A FUZZY FRAMEWORK FOR SEGMENTATION, FEATURE MATCHING AND RETRIEVAL OF BRAIN MR IMAGES Archana.S 1 an Srihar.S 2 1 Department of Information Science an Technology, College of Engineering, Guiny archana.santhira@gmail.com 2 Department of Information Science an Technology, College of Engineering, Guiny ssrihar@annauniv.eu ABSTRACT This paper proposes a complete framework base on fuzzy logic, for the retrieval of MRI images of the brain. In our system, the MRI image is segmente using the fuzzy local information c means (FLICM) algorithm. Region, location an shape properties are extracte from each region. These properties are represente as fuzzy features. Thus each image is represente by a set of fuzzy features corresponing to the various regions obtaine after segmentation. To compute the overall similarity between two images, unifie feature matching (UFM) is use. This measure integrates the properties of all the regions in an image. Our system can be applie to istinguish between the tumor an normal magnetic resonance images. Also the location feature of each region, especially of the tumor region, obtaine after segmentation can be use to etermine the type an symptoms of the brain tumor. Images from the brain web were use for experimentation. KEYWORDS Fuzzy logic, Segmentation, Clustering, Feature Matching, Fuzzy Similarity 1. INTRODUCTION These ays, the amount of meical ata generate in the form of images is increasing tremenously. These images of various imaging moalities like computer tomography an Magnetic Resonance Imaging are generate aily. They are use by researchers an octors for ecision making. The evelopment of systems for managing an archiving of these ata is becoming a strenuous ob. These ata if organize properly can be use for aiing the octors in iagnosis proceures. Magnetic Resonance Imaging is an imaging technique that is useful for istinguishing the various soft tissues of the boy. This technique is use for imaging brain, heart an other muscles. While this technique is avantageous compare to Compute Tomography (CT) an X ray techniques, there arise certain issues. The brain tissue can be classifie into three types on the basis of the brain MR image. They are Grey Matter (GM), White Matter an Cerebro Spinal Flui (CSF). These tissues can be ifferentiate by their intensity ranges. There are ifferent clustering algorithms which segment the MR image base on intensity values. But uncertainties arise in MRI images ue to the noise an partial volume effects. Simple har clustering algorithms will not be effective. These uncertainties can be moelle effectively using fuzzy logic [1]. Thus the image can be segmente by incorporating fuzzy techniques into segmentation algorithms. DOI : 10.5121/icsit.2011.3404 45
For effective organization these MR images are store in atabases. The existing atabases make use of a simple keywor base querying techniques. These atabases are epenent on ata entry. There may be manual errors in these atabases. This kin of retrieval is highly insufficient. In orer to make the atabase more useful we nee a content base retrieval system (CBIR). The content base retrieval systems make use of the properties in the images for inexing an retrieval. The features of the images nee to be use rather than simply labels for efficient iagnosis. The aim is to evelop a fuzzy framework for effective retrieval of images.the first step in any kin of image analysis is the image segmentation. The segmentation of the brain MRI to istinguish between the tumour an normal tissues has been uner research for long years. To automate the entire system unsupervise clustering [2] is use for the segmentation of the brain MRI images. This partitions the image into ifferent regions. After segmentation the features are extracte from each region an a fuzzy representation of the image is obtaine. Then the unifie feature matching [3] is use for integrating the properties of all the regions in the image. 2. FUZZY FRAMEWORK The main aim is to evelop an operational fuzzy framework for the retrieval of brain MR Images. This framework helps to istinguish between tumour an non tumour images. Also retrieves tumour images base on the location of the tumour. This location information is very useful to the iagnosis as the symptoms, type an treatment of the tumour varies with the location of the tumour. The framework consists of the following stages: 2.1 Image Pre-processing an Segmentation 2.2 Feature Extraction 2.3 Fuzzy Feature Representation 2.4 Unifie Feature Matching The system architecture is given below: SIMILAR IMAGES FUZZY PROCESSING UNIT UNIFIED FEATURE MATCHING FEATURE DATABASE QUERY IMAGE IMAGE DATABASE FUZZY PROCESSING UNIT FUZZY REPRESENTATION Figure 1.System Architecture 46
The Fuzzy Processing Unit is given below: FUZZY IMAGE SEGMENTATION FEATURE EXTRACTION FUZZY FEATURE REPRESENTATION Figure 2. Fuzzy Processing Unit 2.1. Image Pre Processing an Segmentation 2.1.1 Image Pre Processing The pre processing is one to improve the contrast of the brain MR image. This is one using a Fuzzy Intensification operator [4]. For an image I, the grey level at (m, n) is given by g, an the maximum an minimum value of the grey level is given by function is efine as µ g = + m n 1 max g F F e gmax an g min. The membership Where F an Fe are the enominational an the exponential fuzzifiers respectively that controls the amount of greyness ambiguity. The membership values are moifie using the intensification operator, [ ] 2 µ = 2 µ,0 µ 0.5 ' µ µ 2 = 1-2[1- ],0.5 1 Once the membership values are moifie, moifie grey values are transforme to spatial omain using an inverse function, g = G ( µ ) ' 1 ' Where, ' F ( ) = gmax F [( ) e µ 1] 1 ' G µ 1 47
Fuzzy Intensification Brain MR Image Contrast Enhance Image 2.1.2 Fuzzy Segmentation Figure 3. Contrast Enhancement The segmentation of the contrast enhance image is one using five algorithms, Fuzzy C Means, Enhance Fuzzy C Means, Suppresse Fuzzy C Means, Fast Generalize Fuzzy C Means an Fuzzy Local Information C Means. The algorithms iffer by their obective functions. All the algorithms, except fuzzy c means, incorporate the local an spatial information into their obective functions. The results of segmentation are compare with the manually segmente images from the Internet Brain Segmentation Repository. Of the five algorithms consiere the Fuzzy Local Information C Means is foun to be the most effective in terms of the Correlation Inex an Mean Square Error. The correlation inex is given by, r = m n ( A A)( B B) 2 2 ( ( A A) )( ( B B) ) m n m n Table 1. Correlation Inex of various images Image # FCM EnFCM FCM_S FGFCM FLICM 1 0.5117 0.6568 0.6569 0.6971 0.8234 2 0.5665 0.6490 0.6565 0.7012 0.7980 3 0.5521 0.6321 0.6754 0.7080 0.8010 4 0.5612 0.6540 0.6901 0.7122 0.8011 5 0.5211 0.6321 0.7010 0.7432 0.8760 The above table inicates that the Fuzzy Local Information C Means is superior to the other algorithms uner consieration. Thus we use the same for the segmentation moule. 48
Fuzzy Local Information C Means Normal Image Segmente Regions of the Image Figure 4. Fuzzy Local Information C Means Segmentation of normal image Fuzzy Local Information C Means Tumour Image Segmente Regions Figure 5. Fuzzy Local Information C Means Segmentation of the Tumour Image 49
2.2. Feature Extraction The regions of the image are then subect to feature extraction. The features such as the Hu moments [7], normalize inertia an location features [8] are extracte. Normalize inertia of orer γ is calculate for each region R given as, I $ $ ( x x) + y y ( x, y) R ( R, γ ) = 1 + γ /2 V ( R ) ( ) 2 2 γ /2 Where ( x$, $ y ) is the centroi of is invariant to rotation an scaling. 2.3. Fuzzy Feature Representation R, V( R ) is the volume of R. The normalize inertia The feature vectors obtaine above are very long since each region is subecte to feature extraction. Also the features are subecte to segmentation relate uncertainties. To overcome these rawbacks we evelop a fuzzy feature representation base on the Cauchy membership function. The segmente image can be viewe as a collection of regions {R 1, R 2,.., R n }. Equivalently in the feature space as {F 1, F 2,.., F n }. Each region R is represente by the centre uur ( f ) efine as ur f uur ur f F ( f ) = V ( F ) uur ( f ) of the corresponing feature set ( F ) with Which is essentially the mean of all elements of ( F ). Accoringly, region is represente by fuzzy feature F whose membership function, ur 1 µ ( f ) = F ur uur f f 1 + ( ) f Where, C 1 C 2 ur uur f = fi fk C( C 1) i= 1 k = i+ 1 Is the average istance between the cluster centers.. For an image with regions R,1<<C, F is calle the fuzzy feature representation (signature) of the image where F={ F :1<<C}, C is the total number of regions in segmente image. 50
2.4. Unifie Feature Matching The UFM scheme characterizes the resemblance between images by integrating the properties of all the regions in the images. Let us assume we have 2 fuzzy feature sets A an B with Cauchy membership functions An 1 ( x r µ ) = r r A x u 1 + ( ) r 1 µ ( x) = r r B x v 1 + ( ) The fuzzy similarity can be written as, ( a + b ) S( A, B) = r r ( + ) + u v a b b Having the similarity measure, the similarity vector is constructe. For every Ai A, we efine the similarity measure for it an B as C b B l ( i = S Ai, U B ) 1 Combining we get a vector, rb l = [ l1, l2,..., l C ] B B B T Similarly for every, B C a A l ( = S B, U Ai ) i 1 Combining we get a vector, ra l = [ l1, l2,..., l C ] A A A T a B,the similarity measure for it an A as, b Thus, we efine a similarity vector for A an B enote by ur L ( A, B) r B l = A r. l Also weights can be given to each feature to give more importance to that feature while retrieving the results. In our system we give importance to the location of the regions. This is highly useful in the etection of the type of brain tumour an its symptoms. The UFM scheme is more avantageous than the other features. 51
It can be analyse by the following graphs: Precision for ifferent types of Feature Vectors Precision 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 2 3 4 No. of Query Images Location Feature Hu Moment Haralick Texture UFM measure Figure 6. Precision Graph of UFM an other feature vectors 3. EXPERIMENTAL RESULTS In orer to test the frame work the MR image slices from the Internet Brain Segmentation Repository an the Brain Web are use. Aroun 400 image slices of resolution 256x256 are use. The system is implemente using MATLAB on a 1.6Ghz Intel(R) Core(TM) 2Duo CPU with 3GB of RAM. The Matlab is a Matrix laboratory an it is a high-performance language for technical computing integrates computation, visualization, an programming in an easy-to-use environment. We evaluate the system using the stanar precision an recall measures. Precision is efine as the proportion of the relevant ocuments in a set of ocuments retrieve by the search. No of Relevant Images P = No of Retrieve Images Recall is efine as the number of relevant ocuments retrieve as fraction of all relevant ocuments. No of relevant images retrieve R= Total No. relevant images in Database We assume that we have 4 categories of tumour images an there are 10 images in each category accounting to a total of 40 images. The precision recall graph is shown in fig 7. Table 2 gives the precision an recall values for 4 ifferent queries. 52
PRECISION RECALL GRAPH PRECISION 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.5 0.6 0.4 0.3 RECALL Figure 7. Precision Recall Graph Table 2. Precision Recall Values Image Query Precision Recall 1 0.714 0.5 2 0.851 0.6 3 0.571 0.4 4 0.428 0.3 FUZZY FRAMEWORK Query Image with Tumour on the left Output containing best six results with tumour on the left Figure 8. A test run isplaying the first six results 53
4. CONCLUSIONS AND FUTURE WORK Thus the system was implemente successfully in MATLAB. The atabase contains 250 normal an 150 tumour images. The resolution of these images is 256x256.A query by example scheme of retrieval was use. Since the location feature of each region is given more importance, the similar images with tumour on the same sie as the test image are retrieve. The location of the tumour is very important in etermining the type, treatment an symptoms of the tumour. The system uses fuzzy logic at all stages in the framework. The Fuzzy Local Information C Means (FLICM) algorithm use for segmentation yiele an average correlation inex of 0.80. This was higher than the rest of the segmentation algorithms consiere. The combination of various features for matching yiele better results than iniviual features. The precision factor was as high as 80 percentage. The experimental results at each stage (from segmentation to feature matching) show that the fuzzy logic yiels better results than the conventional techniques. The Unifie Feature Matching technique helps us in giving more importance to specific features while incluing all the features. The system can help or ai in the iagnosis if the profile information of the patient an the report of the MR image are inclue in the atabase along with the images. Also the fuzzy clustering algorithms may be improve by using the features of the MR images rather than ust intensity alone. Other membership functions can be use in place of the Cauchy membership function for the feature matching. Weights can be varie for ifferent features epening on the application. Our system gives importance to the location of the tumour. Another application may give importance to size of tumour. REFERENCES [1] L. Zaeh, Fuzzy sets, IEEE Transactions on Information Control, vol. 8,1965, pp. 338 353. [2] Zheru Chi, Hong Yan, Tuan Pham, Fuzzy algorithms: with applications to image processing an pattern recognition, Worl Scientific Publishing Co. Pvt. Lt, vol. 10,1996, pp. 85-91. [3] Yixin Chen, James Z. Wang, A region base Fuzzy Feature Matching Approach to Content Base Image Retrieval, IEEE Transactions on Pattern Analysis an Machine Intelligence,vol. 24,2002,pp. 1252-1266. [4] Tamalika Chaira, Aoy Kumar Ray, Fuzzy Image Processing an Applications with MATLAB, CRC Press, vol. 1, 2010,pp. 47-55. [5] Stelios Kriniis an Vassilios Chatzis, A robust fuzzy local information C-Means Algorithm, IEEE Transactions on Image Processing, vol.19,2010,pp. 1328-1337. [6] Robert.M.Haralick, K.Shanmugam,Its hak Dinstein, Textural Features for Image Classification,IEEE Transactions on Systems, Man an Cybernetics,vol. 6,1973,pp.610-621. [7] Hu, M. K., Visual pattern recognition by moments Invariants, IRE Trans. Information Theory,vol. 8,1962,pp. 179-187. [8] B.G.Prasa, K.K.Biswas an S.K. Gupta, Region-base image retrieval using integrate color, shape, an location inex, Computer Vision an Image Unerstaning, vol. 94, (2003),pp 193-233. [9] Images:http://www.bic.i.mcgill.ca/brainweb/ [10] C.A. Cocosco, V. Kollokian, R.K.-S. Kwan, A.C. Evans : "BrainWeb: Online Interface to a 3D MRI Simulate Brain Database,NeuroImage, vol.5, no.4, part 2/4, S425, 1997 -- Proceeings of 3-r International Conference on Functional Mapping of the Human Brain, Copenhagen, May 1997. 54
[11] Images-http://www.cma.mgh.harvar.eu/ibsr/ Authors [1] S.Archana has complete her Post Grauation in Multimeia Technology. Her area of interest inclues Image processing an Data Structures. [2] Dr. S. Srihar is an Associate Professor in Department of Information Science an Technology at College of Engineering, Guiny. His area of interest inclues Image Processing, Pattern Recognition, Data Mining, Artificial Intelligence. 55