Feature Extraction and Rule Classification Algorithm of Digital Mammography based on Rough Set Theory

Transcription

1 Feature Extraction an Rule Classification Algorithm of Digital Mammography base on Rough Set Theory Aboul Ella Hassanien Jafar M. H. Ali. Kuwait University, Faculty of Aministrative Science, Quantitative Methos an Information Systems Department P.O..Box 5969 Safat, coe no Kuwait Abstract: Breast cancer represents the secon leaing cause of cancer eaths in women toay an it is the most common type of cancer in women. This paper presents an efficient classification algorithm in igital mammograms in the context of rough set theory. Feature extractions acquire in this work are erive from the gray-level co-occurrence matrix. The features are extracte, normalize an then the rough set epenency rules are generate irectly from the real value attribute vector. Then the classification is built an the quaratic istance function use to etermines similarity between a query an atabase image. The experimental results show that the propose algorithm performs well reaching over 85% in accuracy. Keywors: rough sets, mammograms, classification, feature extraction, rule generation, similarity measure, gray-level co-occurrence matrices 1. Introuction Breast cancer is a ealy isease that aversely affects the lives of far too many people, primarily women. Accoring to the National Cancer Institute [5], each year about 180,000 women in the Unite States evelop breast cancer, an about 48,000 lose their lives to this isease. It is also reporte that a woman's lifetime risk of eveloping breast cancer is one in eight. Currently, igital mammography [11,13,15,16] is one of the most promising cancer control strategies since the cause of breast cancer is still unknown. Mammography is a specific type of imaging that uses a low-ose x-ray system an high-contrast, high-resolution film for examination of the breasts. Most meical experts agree that successful treatment of breast cancer often is linke to early iagnosis. Mammography plays a central part in early etection of breast cancers because it can show changes in the breast up to two years before a patient or physician can feel them. Classification is a form of meical ata analysis, which can be use to extract moels escribing important ata classes or to preict future ata trens. In other wors, classification is to ientify essential features of ifferent classes base on a training set an then classify new instances into the appropriate classes [8,1]. Classification of igital mammogram of breast lesions as malignant or benign must be base on information present in the mammogram images. In aitional, clinical ata such as patient age can also be use. A common approach to this task is to extract information (calle features) from mammograms an then buil a classifier moel base on rough set theory to make the malignant versus benign assessment. In this paper, we introuce an efficient classification approach base on the context of rough set theory which applie on five statistical extracte features from the igital mammograms. Rough set concept was introuce by Polish logician, Professor Zzisław Pawlak in early eighties [17]. This theory become very popular among scientists aroun the worl an the rough set is now one of the most eveloping 1

2 intelligent ata analysis. Rough sets ata analysis was use for the iscovery of ata epenencies, ata reuction, approximate set classification, an rule inuction from atabases. The generate rules represent the unerlying semantic content of the images in the atabase. A classification mechanism is evelope by which the images are classifie accoring to the generate rules. The analysis of meical, biological an health-relate ata has been one of the successful applications of rough sets. In some cases, the rough set analysis has given meical octors new insight into their practice an training. This paper is organize as follows. Section epicts the general classification process. Section 3 gives a brief escription of the image pre-processing phase. Feature extraction base on Gray-level Co-occurrence Matrix is presente in Section 4. In section 5, the funamental of rough set theory is introuce. The rule generation an rule classification algorithm is iscusse in Section 6. Experimental results are given an iscusse in section 7. The paper is conclue in section 8.. Image Classification Process Classification is a function that classifies a ata item into one of several preefine classes. In other wors, classification is the process that establishes classes with attributes from a set of instances (calle training sets). The class of an instance must be one from a finite set of possible, pre-etermine class values, while attributes of the instance are escriptions of the instance potentially affecting its class. A training set is a set of tuples efine on a set of multiple attributes (also, calle features). Each tuple has a known class label, which is calle a target class, associate with it [14]. The aim of classification process is to inuce a classifier that can be use to classify future ata items whose classification is unknown. The classification is base on a well-efine set of classes an a training set of pre-classifie examples (tuples). The knowlege prouce uring the classification process can be extracte an represente in the form of rules. These classification rules allow one to evelop a profile of items belonging to a particular group accoring to their common attributes. This profile can then be use to classify new ata items that are ae to the atabase. The accuracy of a classifier on a given test set is the percentage of test set samples that are correctly classifie by the moel. If the accuracy of the classifier is consiere acceptable, the classifier can be use to classify future ata tuples or obects for which the class label is not known. Figure (1) illustrates the classification scheme. The architecture for image classification uner the rough set theory framework [] is compose by four funamental builing phases: pre-processing, feature extraction, rule generation an similarity function. Mammograms are images ifficult to interpret, an a pre-processing phase of the images is necessary to improve the quality of the images an make the feature extraction phase more reliable. After enhancing the images, feature relevant to classification are extracte an represente in a atabase. Given a feature representation for each atabase image, retrieval consists of extracting a set of feature vectors from the querie image an relying on a similarity function to evaluate which feature representation best explains those features. In feature representation, a scheme is neee to convert an attribute vector to be rea by the rough set rule generator. Rule generation forms the core of the rough set framework. The preiction of new obects is epenent on the type of rules generate an some strategies will work well for preiction. Rules generate can be

3 use to preict values of ecision attributes (classify) for new obect. The rules of the query specification are compare with rules of the image atabase to etermine which images match correctly (similar) with the given rules. The matching process is base on similarity measure between query image an images in atabase. Image Acquisition Image Enhancement Feature Extraction Rule Generation Classification moel Rule Classification Figure 1: classification scheme 3. Pre-Processing phase Diagnosing cancer tissues using igital mammograms is a time consuming task even for highly skille raiologists because mammograms are low contrast, noisy images. Therefore, in igital mammogram there is a nee for enhancing imaging before a reasonable feature extraction can be achieve. Image enhancement in meical computing is the use of computers to make an image clearer [3]. This may be to ai interpretation by humans or computers. Types of image enhancement inclue, noise reuction, ege enhancement an contrast enhancement. Contrast enhancement is useful when an area of the image that is of particular importance has only subtle changes in pixel intensity. In these cases it may be ifficult for the human eye to make out the structures clearly, especially if the image is being isplaye on a low quality screen, or limite contrast like mammograms image. Applying the contrast enhancement filter will improve the reaability of areas with subtle changes in contrast but will also estroy areas of the image where the intensity of the pixels is outsie the range of intensities being enhance. In this work, we applie pre-processing techniques base on mathematical morphology theory to enhancing the contrast of the original mammogram image to help in feature extraction process [3, 6]. 4. Texture Feature Extraction Texture is one of the most important efining characteristics of an image. It is characterize by the spatial istribution of gray levels in a neighborhoo [9]. In orer to capture the spatial epenence of gray-level values which contribute to the perception of texture, a two-imensional epenence texture analysis matrix is iscusse for texture 3

4 consieration. Since texture shows its characteristics by both each pixel an pixel values. There are many approaches using for texture classification. The gray-level co-occurrence matrix seems to be a well-know statistical technique for feature extraction [4, 14]. 4.1 Gray-level Co-occurrence Matrix Co-occurrence matrix, the secon-orer histogram, is the two imensional matrix of oint probabilities P, r ( between pairs of pixels, separate by a istance in a given irection r. Haralick et al., [9] efine 14 texture features erive from the co-occurrence matrix. In this paper, five features were selecte for further stuy; maximum probability, contrast, inverse ifferent moment, angular secon moment, an entropy. Table (1) gives the escription of these selecte features. Where P ( refers to the normalize entry of the co-occurrence matrices. That is = P ( / R where R is the total number of pixel pairs ( i,. For a isplacement vector = ( x, y) an image of size NxM R is given by ( N x)( M y). Feature Equation Description Maximum probability max This is simply the largest entry in the matrix, an correspons to the strongest response. Contrast ( i A measure of the image contrast or the amount of local variations present in an image. 1 This escriptor has large values Inverse ifferent moment in cases where the largest 1+ ( i elements in P are along the principal iagonal. Angular secon moment Entropy log( ) A measure of the homogeneity of an image. Hence it is a suitable measure for etection of isorers in textures. For homogeneous textures value of angular secon moment turns out to be small compare to nonhomogeneous ones. Entropy is a measure of information content. It measures the ranomness of intensity istribution. Table 1: texture features extracte from Co-occurrence Matrix 5. Rough Set Theory: Theoretical Backgroun Let us present here some preliminaries of rough set theory, which are relevant to this work. For etails one may refer to [7, 17, an 18]. 4

5 5.1 Information Systems Knowlege representation in rough sets is one via information systems, which are a tabular form of an OBJECT ATTRIBUTE VALUE relationship. More precisely, an information system, Γ =< U, Ω, V q, f q > qεω, where - U is a finite set of obects, U = { x1, x, x3,..., xn} - Ω is a finite set of attributes (features), the attributes in Ω are further classifie into isoint conition attributes A an ecision attributes D, Ω = A U D - For each q Ω, V is a set of attribute values for q, q Each f q: U Vq is an information function which assigns particular values from omains of attributes to obects such that f ( x ) V for all U an q Ω. q i q x i With respect to a given q, the functions partitions the universe into a set of pairwise isoints subsets of U: R q = x : x U f ( x, q) = f ( x, q) x } (1) { 0 0 U Assume a subset of the set of attributes, P A. Two samples x an y in U are iniscernible with respect to P if an only if f ( x, q) = f ( y, q) q P. The iniscernibility relation for all P A is written as IND (P). U / IND( P) is use to enote the partition of U given IND(P) an is calculate as follows: U / IND( P) = { q P : U / IND( P)({ q})}, () A B = { X Y : q A, Y B, X Y {}}. (3) 5. Approximation Spaces A rough set approximates traitional sets using a pair of sets name the lower an upper approximation of the set. The lower an upper approximations of a set P U, are efine by equations (4) an (5), respectively. PY = U { X : X U / IND( P), X Y} (4) PY = U { X : X U / IND( P), X Y {}} (5) Assuming P an Q are equivalence relations in U, the important concept positive region POS P (Q) is efine as: P U POS ( Q) = PX (6) X Q 5

6 A positive region contains all patterns in U that can be classifie in attribute set Q using the information in attribute set P. 6. Builing the Classifier The goal of classification is to assign a new obect to a class from a given set of classes base on the attribute values of this obect. To classify obects, which has never been seen before, rules generate from a training set will be use. These rules represent the actual classifier. This classifier is use to preict to which classes new obects are attache. Given a new image, the classification process searches in this set of rules for fining the class that is the closest to be attache with the obect presente for classification [1]. This section escribes how the classification system is built an how a new pattern can be classifie using this system. Given an obect to classify, the features iscusse in section (4) are extracte. The features in the obect woul yiel a list of applicable rules. Then the applicable rules are groupe by class in their consequent part an the groups are orere by the sum of rules confiences, the orere groups woul inicate the most significant class that shoul be attache to the obect to be classifie. Figure () illustrates the rule classification algorithm. We use the alreay generate rules to classify new obects. Given a new image its feature vector is first extracte an then the attribute vector is compute. The nearest matching rule is etermine as the one whose conition part iffers from the attribute vector of re-image by the minimum number of attributes. When there is more than one matching rule, we use a voting mechanism to choose the ecision value. Every matche rule contributes votes to its ecision value, which are equal to the t times number of obects matche by the rule. The votes are ae an the ecision with the largest number of votes is chosen as the correct class. Here t represents the valiity of the rule. Algorithm: Classification of a new obect Input: A new image to be classifie, the attribute vector of the new image, an the set of rules Output: The final classification Processing: Begin For each rule in Rule set Do If match (rule, new obect) Then Measure = Obects, K Classes ; For i=1 to K Do Collect the set of obects efining the concept X i Extract Mrule(X i,u t ) = {r Rule} For any rule r Mrule(X i,u t ) Do T=Match A (r ) I X i an LL=LL U T ; Strength =Car(LL)/Car(X i ) Vote = Measure* Strength Give Vote(Class(Rule),Vote) Return Class with highest Vote En Figure : rule classification algorithm 6

7 6.1 Similarity Measure To select the top N matches to a query image, the atabase is ranke base on the quaratic istances between the query moel an moels of each caniate images. The h h quaratic istance between two m-imensional histograms e an p is efine as: Quaratic T ( e, p) = ( h h ) A( h h ) (8) e p e p Where A is a matrix of similarity weights, A = a ], 0 a < 1, a = 1 an [ i i ii a = 0 1 / max T > T (9) i is the Eucliean istance between colors i an, an max is the greater istance between gray colors. That is, coefficients a i for two gray colors are efine by: m = ( h, s, v ) an m = ( h, s, v ). 0 e e e 7. Results an Discussion 1 p p p 7.1 Data sets The ata sets that we use in this work were taken from the Mammography Image Analysis Society (MIAS) [10]. It contains 30 images, which belong to three normal categories: normal, benign an malign. There are 06 normal images, 63 benign an 51 malign, which are consiere abnormal. In aition, the abnormal case are further ivie in six categories: microcalcification, circumscribe masses, speculate masses, ill-efine masses, architectural istortion an asymmetry. All the images also inclue the locations of any abnormities that may be present. We ivie the 30 samples of mammogram images into 10 equal size folers, such that a single foler is use for testing the moel that has been evelope from the remaining nine sets. The evaluation statistics for each metho is then assesse as an average of 10 experiments. 7. Visual features The query was performe by proviing a query image from a ata set an the selecte five texture features: maximum probability, contrast, inverse ifference moment, angular secon moment an entropy calculate from each occurrence matrix an their values are save in the feature vector of the corresponing image. Then the rules will be generate an orere. The similarity between images is estimate by summing up the istance between corresponing features in their feature vectors. Images having feature vectors closest to feature vector of the query image are returne as best matches. The results were then numerically sorte an the best 1 images were isplaye along with the query image. 7

8 (a) original image (b) enhance result Figure 3: Pre-processing phase on an example image 7.3 Performance Evaluation for the classification To assess the performance of the moel on the test atasets, in this paper we are use the following accuracy (ACC) measure to estimate the classifier accuracy of the propose algorithm. The accuracy is efine as follows: ACC = [(T 1 +T + T 3 )/N] *100% (10) Where T 1, T, an T 3 are the number of correctly classifie normal, benign an malign cases; respectively. N is the total number of test samples. Table () shows the accuracy results of the classification algorithm. Foler Avg(%) Number of rules % Accuracy % Table : Classification accuracy over the 10 foler 8. Conclusion Mammography is one of the best methos in breast cancer analysis, but in some cases, raiologists can not analysis tumors espite their experiences. Such computeraie methos like those presente in this paper coul assist meical staff an improve the accuracy of etection. This paper presents an efficient classification base on texture features to classify from meical atabases in the context of rough set theory. Five features generate form the co-occurrence matrix are extracte an represente in attribute vector, an then the ecision rules within the ata are extracte. Therefore, the classifier moel was built an the quaratic istance similarly is use for matching process. The experimental results show that the algorithm performs well reaching over 85% in accuracy. 8

9 References 1. Aboul Ella H., Rough Sets Analysis for Reuct Generation of Breast Cancer Data International Journal of Artificial Intelligent Tools, IJAIT, Scientific Worl Publisher, 003 (to appear).. Aboul Ella H. an Jafar M. H. Al An Efficient Classification an Image Retrieval Algorithm base on Rough Set Theory 5th International Conference On Enterprise Information Systems École Supérieure ' Électronique e l' Ouest - Angers - France April, pp , Aboul Ella H. Taha E.A. an Tarek A. H., Digital Mammography Image Analysis System Base on Mathematical Morphology IEEE computer society 7th International Conference on Intelligent Engineering Systems INES 003, March 4-6 Assiut-Luxor, Egypt, Ahua N. an Rosefel A., A Note on the Use of Secon-orer Gray-level Statistics for Threshol Selection IEEE Trans. Systems, Man, an Cybernatics, SMC-891, pp , Cancer Net Home Page of the National Cancer Institute Dougherty E. R, Mathematical Morphology in Image Processing Marcel Dekker, New York, Grzymala-Busse J., Pawlak Z., Slowinski R. an Ziarko W., Rough Sets, Communications of the ACM, vol.38, no. 11, Han D.J., Mannila H. an Smyth P., Principles of Data Mining MIT Press, Haralick R.M., Statistical an Structural Approaches to Texture Proceeing of the IEEE, vol.67, no.5, May "The Mammographic Image Analysis Society: Mini Mammography Database Kok S.L., Computer Interpretation of X-ray Mammography Department of Engineering Science, University of Oxfor, web site Lavrac N., Machine Learning for Data Mining in Meicine In AIMDM 99, pp.47-64, Liu S. Babbs C.F. an Delp E.J., Normal Mammogram Analysis an Recognition Proceeings of the IEEE International Conference on Image Processing, October 4-7, Chicago, Illinois, Mari P., Bogan C. Moncef G. an Ari V., Rock Texture Retrieval Using Gray Level Co-Occurrence Matrix. NORSIG-00, 5th NORDIC Signal Processing Symposium, Oct. 4-7, Norway, Martin J. Yaffe Digital Mammography The fifth International Wrkshops on Digital Mammography, IWDM'000, June 11-14, Toronto, Canaa, Nico K. Martin T. Jan H. an Leon V. E., Digital Mammography: Computational Imaging an Vision Kluwer acaemic publication, Pawlak Z., Rough sets Int. J. Computer an Information Science, vol. 11, pp , Slowinski R., Rough Set Approach to Decision Analysis AI Expert, pp. 19-5, March