AUTOMATED DETECTION OF HARD EXUDATES IN FUNDUS IMAGES USING IMPROVED OTSU THRESHOLDING AND SVM Wewe Gao 1 and Jng Zuo 2 1 College of Mechancal Engneerng, Shangha Unversty of Engneerng Scence, Shangha, 201620, Chna 2 Department of Ophthalmology, Jangsu provnce hosptal of TCM, Nanjng, 210029, Chna ABSTRACT One common cause of vsual mparment among people of workng age n the ndustralzed countres s Dabetc Retnopathy (DR). Automatc recognton of hard exudates (EXs) whch s one of DR lesons n fundus mages can contrbute to the dagnoss and screenng of DR.The am of ths paper was to automatcally detect those lesons from fundus mages. At frst,green channel of each orgnal fundus mage was segmented by mproved Otsu thresholdng based on mnmum nner-cluster varance, and canddate regons of EXs were obtaned. Then, we extracted features of canddate regons and selected a subset whch best dscrmnates EXs from the retnal background by means of logstc regresson (LR). The selected features were subsequently used as nputs to a SVM to get a fnal segmentaton result of EXs n the mage. Our database was composed of 120 mages wth varable color, brghtness, and qualty. 70 of them were used to tran the SVM and the remanng 50 to assess the performance of the method. Usng a leson based crteron, we acheved a mean senstvty of 95.05% and a mean postve predctve value of 95.37%. Wth an mage-based crteron, our approach reached a 100% mean senstvty, 90.9% mean specfcty and 96.0% mean accuracy. Furthermore, the average tme cost n processng an mage s 8.31 seconds. These results suggest that the proposed method could be a dagnostc ad for ophthalmologsts n the screenng for DR. KEYWORDS Dabetc retnopathy, Fundus mages, Hard exudates, Improved Otsu thresholdng, SVM, Automated detecton. 1. INTRODUCTION At present, a man challenge of current health care n the world s fast progresson of dabetes. The number of people wth dabetcs would ncrease to 4.4% of the global populaton by 2030 expected by the world health organzaton [1]. However, the fact s that only one half of the patents are aware of the dsease. And n the medcal perspectve, dabetes wll lead to severe late complcatons. These complcatons are macro and mcro vascular changes whch would result n heart dsease, renal problems and retnopathy. Dabetc retnopathy (DR) whch s one of the common complcatons and t remans the most common cause of blndness among adults greater than 65 years n the developed countres [2]. Although early detecton and laser treatment of DR have proved effectve n preventng vsual loss [3], too many dabetc patents are not treated n tme because of nadequaces of the currently avalable screenng programmes [4]. It s now a recognsed urgent worldwde prorty that nsttutes an effcent screenng programme for the detecton of at-rsk patents at a stage when they can stll be effectvely treated because 75% of the blndness due to dabetes can be preventable. Internatonal gudelnes recommend annual DOI:10.5121/jcses.2016.7101 1
fundus examnaton for all dabetc patents to detect the early stage of DR [5]. Now, t s certan that the most effectve treatment for DR s only n the frst stages of the dsease. So, early detecton by regular screenng s of paramount mportance. Dgtal mage capturng technology must be used because t can lower the cost of such screenngs, and ths technology enables us to employ state-of-the-art mage processng technques whch automate the detecton of abnormaltes n fundus. DR clncal sgns nclude red lesons and brght lesons. The former nclude ntraretnal mcrovascular abnormaltes (IRMAs) and hemorrhages (HEs), the latter nclude hard exudates (EXs) and soft exudates or cotton-wool spots (CWs) [6]. EXs are yellowsh ntraretnal deposts and usually located n the posteror pole of the human fundus whch s shown n fgure 1. Hard exudates are made up of serum lpoprotens whch leak from the abnormally permeable blood vessels, especally across the walls of leakng MAs. It can represent the only vsble sgn of DR for some patents [6]. Because dfferent lesons have dfferent dagnostc mportance and management mplcatons, t s very mportant to dstngush among leson types. We dd research on EXs detecton and ther dfferentaton from other brght areas n fundus mages n ths paper. EXs detecton plays a very mportant role on DR screenng tasks, as EXs are among the common early clncal sgns of DR [7]. Addtonally, EXs detecton could act as the frst step for a complete montorng and gradng of DR. Many attempts to detect EXs from fundus mages can be found n the lterature. Some of them use the hgh lumnosty of EXs to dstngush the lesons from background by the method of thresholdng.ward [8] used shade correctng to reduce the shade varaton n the color fundus mage. EXs were detected by thresholdng. It requred the user to select the threshold manually accordng to the hstogram. Phllps [9] detected the large EXs by a global threshold and segmented the smaller, lower ntensty ones by local threshold. The thresholds were selected automatcally, but the regon of nterest must be chosen manually. Snthanayothn[10] appled a recursve regon growng segmentaton to detect EXs after the color mage standardzaton. L[11] modfed the regon growng method by ntroducng the Luv color space. Walter[12] proposed the approach on EXs detecton by ther hgh grey level varaton and ther contours determned by morphologcal reconstructon. S anchez[13] rased the algorthm employng statstcal recognton, Fsher s lnear dscrmnant analyss, colour nformaton and the edge sharpness by applyng a Krsch operator to detect hard exudates and based on the detecton to perform the classfcaton. Jaafar[14] proposed the algorthm that ncluded a coarse segmentaton whch s based on a local varaton operaton to outlne the boundares of all canddates wth clear borders and a followng fne segmentaton whch s based on an adaptve thresholdng as well as a new splt-and-merge technque to segment all brght canddates locally. Other approaches were based on neural network classfers, such as multlayer perceptrons as well as support vector machnes (SVM) [15-18]. In the followng secton 2, the characterstcs of the mage database under study s presented. Secton 3 descrbes the detecton method. We test the performance of the proposed method and compare our method wth other methods n secton 4. In secton 5, dscusson and concluson of ths paper s made. 2
Optc dsc EXs 2. MATERIALS Fgure 1. Color fundus mage and close-up of EXs In ths work, a total of 120 fundus mages whch came from department of Ophthalmology, Jangsu provnce hosptal of TCM were obtaned from a non-mydratc retnal camera wth a 45 feld of vew. The mage resoluton was 800600 at 24 bt RGB. An ophthalmologst manually marked all EXs n these mages. The results obtaned by our automatc method were compared wth these hand-labeled ones. Accordng to the ophthalmologst, these mages n the database were dvded as follows: (1) 68 of these mages belong to patents who suffered from mld to moderate nonprolferatve DR. The mages from DR patents all contaned EXs. (2) The remanng 52 mages belonged to healthy patents. Drusens were not present n any of these mages. The total 120 fundus mages were dvded at random nto two subsets, one s a tranng set the other s a test set: (1) The tranng set contaned 70 mages of whch 40 mages were from DR patents. 2560 segmented regons were extracted from these mages after the segmentaton step of the method named mproved Otsu threshold segmentaton. They were labeled as EXs, or non-exs accordng to the annotaton of the ophthalmologst. Consequently, a fully labeled ground truth database was created whch would be used to determne the parameters of SVM. The test set contaned the remanng 50 mages (22 from healthy retnas and 28 from DR patents). It was used to valdate the effectveness of the complete algorthm by comparng our result wth the expert annotaton. For ths study we have used a Intel(R) Core(TM) Duo E7500 CPU PC wth 6.0 GB RAM and the platform of MATLAB R2009a. 3. METHODS 3.1. Improved Otsu thresholdng segmentaton stage Pgments contaned n the fundus structures have dfferent absorpton characterstcs, and dfferent wavelength monochromatc lghts also have dfferent penetraton performance through fundus, so spectral sgnatures of dfferent layer of fundus structure are dfferent. We found that EXs appear most contrasted n the green channel, and optc dsc appears most contnuous and most contrasted aganst the background n the red channel [19]. So we coarsely segmented EXs n the green channel as shown n Fgure 2(a), and segmented optc dsc n the red channel whch was 3
shown n Fgure 2(b). In color fundus mages, EXs are yellow whte lesons wth relatvely dstnct margns, and n the green channel, they are of hgh lumnosty. Consequently, the most drect and smple approach to dentfy canddate regons of EXs mght be to extract the brght pxels from green channel usng a threshold. But resdual varablty n the lumnosty or n the pgmentaton of the retnal background made the utlzaton of an approach based on a threshold mpractcal.so we opted for an mproved approach whch s called mproved Otsu thresholdng based on mnmum nner-cluster varance [20]. Ths method combnes nner-cluster varance and between-cluster varance of adaptve threshold segmentaton, and when the threshold makes the nner-cluster varance mnmum and between-cluster varance maxmum, t s consdered to be the optmal threshold. Optc dsc s also masked out as shown n Fgure 2(b) usng the method developed by Walter T [12] to reduce the computatonal cost of the classfer because optc dsc has smlar attrbutes n terms of brghtness, color and contrast. Fgure 2(c) s the result after segmentaton of the mage n Fgure 1. Fgure 2. The detecton of EX. (a) green channel of orgnal fundus mage. (b) segmentaton result of optc dsc. (c) canddate regons of EXs. (d) classfcaton result of SVM. (e) detecton result supermposed over the orgnal mage. (f) close-up of detecton result 3.2. Feature extracton stage In order to classfy the canddate regons dentfed n the prevous stage as EXs or non-exs, some sgnfcant features whch help ophthalmologsts to vsually dstngush EXs from other types of lesons as well as the background needed to be extracted from each regon and to be used as nputs of SVM. In ths way, pror knowledge was used to the classfcaton task. We extracted the followng 24 features[16]: (1) Regon sze A (2-4) Mean RGB values nsde the regon µ R, µ G, µ B (5-7) Standard devaton of the RGB values nsde the regon σ R, σ G, σ B N N N (8-10) Mean RGB values around the regon µ R, µ G, µ B 4
N N N (11-13) Standard devaton of the RGB values around the regon σ R, σ G, σ B (14-16) Rato of the mean RGB values between nsde regon area and surroundng area P R, P B (17-19) Homogenety of the regon H R, H G, H B (20-22) RGB values of the regon center C R, (23) Regon compactness CP (24) Regon edge strength ES C G, C B P G, 3.3. Feature selecton stage Many features could be used to desgn and tran the gven classfer. However, msclassfcaton probablty mght be to ncrease wth the number of features and the structure of classfer would much more dffcult to nterpret [21]. Feature selecton tres to avod these problems by pckng out the subset of the extracted features whch s most useful for a specfc problem. Feature analyss for a specfc classfcaton task s based on the dscrmnatory power of the features. Ths s a measure of the usefulness of each feature n predctng the class of an object. Tradtonal feature selecton methods are classfer-drven,.e., whch rely on the results obtaned by a partcular classfer for dfferent subsets of features. In addton, t s more approprate for medcal mage analyss s that classfer ndependent feature analyss whch collects nformaton concernng the structure of the data rather than the requrements for a partcular classfer [21]. Logstc Regresson (LR) s a classfer-ndependent method usually used for feature selecton. LR can be used to descrbe the relatonshp between a response or dependent varable and one or more explanatory or ndependent varables [22]. In our study, there were 24 ndependent varables and the dependent varable was dchotomous. LR could be modelled by functon (1) f he possble values of the dependent varable are 0 and 1 [23]. e Prob( Y = 1) = (1) z 1 + e where Y s the dependent varable, 1 s the desred outcome, z b + b x + b x + Lb p x p s the number of x, the maxmum lkelhood method. T [ b0, b0, Lbp ] z = 0 1 1 2 2 p, B = s regresson coeffcent whch can be dentfed by Model selecton of LR can be carred out by several strateges. A stepwse selecton method was adopted because t can provde a fast and effectve means to screen a large number of varables, and to smultaneously ft a number of LR equatons. Certan crtera are needed to be decded whether an ndependent varable provdes enough valuable nformaton to enter the model or whether, on the other hand, t can be consdered redundant. In our study, an ndependent varable would enter the model f the p-value assocated wth the result of the score test was lower than 0.05 whch was a standard sgnfcance level. Accordngly, a varable would remove from the model f the p-value assocated wth the result of the lkelhood rato test was hgher than 0.10. 3.4. SVM classfcaton stage SVM s a typcally statstcal learnng method based on structural rsk mnmzaton [24], and t has been successfully appled to many pattern recognton problems and the reader s referred to for detals [25]. 5
Consder a labelled two-class tranng set { } y d x,, = 1L,, n, R y +1, 1 s the assocated truth. The separatng hyperplane must satsfy the followng constrant: x, { } y [( ω x ) + b] 1+ δ 0 = 1,2L n, δ 0 (2) Where ω s the weght vector, b s the bas, δ s the slack varable. In order to fnd the optmal separatng hyperplane, functon (3) should be mnmzed subjected to functon (2): n 1 φ ( ω, δ ) = ( ω ω) + C δ (3) 2 = 1 Where C s a parameter chosen by the user and controls the trade-off between maxmzng the margn and mnmzng the tranng error. The classfer can be constructed as: Where ω and s the Lagrange multpler. n { x + b } = sgn y ( x x) f ( x) = sgn ω + b (4) = 1 b denote the optmum values of the weght vector and bas respectvely, and SVM wll map the nput vector x nto a hgh dmensonal feature space by means of choosng a nonlnear mappng kernel f a lnear boundary beng napproprate. The optmal separatng hyperplane n the feature space can be gven by functon (5): n = ( ) + f ( x) sgn y K x x b (5) = 1 Where ( x y) K, s the kernel functon. Gaussan radal bass functon s commonly used. 3.5. Postprocessng stage [ In fundus mages, there may be brght spots or brght ponts of physologcal structures due to reflecton n the human fundus. The bggest dfference between such false postve regons as shown n fgure 3 and EXs s that t usually exsts n solaton, whereas EXs appear n clumps. In order to mprove the performance of ths system, we regarded those mages n whch less than 0.01% of the total number of pxels had been detected as EXs as normal fundus. An example of the false postve regons s shown n Fg. 3. 4. RESULTS 4.1. System performance evaluaton Fgure 3. False postve regons Now, there are no publcly avalable databases whch can be used to test the performance of automatc DR detecton algorthm on fundus mages [26]. The performance of the proposed 6
method was tested usng our test set of 50 fundus mages. Ths evaluaton requres the parameter of the algorthm to be prevously settled usng the tranng set. We vared the parameters and used 10-fold cross valdaton to assess the generalzaton ablty of SVM. For each combnaton of the parameters, we measured the mean senstvty ( SE val ), specfcty ( SP val ) and accuracy ( AC val ) obtaned for the valdaton set, defned as follows: TP SE val = TP + FN (6) TN SP val = TN + FP (7) TP + TN AC val = TP + TN + FP + FN (8) Where TP s the number of classfed EXs correctly, TN represents the number of exactly classfed non-exs, FP s the number of non-exs mstakenly classfed as EXs and FN s the number of EXs classfed as non-exs ncorrectly. Once all the parameters of the algorthm were fxed, we assessed the dagnostc accuracy of t on the test set n terms of a leson-based crteron and an mage-based crteron whch are defned as follows : (1) Usng a leson-based crteron, we measured the mean SE les and postve predct value ( PPV les ) whch s gven by (9) obtaned for the test set. Specfcty s not an nformatve measure for the leson-based crteron. PPV accounts for the probablty that a detected regon s really an EX, and t was regarded as a more sgnfcant crteron to evaluate the system. TP PPV les = (9) TP + FP (2) The mage-based crteron accounted for the ablty of the algorthm to detect pathologcal mages and separate them from normal ones n the test set due to the presence of EXs. Usng the mage-based crteron, each mages n the test set s classfed as belong to a patent wth DR or to healthy subject. We used the mage-based statstcs mean senstvty ( SE m ), mean specfcty ( SP m ) and mean accuracy ( AC m ) to present the fnal results. The statstcs are measured as descrbed n Eqs. (6)-(8). 4.2. Expermental results We coarsely segmented the tranng set of 70 fundus mages wth varable color, brghtness and qualty usng Otsu thresholdng based on mnmum nner-cluster varance. The segmentaton resulted n 2560 canddate regons consstng of 1150 EXs and 1410 non-exs. We computed the 24 features for 2560 canddate regons and used SPSS 18.0 to perform LR on these data. The nputs were normalzed (mean = 0, standard devaton = 1) to nsure that LR was not nfluenced by the numercal strength of a feature rather than by ts dscrmnatory power. The followng 10 features shown n Table 1 were selected and them were wth the hghest dscrmnatory power for separatng EXs from other non-ex yellowsh objects. 7
Table 1. Result of feature selecton Step Parameters Sgnfcance Parameters Sgnfcance 16 A 0.002 CP 0.000 µ 0.000 ES 0.000 R N µ R 0.000 H R 0.000 C G 0.001 P R 0.000 C B 0.000 G 0.000 SVM was used to classfy canddate regons of EXs. At frst, t must be specfed by the obtaned canddate regons wth 10 features. It s that specfyng a SVM requres two parameters: the kernel functon and the regularsaton parameter C. For tranng the SVM classfer, the Kernel-Adatron technque usng a Gaussan Radal Bass kernel was used. To obtan the optmal values for the RBF kernel ( g ) and C, we expermented wth dfferent SVM classfers wth a range of values. We apply genetc algorthm combned wth 10-fold cross-valdaton to fnd the best classfer on the base of valdaton error. Table 2 shows the optmal parameters of ( g ) and C wth the SVM classfcaton structures. The specfed SVM classfcaton result of Fg. 2(c) s shown n Fg. 2(d) whch supermposed over the orgnal mage s shown n Fg. 2(e), and the close-up of the detecton result s shown n Fg. 2(f). Table 2. Valdated parameters Parameters Result C g SE val /% SP val /% AC val /% 2.39 1.25 95.13 98.01 96.72 The proposed method was tested usng a new set of 50 unseen fundus mages: 28 belonged to patents n early stages of DR and 22 corresponded to healthy retnas. The test set contaned 50 mages whch were wth varable color, brghtness and qualty. Usng the specfed SVM to classfy the canddate regons segmented by mproved Otsu thresholdng based on mnmum nner-cluster varance from 50 fundus mages, the performance s shown n Table 3. And t s stated that post processng excluded 2 mages of false postve. In addton, the detecton results of these fundus mages processed by the method proposed by Jaafar [14] are also shown n Table 3. It s found that the automated detecton performance ncludng accuracy and effcency of the method proposed n ths paper s obvously superor to the one proposed by Jaafar. To assess the performance of the SVM we also classfed our segmented canddate regons usng other classfers ncludng RBF classfer and Bayes classfer. In order to estmate the generalsaton error of all classfers, a 10-fold cross-valdaton technque was used agan. The results were summarsed n Table 4 whch are the best results from a selecton of confguratons used for tranng the classfers. These results ndcate that the dagnostc accuracy of SVM s the best n RBF classfer and Bayes classfer. 8
Table 3. Detecton result Method Proposed method Jaafar [14] SE m Image-based crteron SP m AC m Leson-based crteron SE les PPV les Effcency /s 100 90.9 96.0 95.05 95.37 8.31 96.43 90.9 94.0 89.7 90.12 13.58 Table 4. Performances of dfferent classfers Classfer SE m Image-based crteron SP m AC m Leson-based crteron SE les PPV les SVM 100 90.9 96.0 95.05 95.37 RBF 100 90.9 96.0 93.9 95.5 Bayes 96.43 90.9 94.0 90.15 92.06 5. CONCLUSIONS AND DISCUSSION Dgtal magng s becomng avalable as means of screenng for dabetc retnopathy because t can provde a hgh qualty permanent record of the retnal appearance whch can be used for montorng of progresson or response to treatment, and can be revewed by an ophthalmologst. In other words, dgtal mages have the potental to be processed by automatc analyss system. In ths study, we developed a totally automatc method to detect EXs n fundus mages taken from dabetc patents or healthy people wth non-dlated pupls. The fundus mages were segmented by mproved Otsu thresholdng based on mnmum nner-cluster varance to obtan canddate regons of EXs. In order to dstngush EXs from retnal background, the effectveness of a set of 24 features of the canddate regons were examned and the subset wth maxmum dscrmnatory power were selected, accordng to a LR analyss of our data.the optmum subset was used to tran a SVM classfer. The algorthm was tested on 22 mages from healthy retnas and 28 mages of retnas wth DR, wth varable colour, brghtness and qualty. Furthermore, the result was compared wth the method proposed by Jaafar [14]. The SE m n Table 3 shows that we detected all mages wth sgns of DR. We also detected some false postves n the mages of healthy subjects, as the mage-based SP m demonstrates. However, the mpact of wrongly classfyng a person sufferng from DR as a healthy subject s more mportant than the converse. Javtt et al. suggested that a senstvty of 60% or greater maxmzed cost-effectveness n screenng for dabetc retnopathy n ther health polcy model [27]. It means that ncreasng senstvty of screenng from 60% to 100% can not provded addtonal beneft because of the frequency of screenng as well as the lkelhood that retnopathy cases mssed at one vst wll be detected at the next. That means effcency of automated detecton s more mportant when detecton accuracy s adequate. For DR patents, the retnal lesons wll be nfluenced by treatment, control and progresson of dabetes,.e.. Therefore, only hgher detecton frequency s guaranteed, related lesons of fundus can be founded n tme. Hgher detecton frequency s guaranteed by effcent 9
automatc detecton algorthm. The Effcency n Table 3 of the proposed method s obvously superor to the one proposed by Jaafar [14]. However, t may not yet be approprate for clnc. So t s sgnfcant that develop more effcent EXs automatc detecton algorthm n the future. ACKNOWLEDGEMENTS We would lke to thank Department of Ophthalmology, Jangsu provnce hosptal of TCM who suppled all the mages used n ths project for ts great support for the project. REFERENCES [1] Wld S(2004) Global prevalence of dabetes: estmates for the year 2000 and projectons for 2030. Dabetes Care (27): 1047-1053 [2] Wong TY(2006) Dabetc retnopathy n a mult-ethnc cohort n the Unted States. Am J Ophthalmol 141: 446-55. [3] Krstnsson JK(1997) Dabetc retnopathy, screenng and preventon of blndness. A doctoral thess. Acta Ophthalmol Scand Suppl (223):1-76 [4] Fagot-Campagna A(2007) Non-nsuln treated dabetes: Relatonshp between dsease management and qualty of care. The Entred study, 2001 qualty of care. Rev Prat (57): 2209-2224 [5] Dabetes care and research n Europe(1990) The SantVncent declaraton. Dabet Med 7:360 [6] Nemejer, M(2007) Automated detecton and dfferentaton of drusen, exudates, and cotton-wool spots n dgtal color fundus photographs for dabetc retnopathy dagnoss. Invest. Ophthalmol. Vs. Sc. 48(5): 2260-2267 [7] Klen, R(1987) The Wsconsn epdemologc study of dabetc retnopathy VII. Dabetc nonprolferatve retnal lesons. Ophthalmology 94:1389-1400 [8] Ward N P(1989) The detecton and measurement of exudates assocated wth dabetc retnopathy. Ophthalmology, 96(1): 80-86 [9] Phllps R P(1993) Automated detecton and quantfcaton of retnal exudates. Graefe s Archve for Clncal and Expermental Ophthalmology 231: 90-94 [10] Snthanayothn C(1999) Image analyss for automatc dagnoss of dabetc retnopathy. PhD thess, Kng s College London. [11] L H(2002) A model based approach for automated feature extracton n color fundus mages. PhD thess, Nanyang Technologcal Unversty [12] Walter T(2002) A Contrbuton of Image Processng to the Dagnoss of Dabetc Retnopathy-Detecton of Exudates n Color Fundus Images of the Human Retna. IEEE Transactons on Medcal Imagng 21:1236-1243 [13] S anchez C I(2008) A novel automatc mage processng algorthm for detecton of hard exudates based on retnal mage analyss. Medcal Engneerng and Physcs, Elsever, 30: 350-357 [14] Jaafar H F(2010) Automated detecton of exudates n retnal mages usng a splt-and-merge algorthm. 18th European Sgnal Processng Conference. Aalborg, Denmark:EUSIPCO, 1622-1626 [15] Gardner GG.(1996) Automatc detecton of dabetc retnopathy usng an artfcal neural network: a screenng tool. Br. J. Ophthalmol. 80: 940-944 [16] Ege BM(2000) Screenng for dabetc retnopathy usng computer based mage analyss and statstcal classfcaton. Comput. Methods Programs Bomed 62 165-175 [17] Osareh, A(2004) Automatc dentfcaton of dabetc retnal exudates and the optc dsc. Ph. D thess, Brstol [18] Mr HS(2011) Assessment of Retnopathy Severty Usng Dgtal Fundus Images. The Frst Mddle East Conference on Bomedcal Engneerng, Sharjah, UAE [19] Osareh A(2003) Automated dentfcaton of dabetc retnal exudates n dgtal colour mages. Br J Ophthalmol 87(10):1220-1223 [20] Zhou YY(2007) Improved Otsu thresholdng based on mnmum nner-cluster varance. J. Huazhong Unv. of Sc. & Tech. (Nature Scence Edton) 35(2): 101-103 [21] Loew MH(2000) Feature Extracton. n Handbook of Medcal Imagng. Bellngham, WA: SPIE Press, 273 341 [22] Hosmer, D W(2000) Appled Logstc Regresson. New York: John Wley, 307:1989. 10
[23] Xu L(2005) Comparsons of logstc regresson and artfcal neural network on power dstrbuton systems fault cause dentfcaton. Proc. 2005 IEEE Md-Summer Workshop on Soft Computng n Industral Applcatons, Washngton, DC 128-131 [24] Burges CJC(1998) A tutoral on support vector machnes for pattern recognton. Data mnng and knowledge dscovery 2(2): 121-167 [25] Zhang YL (2005) Automated defect recognton of C-SAM mages n IC packang usng Support Vector Machnes. Internatonal Journal of Advanced Manufacturng Technology 25: 1191-1196 [26] Sa nchez CI(2007) A novel automated mage processng algorthm for detecton of hard exudates based on retnal mages analyss. Med. Eng. Phys 30(3): 350-357 [27] Javtt JC(1990) Detectng and treatng retnopathy n patents wth type I dabetes melltus. A health polcy model. Ophthalmology 97: 483-494 AUTHOR My name s Gao Wewe, and I am a teacher n Shangha Unversty of Engneerng Scence. I receved my Master degree and Doctor's degree n Nanjng Unversty of Aeronautcs and Astronautcs. I major n the technology of dgtal medcal equpment. My research nterests nclude medcal mage processng, Bomedcal nformaton analyss and processng, pattern recognton and so on. Specfcally, I do research on medcal mage segmentaton technology whch was appled to automated screenng of dabetc retnopathy. 11