Medcal X-ray Image Classfcaton Usng Gabor-Based CS-Local Bnary Patterns Fatemeh Ghofran, Mohammad Sadegh Helfroush, Habbollah Danyal, Kamran Kazem Abstract As ntensty of medcal x-ray mages vares consderably and texture characterstcs of them are very smlar, feature extracton based on characterstc of these mages and provdng a set of capable features s a dffcult task. To overcome ths dffculty, many researches focusng on feature extracton schemes for ths type of mages. In ths paper, a novel set of features s proposed that are extracted from transform doman. The proposed scheme nvolves three stages: In the frst stage, we extract edges and shape nformaton from orgnal mages. In order to obtan local features, each mage s parttoned n to 5 overlappng submages n the second stage. Then Gabor transform of each submage s computed to extract centre symmetrc local bnary patterns. In order to evaluate our proposed scheme for radography mage classfcaton, two varous SVM classfers are used. The proposed scheme s mplemented on a subset of IRMA dataset from 15 dfferent categores. Expermental results Shows that proposed method provde an effcent performance. Keywords- Medcal X-Ray Images; Gabor Transform; Centre Symmetrc Local Bnary Pattern; Support Vector Machnes. N I. INTRODUCTION OWDAYS, medcal mages are explosvely used by physcans for dagnosng. As large mage databases are creatng, the demands for effcent medcal mage classfcaton schemes are ncreasng. There are two commonly used steps n any classfcaton scheme: Feature extracton and classfcaton. So feature extracton based on characterstc of mages and provdng a set of capable features play an mportant role n mprovng the mage retreval schemes. Varous medcal x-ray mage classfcaton schemes and algorthms are presented n the lteratures [1-6]. For example, authors n [1] proposed a new content-based medcal x-ray mage retreval method usng the GMM-KL framework. In ths method varous features ncludng textures, ntenstes and locatons was extracted and fnally, each mage was presented by 37500 feature vectors. These features were employed as nput to GMM-mage modelng phase. Ths method has been mplemented on 1501 mages over 17 categores and a classfcaton rate of 97.5% was obtaned. Fatemeh Ghofran, Mohammad Sadegh Helfroush, Habbollah Danyal and Kamran Kazem are wth Department of Electrcal and Electronc Engneerng, Shraz Unversty of Technology, Shraz, Iran E-mal:{f.ghofran, ms_helfroush, danyal, kazem} @sutech.ac.r Although, ths method was very effcent, the hgh dmensonal features were extracted that make classfcaton method very complex. In [] an mage classfcaton method has been proposed. In ths method, multlevel features and Support Vector Machnes (SVM) were employed to mage ndexng. Ths method was evaluated on 9000 tranng mages and 1000 test mages. The accuracy rate of ths method for 57 classes was 89%. In ths method a large number of tranng mages was used and manual labelng of them s very tme consumng. A new content based medcal mage retreval scheme has been presented n [3]. In ths scheme, class label of each mage was determned based on mage modalty, body orentaton, anatomc regon and bologcal system. Ths scheme was mplemented on a dataset comprsng 1617 tranng mages and 33 test mages from only sx categores and the accuracy rate n the sx classes was 9% and so ts capablty for large databases stll remans as a challenge. Medcal x-ray mages are characterzed wth ntensty varatons and poor contrast, thus, t s dffcult to extract a set of approprate features accordng to ther characterstcs. To overcome ths problem, n ths paper a novel feature extracton framework for medcal x-ray mages classfcaton s proposed. In ths framework, after some preprocessng, we extract centre symmetrc local bnary patterns from local parts of shape and drectonal nformaton extracted from mages to acheve a set of capable features. We call these features Gabor-based CS-Local Bnary Patterns (GCS-LBP). Expermental results show effcency of extracted features for medcal x-ray mage classfcaton. The rest of ths paper s organzed as follows. In secton II, some necessary backgrounds are presented. In the thrd secton our methodology s descrbed. Expermental results are presented n secton IV. Fnally, our concluson s summarzed n secton V. II. BACKGROUND REWIEW In ths secton, before we present our classfcaton scheme for radographc mages, some essental background nformaton are brefly provded to the reader. A. Gabor Transform Gabor flters [7] are a group of wavelets, where wavelets are used to capture energy at varous frequences and drectons. 84
Dscrete Gabor transform for a gven mage Ixy (, ) s computed usng the followng expressons: G xy I x s y t st * m, n (, ) = (, ) ϕ (, ), (1) s t * where, s and t are the flter mask sze varables, and ϕ s the complex conjugate of ϕ whch s a class of self-smlar Functons generated from expanson and rotaton of the followng mother wavelet: ϕ 1 1 x y = exp[ ( + )].exp( j πωx), () πσ σ σ σ x y x y where ω s modulaton frequency. The self-smlar Gabor wavelets are obtaned through the generatng functon: ϕ ( xy, ) = a ϕ ( xy, ), (3) where m and n ndcate the scale and drecton of the wavelet respectvely, wth m = 0,1,, M 1, n = 0,1,, N 1, and x = a ( xcosθ + ysn θ), y = a ( xsnθ + ycos θ), (4) nπ where a > 1 and θ =. The varables n the mentoned N equatons are defned as follows: 1 U h M 1 m a= ( ), W, = aul, Ul ( a + 1) ln σ x,, = m, (5) π a ( a 1) U σ y,, l 1 U h 1 = ( ) π π tan( ) ln πσ x,, N The followng values are used for mentoned parameters n our mplementaton: Ul = 0.05, Uh = 0.4, M=3, N=4, s and t range from 0 to 33,.e., sze of flter mask s 33 33. B. Centre Symmetrc Local Bnary Patterns Centre Symmetrc Local Bnary Pattern (CS-LBP) [8] s a modfed verson of orgnal Local Bnary Pattern (LBP) [9] to reduce computaton, whle keepng the characterstcs such as robustness aganst ntensty and llumnaton varaton. In order to generate CS-LBP features, a neghbor set of sze 3 3 for each pxel n an mage s consdered. Then, wth respect to dfference n ntensty between centre symmetrc pars of pxels ( x ), a set of bnary values can be obtaned usng the followng functon: 1, x > 0 sx ( ) = (6) 0, x < 0 For each neghbor set, a unque CS-LBP can be computed usng these bnary values as the followng: I 1 I I 3 I 4 I 6 I 7 I 8 4 I 9 CS-LBP = s(i -I ) 1 9 + s(i -I )1 8 + s(i -I 3 7) 3 +s(i -I ) Fgure 1: CS-LBP features for a neghbor set of sze 3 3. 3 = CS LBP s( x ). (7) = 0 Fnally, a hstogram n the range of [0, 15] s created from all generated CS-LBPs. It s mportant to note that hstogram acheved from orgnal LBPs s n the range of [0, 55]. So, usng CS-LBP representaton leads to reduce extracted features. C. Support Vector Mechnes As we use support vector machnes (SVM) to determne label of each test mage, ths classfer s brefly explaned n ths subsecton. SVMs are bascally b-classfer whch fnd an optmal hyperplane to mnmze the classfcaton error n the tranng step. In the majorty of practcal classfcaton applcatons, patterns are not lnearly separable n the orgnal feature space. In these cases, a feature vectors can be map n to hgh dmensonal space usng a nonlnear functon. Work n new space s possble usng a functon that called and can be acheved from dot product of feature vectors n new space. As SVMs are bnary classfer, several schemes such as oneaganst-all, one-aganst-one and herarchcal are proposed for mult-class problems. Here, we use one-aganst-one method [10] for classfcaton. III. METHODOLOGY The proposed method nvolves three stages: preprocessng, feature extracton and classfcaton process. Next, these stages are descrbed n detal. A. Preprocessng As medcal x-ray mages have gray-level varaton and nterpretaton of them s dffcult, to acheve a set of capable features and also to reduce effects of noses, preprocessng s necessary. In ths paper a set of proposed preprocessng n [11] s appled to each mage as the followng: 6 0 85
Hstogram Adjustng In ths step, to enhance of poor contrast n these mages, ntensty values n any mage are mapped to new values n adjusted mage, such that 1% of data s saturated at low and hgh ntenstes of orgnal mage. Nose Removng Ansotropc dffuson flter proposed by P. Perona and J. Malk n order to reduce effects of noses from mages, s employed. They put forward an ansotropc dffuson (AD) equaton to smooth a nosy mage that s gven by the expresson: uxyt (,,) = dvg ( ( uxyt (,, ) ) uxyt (,, )), (8) t where uxyt (,, ) : Ω [0, + ) Rs a scale mage, g( u) s a decreasng the functon of the gradent. Edges Extracton Usng Canny Edge Detecton The Canny method fnds edges by seekng local maxma of the gradent of any mage. The gradent s obtaned usng the dervatve of a flter. To detect strong and weak edges, the method uses two thresholds. In ths method, the weak edges are acceptable only f they are connected to strong edges. So, ths method s very successful n detectng true weak edges. extractng features to acheve more shape and drectonal nformaton. Fnally, n the last stage, CS-CBP features are extracted from fltered mages. In fact Gabor-based CS-LBP features are the novelty of ths work. C. Medcal X-ray mage classfcaton After the feature extracton, the mages are labeled n to ther respectve classes usng mult-class on-aganst-one algorthm. IV. EXPRIMENTAL RESULTS In ths secton, we carry out experments usng a subset of IRMA database comprsng 1169 x-ray mages (15 classes). Image categores and the number of tranng and test mages per class are shown n Table I. Classfcaton usng a small number of tranng mages s one of the promnent features of ths work. A. Characterstc of medcal x-ray mages Intensty of medcal x-ray mages vares consderably, thus classfcaton of them s a dffcult task. Some mages are captured at dfferent postons and have major varatons ncludng ntensty, contrast, scale, angle and etc. A part from ths there are hgh overlaps between dfferent classes. For example, the content varablty wthn hand class and hgh overlaps between Cranum and Neuro cranum classes and also Ankle jont and knee classes are shown n Fg.. Phase Congruency Computaton Edge and corner phase congruency between orgnal mage and resultng mage of prevous step s calculated. Then resultng mages are multpled. The measure of phase congruency developed by Morrone et al. s: Ex ( ) PC1 ( x) = ( ), (9) An x n where Ex ( ),(local energy), s the magntude of the vector from the orgn to the end pont and An ( x ) s the ampltude of local, complex valued, Fourer components at a locaton x n the sgnal. (a) (b) B. Feature Extracton After preprocessng, Gabor-based CS-LBP features are extracted from local parts of each mage. In ths work three stages are employed for feature extracton. At frst, n order to acheve local features, each mage s parttoned to 5 submages. Also, because of preserve the nformaton of boundares of submages, mage parttonng to overlappng submages s preferable to non-overlappng. As x-ray mages have complex edges and ntensty varatons, extracted features from orgnal mages may not be approprate for classfyng these mages. Thus, n the second stage, we proposed Gabor transform computaton before (c) Fgure.. (a) Sample mages wth content varatons wthn a class. (b) representaton of hgh overlaps between Cranum and Neuro cranum classes. (c) representaton of hgh overlaps between Ankle jont and knee classes. B. Features Reducton Each mage s represented by 4800 features n the mage space. To reduce dmenson of feature vectors and also ncrease ther capablty, Prncpal Component Analyss (PCA) 86
[1] algorthm, a known transform for feature reducton, s utlzed. After feature reducton, 35 features were consdered for descrpton of each mage. C. Proposed Feature Evaluaton The results obtaned from mplementaton of our proposed features wth two varous SVM classfers are shown TABLE II. To verfy capablty of our features, two feature extracton schemes are mplemented and evaluated n ths study. These dfferent schemes are: Scheme A: Extractng CS-LBP from orgnal mages. Scheme B: Extractng CS-LBP from submages wthout applyng Gabor transform and preprocessng. Fg. 3 and Fg. 4 shows results obtaned wth these schemes. Also, the accuracy rates per class for scheme A and scheme B are lsted n Table III and IV respectvely. It can be notced that our proposed scheme for feature extracton provdes the best accuracy rate wth SVMs. The total accuracy rate of scheme B s 4.6% lower than our scheme. Comparng the results acheved wth the dfferent feature extracton schemes, t can be derved that extractng features from Gabor transform of submages and preprocessng stages are very effectve. The lowest accuracy rate was obtaned wth extractng features from orgnal mage. Ths can be attrbuted to the fact that global features are not approprate for descrpton of medcal x-ray mages. accordng to characterstcs of these mages, a set of capable features s ntroduced. We call extracted features GCS-LBP. Expermental results demonstrated that performance of proposed scheme s very effectve. Future works nvolve to applyng proposed features to other mage classfcaton applcaton. TABLE II. FEATURES. CLASSIFICATION ACCURACY RATES OBTAINED BY PROPOSED SVM classfers Class Number Polynomal 1 86% 86% 100% 98% 3 9% 9% 4 100% 100% 5 86% 86% 6 78% 78% 7 80% 80% 8 88% 9% 9 9% 9% 10 88% 88% 11 90% 90% 1 96v 98% 13 9% 9% 14 97.06% 97.06% 15 94.8% 94.8% Accuracy 90.6% 90.88% TABLE I. X-RAY I MAGE CLASSES AND NUMBER OF TRAINING AND TEST IMAGES. Class Number Body part Drecton No. of tranng mages No. of test mages 1 Cranum Coronal 30 50 Lumbar spne Coronal 30 50 3 Hand Coronal 30 50 4 Chest Coronal 30 50 5 Foot Coronal 30 50 6 Ankle jont Coronal 30 50 7 Knee Coronal 30 50 8 Neuro cranum Sagttal 30 50 9 Thoracc spne Sagttal 30 50 10 Chest Sagttal 30 50 11 Ankle jont Sagttal 30 50 1 Rght breast Axal 30 50 13 Left breast Axal 30 50 14 Pelvs (cardovascular system) Coronal 30 34 15 Rado carpal jont Sagttal 30 35 Fgure. 3: Performance comparson between three dfferent scheme usng Polynomal SVM classfer. V. CONCLSION In ths paper, a novel feature extracton scheme for medcal x- ray mage categorzaton s proposed. As x-ray mages have hgh overlaps between dfferent classes, conventonal features may not be approprate for representaton these mages. So, VI. CONCLUSION Fgure. 4: Performance comparson between three dfferent scheme usng SVM classfer. 87
TABLE III. TABLE IV. CLASSIFICATION ACCURACY RATES OBTAINED BY SCHEME A SVM classfers Class Number Polynomal 1 8% 8% 98% 98% 3 4% 4% 4 7% 7% 5 8% 8% 6 4% 58% 7 90% 90% 8 88% 88% 9 66% 66% 10 76% 80% 11 34% 3% 1 60% 60% 13 44% 4% 14 100% 100% 15 77.15% 70.13% Accuracy 67.68% 68.74% CLASSIFICATION ACCURACY RATES OBTAINED BY SCHEME B SVM classfers Class Number Polynomal 1 86% 90% 100% 100% 3 8% 8% 4 86% 86% 5 98% 96% 6 66% 6% 7 78% 80% 8 90% 88% 9 9% 94% 10 88% 90% 11 90% 84% 1 86% 86% 13 8% 8% 14 97% 100% 15 68.6% 74.3% Accuracy 85.97% 86.8% [3] D. Keysers, J. Dahmen, H. Ney, B. B. Wen, Lehmann TM, Statstcal framework for model-based mage retreval n medcal applcatons. J. Electronc Image, vol. 1, no. 1, pp.59 68, 003. [4] V Jacquet, V Jeanne, D Unay, Automatc Detecton of Body Parts n X- Ray Images, Mam, FL, pp.5-30, 009. [5] Bertalya, Prhandoko, Djat Keram, Tb. Maulana Kusuma, Classfcaton of X-Ray Images Usng Grd Approach, Sts, pp.314-319, 008. [6] Quan Zhang Xaoyng Ta, The X-ray Chest Image Retreval Based on Feature Fuson, Shangha, pp. 899 903, 008. [7] D. Zhang, A. Wong, M. Indrawan, G. Lu, Content-based mage retreval usng gabor texture features,, Proceedngs of the Frst IEEE Pacfc-Rm Conference on Multmeda, Sydney, Australa, pp. 39 395, 000. [8] M Hekklä, M Petkänen, C Schamd, Descrpton of nterest regons wth Local Bnary Patterns, Pattern Recognton, vol. 4, pp.45-436, 009. [9] T. Ojala, M. Petkanen, and T. Maenpaa., Multresoluton gray-scale and rotaton nvarant texture classfcaton wth local bnary patterns, IEEE Trans. Pattern Analyss and Machne Intellgence, vol. 4, no. 7, pp. 971 987, 00. [10] M. C. Morrone,J. R. Ross, D. C. Burr, R. A.Owens, Mach bands are phase dependent, Nature 34, pp.50-53, 1986. [11] S. M. Mohammad, M. S. Helfroush, K. Kazem, Fuson of multple sets of novel features for medcal x-ray classfcaton, ICINC Conference, Malaysa, 010. [1] A. K. Jan,B. Bhandrasekaran, Dmensonalty and sample sze consderatons n pattern recognton practce, North Holland Amsterdam, pp. 835 55, 1987. ACKNOWLEDGEMENT The authors would lke to thank TM Lehmann, Dept. of Medcal Informatcs, RWTH Aachen, Germany, for makng the database avalable for the experments. REFERENCES [1] H. Greenspan and A. T. Pnhas, Medcal mage categorzaton and retreval for PACSusng the GMM-KLframework, IEEE Trans. Inf. Techol. Bomed., vol. 11, no., pp. 190 0, Mar. 007. [] A. Mueen, M. Sapyan baba, R. Zanuddn, Multlevel feature extracton and x-ray mage classfcaton, J. Appled Scence, vol. 7, no. 8, pp. 14-19, 007. 88