MRI Brain Abnormality Detection Using Fuzzy Neural Network

Transcription

1 MRI Bran Abnoralt Detecton Usng Fuzz Neural Network Kasra Haghgh Address: Dr.Behesht 36th, No. 81, Mashhad, Iran Abstract In ths paper an expert sste for detecton of bran abnoraltes s proposed. Frst precedng ethods for segentaton of MR ages are revewed and ther ltatons are dscussed. In the proposed ethod, MR ages (three ages fro one slce: 1, and Proton Denst) are acqured fro a scanner or drectl fro MRI sste. For nose deletaton two flters (edan and bandreject lowpass) are used (hs stage s optonal). he ake a clean vew of MR ages. It s necessar to have precse detectons. So b pleentng a gra-scale to color transforaton algorth (t s a radall setrc butterworth band-reject flter), sste can recognze the dfferences between tssues accuratel. Now we have three colored ages (1, and Proton Denst) fro the last secton that better represent tssues and t s possble to sa that those tssues wth the sae color n each of these three ages a be sae tssues. he cobnaton of fuzz sstes and neural Networks ake a powerful tool for pattern recognton probles. So a fuzzfed neural network wth outputs to a back-propagaton network for tssues recognton ust be used. herefore a fuzz neuron and a fuzzfed network are ntroduced. he output of the back-propagaton network s the tpe of tssue under process. he results of the last secton are fed nto another network that uses a knowledge-base to ake a suggeston for treatent (hs level s also optonal). Because of the te ltaton MALAB 4. for Wndows s used for expert sste sulaton. It has several abltes for atrx calculatons and graphng that ake the work easer (soe of the base odules are wrtten n C++). A. Introducton S o far several ethods have been proposed for tssue segentaton and classfcaton of MR ages. he used several ethods for ths purpose that have soe ltatons and boundares. Soe of the conventonal ethods are: 1. Edge Detecton: hat uses a gradent operator for detecton of the age edges [1],[].. Regon Detecton: In ths ethod the gra level of each voxel s copared wth a threshold [3],[4]. 3. Statstcal approaches: he a be dvded as supervsed and unsupervsed. In supervsed segentaton the sets of tranng saples are

2 labeled accordng to pror known ebershp. he other ethod usuall assues a odel (such as GMRF) for gra-value dstrbuton or texture features wthn each class and estates the odel paraeters b fttng the ages to the odel. [6],[7] 4. Fuzz clusterng technques: In ths ethod the [5] use Fuzz c-ean algorth for segentng MRI volues of the bran. hese ethods all have certan ltatons for detecton and segentaton of tssues wth the sae gra-scale but dfferent tpe n MR Iages. Although neural networks are wdel used for a nuber of pattern recognton probles but for ths specfc proble a specfc tpe ust be used. On the other hand fuzz logc s a powerful tool for dealng wth uncertan data. For ths case we are dealng wth the data wth no fx specfcatons for all cases, so a fuzzfed neural network s requred. Ipleentaton of the expert sste requres the followng processes. In secton I data acquston, flterng and the state of data s deonstrated. In secton II a gra-scale to color transforaton algorth s consdered and dfferences between gra-scale MR-age and ts colored age s copared. In secton III fuzz neurons and fuzzfed neural networks s ntroduced and a specal network consderaton s suggested. In secton IV fnall b applng the results of the secton III, an expert sste for detecton of bran s abnoralt and ts treatent s consdered. B. Methods and Data I. Data In the frst stage the nput data of the sste ust be represented. In ths level for each slce of a bran 3 ages (1, and Proton Denst) ust be obtaned. So these ages are acqured fro a sheet scanner (wth 6 DPI resoluton) or drectl fro MRI sste (b a hardware nterface). he ages have 16-bt gra level resoluton. hese ages a have several artfacts and unwanted noses, so for deletng these noses a flter should be used [8]. Frst we use a 3 3 (for hgh resoluton ages) edan flter. hen a bandreject lowpass flter wll be run, and after that soe of reconstructon ethods such as [9] a be used. We can easl copare dfferences between orgnal and fltered ages n fgure 1. hese three ages (1, and PD) could show tssues' states clearl. he ages were acqured wth followng paraeters: the feld of vew was 3. c, wth 7. slce spacng, and three exctatons.

3 Fgure 1: Orgnal and ts fltered MR age II. Colorng MR gra-scale ages In the next secton we use a ethod for better recognton of the dfferences between tssues. In ths ethod the gra-scale age s transfored to a color age. hs stage anfests the dfferences between closer gra-scales. For pleentng ths sesson a flter for pseudo-color age enhanceent s used [8]. hs flter uses Fourer transfored for of an age. hen a radall setrc Butter-worth bandreject flter separates red, green and blue colors fro the ages after the nverse Fourer transfor. B settng the varables to the approprate values, a colored age s obtaned. he forulas for ths transfor are: G( u, v) = H( u,v)f( u, v) H(u, v) = 1+ L M N 1 D( u,v)w D (u, v) D P Q n O D( u,v) = u u + v v b g b g 1 where F(u,v) s the Fourer transfored of orgnal age, W s defned as the wdth of the band, D s ts center and n s the flter s order. In fgure we can easl dstngush between the sae gra-scaled tssues n colored age. hs algorth s run over each set of three ages. hs secton akes the recogntons ore accurate because tssues' dfferences ncreased.

4 Fgure : Orgnal and ts colored MR ages III. Fuzz Neurons and Fuzzfed Neural Networks he cobnaton of the features ofuzz sstes (the reason wth ultvalued sets and work wth lngustc nput and output. he have had an applcatons n pattern recognton n recent ears [1],[11].) and the features of neural networks (wth learnng ablt and a hgh speed parallel structure) akes a powerful tool for pattern recognton. Man works have been carred out on fuzz neural sstes for pattern recognton. A tpcal nonfuzz neuron wth N nputs sus these nputs through the correspondng weghts and expressed as: N F G H 1 = f w x = I J K A fuzz neuron wth N weghted nputs (real nubers) and M outputs (n nterval [,1]) could assocate each output wth the ebershp values (the degree that nputs belong to a fuzz set) of a fuzz concept. herefore four tpes of fuzz neurons that construct a four laer feedforward fuzz neural network as shown n fgure 3, are used. [11] In the frst laer nputs are fed nto the network. In ths level all the nput values dvded b the axu value aong all nput patterns. he second portant laer s known as MAX Laer, that fuzzf nput patterns through a weght functon w(,n) (as defned next). he state of (p,q)th ax n ths laer s:

5 F s = ax ax dwbp q, jg pq H j I j K that =1 to N1, j=1 to N, p=1 to N1, q=1 to N w, n = exp e + n b g β c hj that = --(N1-1) to (N1-1), n= -(N-1) to (N-1) where j s the output of the prevous level, β s the fuzzfcaton degree and N1 and N are the sze of nput laer. w(,n) s called the fuzzfcaton functon. he outputs of (p,q)th MAX n ths laer are: pn R s = 1 S pq Θ pq that α, p=1 to N1, q=1 to N, =1 to M α f α s Θ pq pq f otherwse where Θ pq s the central pont of functon pq, α, Θ pq, for ever set of p,q deterned b the learnng algorth. he thrd laer that uses for fuzz deducton has the output such as: F = n p H nd q I pq K that =1 to M, p=1 to N1, q=1 to N. where pq s the output of the prevous level and M s the sze of thrd laer. he fourth laer provdes nonfuzz outputs at the end of the network. (hs laer a be otted).he algorth of the th cop fuzz neuron n the fourth laer s: Out = = axb R S 1 f f g < = Input laer Max Laer Mn Laer Co Laer Fgure 3: Fuzz Neural Network Structure hs network s ver strong, although for accurate and hgh speed recognton a back-propagaton network s used. hs s connected to the output of last laer (conventonal ethods for ths network are used).

6 IV. Expert Sste and Knowledge-base In the precedng stage a neural network cobnaton s defned. B usng ts applcatons an expert sste for ths work could be defned. For ths purpose the followng procedures ust be followed (fgure 4): 1. Frst, sste searches n three ages for the sae colored (In one age) parts. he ght be the sae tssues. hen these parts have a value correspondng to the color n these ages, therefore we have three values for each tssue (x 1,x,x PD ), so we put the n a feature vector. hese feature vectors are used n the next step. he represent the tssues' characterstcs better.. Now the output vectors ust be classfed to the tssue classes and see f t s noral or abnoral. So for pleentng ths secton the ntroduced network s used. A network wth 3 1 nput laer that outputs on ffteen tssues (noral and abnoral) s defned. For recognton of the knd of abnoralt and ts state, another backpropagaton neural network s traned, where ts nput s the output of the precedng network (tssues' tpes) and ts output s the knd of patholog. 3. After recognton of the knd of patholog a proper suggeston regardng surgcal operaton, radaton therap or checal therap ust be ade. In ths level a knowledge-base that could propose an approprate treatent usng the output of prevous level, s used. Feature vec. Fuzz Net. Backprop Net. Suggeston Fgure 4: Expert Sste Structure C. Sulaton Results and Concluson Because of the te ltaton I use MALAB 4. for Wndows for sulaton of the expert sste. he use of MALAB Neural Network oolbox akes the research easer, and the MALAB s ablt to perfor nterestng graphs and longer calculatons could be used easl. But the sste s speed grows hgher f a hardware neural processor and ts software s used. hs sste has the perforance and accurac to recognze abnoralt 88 percent correct and t grows b ore tranng sessons. Acknowledgents he author s grateful to Dr. Nader Jahngr for paper edtng and to Dr. Bahra Bahra Motlagh for bologcal and radologcal exanatons and hs helpful suggestons.

7 References 1. Boans M., Hohne K.H., ede U., Reer M., 3-D segentaton of MR ages of the head for 3-D dspla, IEEE rans. Med. Iag. 9,199.. Chen C.., sao C.K., Ln W.C., Medcal age segentaton b a constrant satsfcaton neural network, IEEE rans. N. S. 38, Shapoo P.K., Soltan S., Wong A.K., Chen Y.C., A surve of thersholdng technques, Coput Vson Graph Iage Proc 41, Young I.R., Fu K.S., Handbook of pattern recognton and Iage processng, Acadec Press, New York, Clark M., Hall L.O., Velthuzen R.P., Slbger M.S., MRI segentaton usng fuzz clusterng technques, IEEE Eng. n Med. Bo., Nov Lang Z., ssue Classfcaton and Segentaton of MR Iages, IEEE Engneerng n Medcne and Bolog, March, 1993, axt., Lundervold A.,Multspectral Analss of the Bran Usng MRI, IEEE rans. on Medcal Iagng, 1994, 13, 3, Gonzalez R.C., Wntz P., Dgtal Iage Processng, Addson-Wesele Pub. Co., Kraer D. M., Kaufan L., Guzan R.J., Hawrszko C., A General Algorth for Oblque Iage Reconstructon, IEEE Coputer Graphcs & Applcatons, March, 199, Kwan H.K., Ca Y., A fuzz neural network and ts applcaton to pattern recognton, IEEE rans on Fuzz Ss,, 1994, Keller J.M., Gra M.R., Gvens J.A., IEEE rans. on Sste, Man and Cbernetcs,1985, 15, 4,