted MRI images. Thetechnique is evaluated using MR brain imagesgenerated by the BrainWeb simulator [22] in order to

establish a common foundation for comparison to othertechniques that have utilized the same data sets. Theperformance of the proposed technique is comparedwith other techniques utilizing the same MRI data sets,with one distinguishing characteristic of the proposedtechnique. Namely, the proposed technique generatesthe segmentation of MS lesions as well as thesegmentations of other brain regions, instead oflumping them into one category of non-MS lesions,which is beneficial to the evaluation of brain atrophyresulting for the onset of the MS disease.

The paper is organized as follows. Section 2introduces the basic grouping artificial immunenetwork (GAIN) model. Section 3 addresses thetraining of the GAIN model. Section 4 introduces theproposed GAIN-based technique for MS lesionsegmentation. Section 5 presents the experimentalresults and discussion while section 6 presents theconclusions and suggested future work.

2. Grouping artificial immune networks

Various aspects of biology have inspired thedevelopment of computational models. Morespecifically, the biological immune system has drawnsignificant attention and, as a result, Artificial ImmuneSystems (AISs) have emerged as viable computationalmodels. The viability of AISs draws from the powerfulinformation processing capabilities of the immunesystem including feature extraction, patternrecognition, learning, memory, and distributive nature,which provide rich metaphors for its artificiallyinspired equivalents. The most prominent immunesystem principles that are utilized are the immunenetwork theory [23]-[28] along with the mechanisms ofnegative selection [29]-[32] and/or the clonal selection[33].

In order to meet the requirements of MS lesionsegmentation from MRI in clinical environments, thetechnique developed in this paper aims to seekaccuracy, reduced computational load, i.e., reducedsegmentation time, and bounded space usage. Many ofthe previously stated artificial immune algorithmseither rely heavily on floating point computation,which entail increased accuracy as well as a highcomputational load. AISs such as the one introduced in[34] are good primary candidates for their binarynature. However, the number of generated detectorsthat results in similar accuracy to using floating pointcomputations was sometimes larger than the trainingset. Modeling recognition through energies andreactions, introduced in [35], provided an inspirationfor some part of the artificial immune activated neuralnetwork (AIANN) model that we successfully applied

1-4244-1509-8/07/$25.00 02007 IEEE 1140

to normal brain MRI segmentation [39]. In the AIANNmodel, the term detector is used to refer to thecomputational model of receptors on T-cells. Aclassifier represents a model of the T-cell that includesmultiple detectors for different pathogens that can

afflict the body. The self set is the set of classes to berejected by the classifier, just like the immune systemdoes not recognize the self proteins of the body. Thenon-self set is the class of pathogens to be recognizedby the classifier. Each T-Cell performs its own specificpermutation of the pathogen space to accurately bind tothe pathogens that it aims to recognize.

The grouping artificial immune network (GAIN),which is based on the AIANN model [39], was firstintroduced in [21] for the segmentation ofnormal brainMR images, where it has demonstrated appealingfeatures of accuracy and reduced computational load.In the AIANN, the reaction between the detector andthe input vector representing the pathogen occurs

between corresponding pairs of bits. Alternatively, theGAIN model is based on the grouping of bits in theinput pathogen vector into groups of size s that attain a

high degree of mutual information, which are notnecessarily composed of neighboring bits. Then, thereaction takes place between corresponding pairs ofbit-groups. In the context of MRI segmentation, theinput vector is composed of k voxel intensities,representing the voxel to be segmented and those in itsneighborhood. Assuming that a voxel intensity isrepresented using an 8-bit value, 8k/s potential groups

can be created. As a result, the general form for theGAIN output per T-Cell is formulated as follows:

(8k s)-1

RX(P) = X(X, g, P(g))(1g=O

where P(g) represents the value of group g, x denotesthe weights X that are specific to the T-Cell x, Rxrepresents the affinity (or reaction result) of the T-Cellfor the input pattern P. The pattern will be recognizedby the T-Cell showing the highest affinity. If one T-Cell corresponds to each class, the parameters listed onthe right hand side of (1) except for the input P are

specific to each T-cell/class x.

Figure 1 depicts the procedure in which (1) isimplemented. Starting with the voxel's input vector,the bits are regrouped into groups of P(O) to P((8k/s) -1). Forming the groups is achieved by masking therequired bits and shifting them in the correct position,then bitwise OR-ing their values to form a group. Thevalue of that group is used as the index into a lookuptable specific for that group, where the weights of eachpossible pattern are stored. The affinity Rx is computedby adding all the weights that are generated from thelookup process. The computation of the affinity is

efficient in terms of complexity since it involves only(8kls - 1) additions and the rest of required operationsare logical bitwise operations. The bit groupingprocedure is summarized in the following steps:

Input vector of bytes

Re-arranging the bits inlt,

Looklup the weight of ea(based on Its value

Figure 1. Computation procedure of GAIN affinity

The bit groups are defined by finding the pairs ofbits with the highest mutual information with theoutput, which is defined as,

I(X; Y) = E p(x, y) log p(x)P(Y) (2)x y p(x,y)

For each T-Cell x, all the possible training patternsfrom all classes are stored in an arrayX and the outputof that T-Cell x for every one of those patterns is storedin another array Y. The values in Y are thus equal toone only for the non-self patterns of the T-Cell x,

otherwise they are equal to zero. In an exhaustivemanner for each possible group of bits of size s, themutual information is computed with the output Y andstored in a 2D matrix M. If groups are of size 2 bits andthe voxels have 8-bit intensity, M will be an 8x8 squarematrix. Each of the (a, b) positions in M describes themutual information of the group composed of bits a

and b with the output. This process is performed whentraining the GAIN model and the group definitions are

used during segmentation, as shown in Figure 1.

3. GAIN training

This section introduces the GAIN trainingalgorithm. The metric employed to guide the training

1-4244-1509-8/07/$25.00 02007 IEEE

1. Set Rx = 0, where x is the T-Cell number2. For each group g, set v = 0

(a) For every bit b from zero to one in the group g(b) Mask the bit out of its location y in the byte(c) Apply a left shift of b - y(d) Bitwise OR the result with v(e) Repeat for the next bit b

3. Rx= Rx+X(x, g, v)4. Repeat for the next group g

1 141

algorithm, the Dice similarity coefficients [36], isknown for its objective assessment of overlap as wellas visual quality in the context of brain segmentation.For each GAIN model to be trained for an MRIchannel, e.g., TI-weighted or T2-weighted, in order torecognize the voxels belonging to one of n possibleclasses, the training data set is organized into vectorscorresponding to the voxels in the 3D MRI channelunder consideration. Then, n T-Cells of affinities Rxwhere x = 0, ..., n-i are created for each of the classesC,. Training GAIN entails reducing the false negativesand false positives during the classification of eachclass. However, judging the quality of the classificationis specific to the nature of the problem. In the contextof 3D MRI segmentation, the Dice similarity [36]metric D is regularly used to compare the segmentationperformance against the manual segmentation, which isdefined as follows:

D(A, B) = (3)

where A, B are two sets for which similarity is beingcomputed and 1.1 represents the number of elements inthe set. The Dice coefficient is equal to one ifA and Bare identical. As A and B drift apart, D decreases invalue until it reaches zero for two totally different sets,i.e., intersection of the two sets is equal to 0.

In the context of training GAIN to recognize braintissues, A represents the given training set, while B isthe segmentation results of GAIN. Since the Dicecoefficient is used to evaluate the quality of thesegmentation for each of the n classes that are to berecognized, the corresponding D, is defined torepresent the overlap, or lack thereof, of theautomatically segmented class x voxels against thecorresponding class voxels defined in the training setwhere xe 0 (n - 1). Then, an average similaritycoefficient is defined as,

n-lD =ZDx (4)

x=O

to represent the average overlap of all classes. The goalof the training algorithm is to maximizeD , which willbe termed the average overlap.

Training is performed using a numerical version ofthe iterated conditional modes algorithm, where thesought after modes' maxima are the weights X. Since(1) is implemented as a sum of weights, the trainingalgorithm pre-calculates the responses of each cell overthe training set. When the training algorithm varies theweights instead of recalculating the whole Rx, Rx isonly updated using the difference between the oldweight and the new weight as follows:

zIR (X(x, g, v)) = X, (x, g, v) - XO (x, g, v) (5)

where X,(x, g, v) is the new value of the weight andXO(x, g, v) is the old value of the weight. Thus, thecomputation time is minimized during training. In eachiteration, the average overlapD is recalculated andfinally the weight corresponding to the highest value ofthe average overlap is kept.

The training procedure is summarized as follows:1. Arrange all training vectors in one array.2. Apply the procedure in section 2 to find the groupdefinitions.3. Initialize all the weights X(x, g, v) of the T-Cellsrandomly.4. Calculate the responses Rx of all the T-Cells forevery training vector, and store them in an array O(x,y) where x is the T-Cell number, ye O .... Y - 1 is thetraining vector number and Y is the total number ofraining vectors5. Compute the average overlap D and store it in DIast6. Initialize the range of search S = 17. For every weight X(x, g, v)

7.1 Change X(x, g, v) from Xn = Xo(x, g, v) - S. . .Xo(x, g, v) + S in steps of 2S/B where B is thenumber of steps

7.2 For each value ofX, evaluate JRx in (5)7.3 Update O(x, y) and calculate D in (4)7.4 Set X to the value of X, that resulted in the

highest D and update all O(x, y)8. Set S = Sa where a< 19. If Dlast -D #& 0 set Dlast = D and repeat from step 7.

Practically, satisfactory results were obtainedwhen the training parameters were set to a = 0.9 and B= 10 and the algorithm was stopped after threeiterations due to marginal changes of the weights.

4. GAIN-based MS lesion detection

The proposed GAIN-based technique for MS lesionsegmentation in MRI data is motivated by the graylevel distribution of MS lesions in the differentchannels of an MRI study. In Ti -weighted MRI data,MS lesions voxels tend to have higher intensity ratherthan in T2-weighted data. Figure 2 shows thehistograms for WM, CSF, GM and MS lesions in asample TI-weighted and T2-weighted data sets. Thehistogram for the TI-weighted data set indicates thatthe CSF has the minimum overlap with other tissues.Therefore, the TI-weighted data provides the best MRchannel to segment the CSF regions from other braintissues. In addition, the TI-weighted data histogramindicates that the MS lesions are highly overlappedwith both the GM and WM regions, which makes itdifficult to identify the MS lesions from just the Ti-weighted data. This also suggests that a combined

1-4244-1509-8/07/$25.00 02007 IEEE 1142

approach that involves other channels of an MRI studyis needed to effectively segment MS lesions.

On the other hand, the T2-weighted data histogramhas less overlap between the MS lesions and both GMand WM regions. However, the MS lesionssignificantly overlap the CSF regions. In addition, theMS lesions overlap both the GM and WM regions inthe T2-weighted data histogram, albeit having a greateroverlap with GM than with WM regions.

Based on these general characteristics of the Ti-weighted and T2-weighted MRI data histograms asthey relate to MS lesions, a cascaded approachinvolving a pair of GAIN networks is employed foraccurate segmentation of MS lesions. First, the Ti-weighted MRI data is segmented using a GAINnetwork, termed TI-GAIN, to detect the CSF regions.The result of the TI-GAIN segmentation is used tomask the CSF from the T2-weighted MRI data set. Themasked T2-weighted data set is then segmented, usinga GAIN network, termed T2-GAIN, to detect the MSlesions.

The proposed cascaded GAIN technique for MSlesion segmentation consists of four main stages thatare described as follows:Step 1. Preprocessing

This stage involves noise filtering of the sourceMRI data sets to be segmented. For that purpose, anonlinear anisotropic filter is used to reduce the noisein the MRI data sets prior to segmentation [37][38].The filter is tuned according to the noise standarddeviation using parametric adjustment of the factorKappa that determines the extent of noise in the data.As a result, Kappa is set to be proportional to the noisestandard. Different values for the proportionalityconstant have been experimentally utilized and a valueofKappa = 1.5 * noise standard deviation was found togive the best results. In the experimental analysis ofthe GAIN-based MS lesion segmentation, the noisestandard deviation was known a priori since the MRIdata sets were generated using the BrainWebsimulator. In other settings where the noise standarddeviation is unknown, the noise estimation algorithmdescribed in [21] can be applied to estimate the noisestandard deviation before applying the filter.Step 2. Ti-GAIN Segmentation

The TI-weighted data set is segmented using theTI-GAIN as explained in section 2. The features usedare the voxel intensity along with the intensities of thesurrounding six neighbor voxels in the x, y, and zdirections. The output of this stage is the class label ofevery voxel in the data set. The class label can beeither CSF, WM, GM or MS.Step 3. CSF Masking

MS lesions are shown better in T2-weighted dataas the lesions have higher intensity relative to WM.

However, in T2-weighted data the histograms for theMS lesions highly overlap the CSF regions. The resultof segmenting the TI-weighted data can be used toovercome this issue. In this step, the overlap the CSFand MS histograms is resolved by masking out thevoxels labeled as CSF by TI-GAIN the T2-weighteddata. The output of this stage is a T2-weighted data setthat contains only WM, GM, and MS lesions voxels.

r igure L. Histograms of a typical il- and * L-welghtedMRI data set from BrainWeb simulator

Step 4. T2-GAIN SegmentationThe masked T2-weighted data is fed to a GAIN

trained to classify three classes: GM, WM, and MSlesions. The purpose of T2-GAIN is to label theremaining voxels in the T2-weighted data to eitherWM, GM or MS lesions. At this stage, MS lesionsvoxels are labeled. By combining the outputs of step 2and step 3, every voxel in the MR image is labeled aseither CSF, GM, WM or MS lesions. The output of thetechnique is classifying the whole image rather thanonly detecting MS lesions' voxels.

1-4244-1509-8/07/$25.00 02007 IEEE 1143

5. Experimental Results and Discussion

GAIN-based segmentation was applied to thesegmentation of MRI brain data generated through theBrainWeb simulator [22] in order to asses theperformance of the proposed technique in terms ofaccuracy and computational load in comparison withexisting techniques that used the same data [8][9],which forms a common foundation for objectivecomparison. The simulated data sets included both TI-weighed and T2-weighted MRI brain data. For the TI -weighted data, the following settings were used on theBrainWeb simulator: slice sickness = 1 mm, scantechnique SFLASH (spoiled FLASH), repetitiontime (TR) 18 ms, Flip angle = 30 degrees, echo time(TE) = 10 ms, and noise levels of 0o, 1%, 300, 500,and 700. For the T2-weighted data, the followingparameters were used: slice sickness = 1 mm, scantechnique = DSE LATE (dual echo spin echo, lateecho), repetition time (TR) = 3300 ms, Flip angle = 90degrees, echo time (TE) = 35 and 120 ms, and noiselevels of 0o, 1%, 3%o, 5%o, and 7%o. Both data setsconsist of 181 slices with each slice composed of181x217 pixels.

The noiseless TI-weighted and T2-weighted datasets were used for the training of the pair of cascadedGAINs, respectively, as detailed in section 4. Thefeatures used for each voxel were the intensities of thevoxel and its surrounding six voxels, two on either sidein the x (intra-slice), y (intra-slice) and z (inter-slice)directions). After training the cascaded pair of GAINs,they were used in the segmentation of the noisycombined MRI data sets (TI-, T2- applied to eachGAIN). The Dice similarity coefficient [36] was usedto measure the efficiency of segmentation incomparison with the ground truth produced by theBrainWeb simulator since the Dice coefficient wasutilized in published studies using the same data sets,which enables the comparison of GAIN-based MSsegmentation to those studies.

Figure 3 shows the Dice similarity coefficient forthe MS lesions, obtained using the GAIN-basedsegmentation, versus the noise level in the MRI dataset. Figure 3 also include the results obtained by thetechniques proposed in [8][9], which used the sameMRI data sets from the BrainWeb simulator. TheGAIN technique shows marginal improvement atlower levels of noise (less than or equal 50 o) and is at amarginal disadvantage at higher levels of noise (70/O).This can be explained by the large training set, tenMRI data sets, used for the technique proposed in [8]in addition to the prior probabilities for each voxelused in [8] that are generated from 53 data sets. In [9],a normal brain probabilistic model is used instead. In

contrast, the proposed GAIN-based technique uses onedata set for training and no prior probabilities. In termsof accuracy, the proposed GAIN-based MS lesionsegmentation achieves an overall accuracy similar tothat obtained using the techniques in [8][9], albeit atlower computational burden of the training andsegmentation processes. For example, an MRI data setis segmented in less than one minute using thecascaded GAIN while it takes hours to segment a dataset using the technique proposed in [8], excluding thepreprocessing time which is larger in [8].

MS Lesions

Figure 3. Dice similarity for MS at different noise levels

One of the key factors that cause the performanceto decrease with high noise levels is that as the noiselevel increases, more voxels have a higher potential tohigh intensity values. As a result, voxels with highintensity values due to noise tend to expand thedynamic range of the noisy image and, as a result, therelevant tissues' dynamic range is shifted to a lowerrange of the overall histogram due to the fixed numberof bits used in the voxel representation. Consequently,the variation in intensity among relevant tissues iscondensed and more MS voxels are incorrectlyclassified, causing the Dice coefficient to decrease.

Figure 4 illustrates a sample T2-weighted MRslice from the BrainWeb data, the same slice with 1%noise and ground truth of the MS lesions (highlightedin white), the same slice with 900 noise, and thesegmented MS lesions using the GAIN-basedtechnique (highlighted in white). Most of the falsenegatives are segmented as gray matter since graymatter has more overlap with the histogram of MSlesions in T2-weighted data than white matter and CSFis mainly masked in the TI-GAIN.

The Dice coefficient for the CSF, WM and GMregions obtained at different levels of noise using theGAIN-based MS segmentation technique are shown inTable 1, where the ability of the technique inaccurately identifying regions, other than MS lesions,is clearly demonstrated. This is particularly beneficial

1-4244-1509-8/07/$25.00 02007 IEEE

0.90

0.8C

0.65

0.600% 1% 2% 3% 4%

Noise Level

5% 6%

1144

to the quantification of brain atrophy resulting for theonset of the MS disease.

())

(c) (t)Figure 4. Sample T2-weighted Slice (a) Original (b) with1% noise and ground truth, (c) with 9% noise, and (d)

with 9% noise and detected MS lesions (ground truth anddetected MS lesions are highlighted in white)

Table 1. Dice Coefricient for brain tissues, other than MS,obtained using GAIN-based MS segmentationNoise Level CSF WM GM

1% 0.963 0.954 0.9563%o 0.936 0.887 0.91450o 0.925 0.813 0.85970o 0.915 0.732 0.815

7. Conclusion

A new automated technique for MS lesionssegmentation was described. The proposed technique isbased on a novel utilization of a pair of cascadedgrouping artificial immune networks (GAIN) thatexploits the nature of the Ti- and T2-weighted MRIdata histograms. The TI-weighted data is segmentedfirst to detect the cerebrospinal fluid. Then, the CSFregions are masked out from the T2-weighted MRIdata, which is segmented to detect the MS lesions. Theoutput of the GAIN-based technique includes labelingthe MS lesions, the white matter tissue, the gray mattertissue and the cerebrospinal fluid. The performance of

the GAIN-based technique was evaluated using MRIdata at different noise levels, where it demonstratedbetter results than another technique based on artificialneural networks in both segmentation accuracy, at lownoise levels, and segmentation time. Training was doneusing one MR image and no prior probabilities wereused while the other technique used ten images fortraining and prior probabilities for each voxel. Theresults demonstrate the efficiency of using GAIN inquantifying MS lesions in dual channel MRI data.Future work will address the investigation of the effectof utilizing other MRI channel on the MS lesionsegmentation as well as evaluation of the GAIN-basedtechnique when applied to real MRI data of MSpatients.

7. References

[1] A. Evans, D. Collins, and C. Holmes, "Computationalapproaches to quantifying human neuroanatomicalvariability," in Brain Mapping: The Methods, A.W. Toga andJ. C. Mazziotta, Eds. New York: Academic,1996, pp. 343-361.[2] J. N. Giedd, J. W. Snell, N. Lange, J. C. Rajapakse, B. J.Casey, P. L.Kozuch, A. C. Vaituzis, Y. C. Vauss, S. D.Hamburger, D. Kaysen, and J. Rapoport, "Quantitativemagnetic resonance imaging of human brain development:Ages 4-18," Cerebral Cortex, vol. 6, pp. 551-560, 1996.[3] J. N. Giedd, J. Blumenthal, N. 0. Jeffries, F. X.Castellanos, H. Liu, A. Zijdenbos, T. Paus, A. C. Evans, andJ. L. Rapoport, "Brain development during childhood andadolescence: A longitudinal MRI study [letter]," NatureNeurosci., vol. 2, pp. 861-3, Oct 1999.[4] T. Paus, A. Zijdenbos, K. Worsley, D. L. Collins, J.Blumenthal, J. N. Giedd, J. L. Rapoport, and A. C. Evans,"Structural maturation of neural pathways in children andadolescents: In vivo study," Science, vol. 283, pp. 1908-11,Mar. 1999.[5] J. C. Pruessner, D. L. Collins, M. Pruessner, and A. C.Evans, "Age and gender predict volume decline in theanterior and posterior hippocampus in early adulthood," J.Neurosci., vol. 21, pp. 194-200, Jan 2001.[6] J. K. Updupa, L. Wei, S. Samarasekera, Y. Miki, M. A.Van buchem, and R. I. Grossman, "Multiple sclerosis lesionsquantification using fuzzy-connectedness principles," IEEETrans. Med Imaging, vol. 16, pp. 598-609, Oct. 1997.[7] K. V. Leemput, F. Maes, D. Vandermeulen, A.Colchester, and P. Suetens, "Automated segmentation ofmultiple sclerosis lesions by model outlier detection," IEEETrans. Med Imaging, vol. 20, pp. 677-688, Aug. 2001.[8] A. P. Zijdenbos, R. Forghani, and A. C. Evans,"Automatic "Pipeline" analysis of 3-D MRI data for clinicaltrials: Application to multiple sclerosis," IEEE Trans. Med.Imaging, vol. 21, pp. 1280-1291, Oct. 2002.[9] 0. Freifeld, H. Geenspan, and J. Goldberger, "Lesiondetection in noisy MR brain images using constrained GMMand active contours," 4th IEEE Int. Symposium on BiomedicalImaging (ISBI 2007), pp. 596-599, April 2007.

1-4244-1509-8/07/$25.00 02007 IEEE 1145

[10] R. He, M. Mehta, and P. Narayana, "Automaticdelineation of enhancements in MS," Proc. Int. Soc. Mag.Reson. Med 10th Scientific Meeting & Exh., p. 1266, 2002.[11] S. Lukes, S. Crooks, M. Aminoff, L. Kaufman, H.Panitch, C. Mills, and D. Norman, "Nuclear magneticresonance imaging in multiple sclerosis," Ann. Neurol., vol.13,pp. 592-601, 1983.[12] D. Uhlenbrock and S. Sehlen, "The value of Tl-weighted images in the differentiation between MS, whitematter lesions, and subcortical arterioscleroticencephalopathy (SAE)," Neuroradiol., vol. 31, pp. 203-212,1989.[13] R. Grossman, S. Gonzalez, S. Atlas, S. Galetta, and D.Silberberg, "Multiple sclerosis: Gadolinium enhancement inMR imaging," Radiol. vol. 161, pp. 721-725, 1986.[14] J. Sheldon, R. Siddharthan, J. Tobias, W. Sheremata, K.Soila, and M. Viamonte, "MR imaging of multiple sclerosis:Comparison with clinical and CT examinations in 74patients," Amer. J Roentgenol., vol. 145, pp. 957-964, 1985.[15] R. Muller, M. Marsh, M. Bemardo, and P. Lauterbur,"True 3-D imaging of limbs by NMR zeugmatography withoff-resonance irradiation," Eur. J Radiol., vol. 3, pp. 286-290, 1983.[16] G. Dugas-Phocion, M. A. Gonzdlez, G. Malandain, N.Ayache, C. Lebrun, S. Chanalet, and C. Bensa, "Hierarchicalsegmentation of multiple sclerosis lesions in multi-sequenceMRI," Proceedings of IEEE International Symposium onBiomedical Imaging, pp. 157-160, Apr. 2004.[17] R. He, P. Narayana, "Detection and delineation ofmultiple sclerosis lesions in gadolinium-enhanced 3D Tl-weighted MRI data," 13th IEEE Symp. on Computer-BasedMedical Systems (CBMS 2000), pp. 201-206, Jun. 2000.[18] L. N. De Castro and F. J. Von Zuben, "An evolutionaryimmune network for data clustering," Proceedings of IEEESBRN, pp. 84-89, 2000.[19] P. Hajela and J. Lee, "Constrained genetic search viaschema adaptation: An immune network solution," StructuralMultidisciplinary Opt., vol. 12, no. 1, pp. 11-15, 1996.[20] E. Hart and P. M. Ross, "An immune system approachto scheduling in changing environments," Proc. GeneticEvolutionary Comp. Conf (GECCO'99), pp. 1559-66, 1999.[21] M. 0. Ibrahim, Segmentation and Classification ofMRBrain Images using Artificial Immune Models. PhD thesis,University of Miami, 2005.[22] M. U. McConnell Brain Imaging Centre, MontralNeurological Institute, "Brainweb: Simulated braindatabase." http://www.bic.mni.mcgill.ca/brainweb/[23] A. B. Watkins and L. C. Boggess, "A resource limitedartificial immune classifier," Proc. Congress on EvolutionaryComputation (CEC '02), pp. 926-931, 2002.[24] A. Watkins and J. Timmis, "Artificial immunerecognition system (AIRS): Revisions and refinements," 1stInt. Conf Artificial Immune Systems, pp. 173-181, 2002.[25] 0. Nasraoui, F. Gonzalez, and D. Dasgupta, "The fuzzyAIS: Motivations, basic concepts and applications toclustering and web profiling," IEEE Int. Conf on FuzzySystems, pp. 711-717, Hawaii, USA, May 12-17, 2002.[26] 0. Nasraoui, D. Dasgupta, and F. Gonzalez, "Thepromise and challenges of artificial immune systems basedweb usage mining: Preliminary results," Workshop on Web

Analytics at Second SIAM International Conference on DataMining (SDM), pp. 29-39, Arlington, VA, 2002.[27] 0. Nasraoui, D. Dasgupta, and F. Gonzalez, "A novelartificial immune system approach to robust data mining,"Proc. Genetic Evolutionary Comp. Conf (GECCO'02), pp.356-363, New York, NY, 2002.[28] L. N. de Castro and J. Timmis, "Hierarchy andconvergence of immune networks: Basic ideas andpreliminary results," 1st Int. Conf Artificial Immune Systems,pp. 231-240, Canterbury, UK, 2002.[29] M. Ayara, J. Timmis, R. de Lemos, L. de Castro, and R.Duncan, "Negative selection: How to generate detectors," 1stInt. Conf Artificial Immune Systems, pp. 89-98, 2002.[30] F. Gonzalez, D. Dasgupta, and J. Gomez, "The effect ofbinary rules in negative selection," Proc. GeneticEvolutionary Comp. Conf (GECCO'03), pp. 198-209, 2003.[31] F. Gonzalez, D. Dasgupta, and L. Nino, "A randomizedreal-value negative selection algorithm," 2nd Int. ConfArtificial Immune Syst., pp. 261-272, Edinburgh UK, 2003.[32] D. Chao and S. Forrest, "Information immune systems,"Genetic Prog. Evol. Mach., vol. 4, no. 4, pp. 311-331, 2003.[33] V. Cutello, G. Nicosia, and M. Pavone, "An immunealgorithm with hyper-macromutations for the Dill's 2Dhydrophobic-hydrophilic model," Congress on EvolutionaryComputation (CEC2004), pp. 1074-1080, 2004.[34] S. A. Hofmeyr, An Immunological Model of DistributedDetection and Its Application to Computer Security. PhDthesis, University ofNew Mexico, 1999.[35] S. Forrest and M. Oprea, "How the immune systemgenerates diversity: Pathogen space coverage with randomand evolved antibody libraries," Proc. Genetic EvolutionaryComp. Conf (GECCO'99), pp. 1651-1656,1999.[36] K. H. Zou, S. K. Wareld, A. Bharatha, C. M. Tempany,M. R. Kaus, S. J. Haker, W. M. W. III, F. A. Jolesz, and R.Kikinis, "Statistical validation of image segmentation qualitybased on a spatial overlap index," Radiology Alliance forHealth Services Research, vol. 1, pp. 178-189, 2004.[37] P. Perona and J. Malik, "Scale space and edge detectionusing anisotropic diffusion," IEEE Trans. Pattern Anal.Machine Intell., vol. 12, pp. 629-639, July 1990.[38] G. Gerig, 0. Kuibler, R. Kikinis, and F. A. Jolesz,"Nonlinear anisotropic filtering of MRI data," IEEE Trans.Med. Imag., vol. 11, pp. 221-232, June 1992.[39] Akmal A. Younis, Mohamed 0. Ibrahim, Mansur R.Kabuka, and Nigel M. John, "An Artificial ImmuneActivated Neural Network Applied to 3D MRISegmentation," Journal ofDigital Imaging. (in press)

1-4244-1509-8/07/$25.00 02007 IEEE 1146