An Intelligent Moile-Bse Automti Dignosti System to Ientify Retinl Diseses using Mthemtil Morphologil Opertions Mohme Omr, Almgir Hossin, Li Zhng n Huert Shum Computtionl Intelligene Group, Deprtment of Computer Siene n Digitl Tehnologies Northumri University, Newstle upon Tyne, NE2 1XE, UK e-mil: {mohme.omr, lmgir.hossin, li.zhng, huert.shum }@northumri..uk Astrt Dieti retinopthy is onsiere in terms of the presene of exutes whih use vision loss in the res ffete. This stuy trgets the evelopment of n intelligent moile-se utomti ignosis integrte with mirosopi lens to ientify retinl iseses t initil stge t ny time or ple. Exute etetion is signifint step in orer otining n erly ignosis of ieti retinopthy, n if they re segmente urtely, lser tretment n e pplie effetively. Consequently, preise segmenttion is the funmentl step in exute extrtion. This pper proposes tehnique for exute segmenttion in olour retinl imges using morphologil opertions. In this metho, fter pre-proessing, the opti is n loo vessels re isolte from the retinl imge. Exutes re then segmente y omintion of morphologil opertions suh s the moifie regionprops funtion n reonstrution tehnique. The propose tehnique is verifie ginst the DIARETDB1 tse n hieves 85.9% sensitivity. The propose tehnique hieves etter exute etetion results in terms of sensitivity thn other reent methos reporte in the literture. In future work, our system will e eploye to moile pltform to llow effiient n instnt ignosis. Keywors Dieti Retinopthy; Moile Tehnology; Exutes; Opti Dis; Morphologil Opertion. the retin, n this region gives most ute. The thir setion onsists of the exutes s expline previously [5]. Morphologil opertions n reonstrution opertions re well-known tehnil tools to ifferentite imges of exutes into ifferent types. Moreover, these opertions re lso ple of istinguishing etween extremely omplex imges with overlpping or extrtion omponents n they hve een use in mny reserh fiels suh s meiine, iomeiine, omputer vision, remote sensing, vieo oing n mny more. Although there hs een extensive reserh on exute segmenttion, it fes vrious hllenges suh s the presene of noise, texture regions, low ontrst, overlpping, n the size n high intensity of the imges [6]. Bloo Vessels Exutes Mul Opti Dis I. INTRODUCTION Dietes is ommon use of eye isese n n hve the mjor effet of linness. Ptients sight n e ffete y trts n gluom use y ietes. Most signifintly, when smll loo vessels in the retin re mge, flui forms on the retin using fetures suh s exutes, onition populrly known s ieti retinopthy (DR) [1]. The mjor symptom of ieti retinopthy is the presene of exutes. These use yellow lesions of ifferent shpes n sizes whih re mjor signs of vision loss in non-prolifertive forms of DR. Aoring to the WHO, ietes ffets more thn 47 million people worlwie, prtiulrly in eveloping ountries []. More thn 75% of ptients who hve h ietes for more thn 20 yers will hve some form of DR. Hene, the etetion of exutes is neessry for the erly ignosis n tretment of DR in orer to preserve vision [4]. The RGB imge of the retin hs three min setions, s isplye in Figure 1. The opti is is the point of exit of the opti nerve n the fove esries the entre of the mul of Fig. 1. Funus imges for norml (left) n norml (right) retin. Computerize sreening systems hve een evelope to lssify norml n norml retinl imges [7]. The utomti etetion of retinl normlities hs shown promising results n n ssist otors in etter eision mking [8]. Therefore, retinl isese etetion se on moile tehnology using mirosopi lens tthe to smrtphone n e operte effiiently for the erly etetion of DR [2, 9]. However, reserh hs shown tht exute segmenttion is hllenging prolem ue to the vriility of illumintion. In prtiulr, the opti is n exutes show gret similrities in terms of high levels of vrition in the intensity n of olour funus imges. Thus, the etetion of exutes is hllenging tsk ut is lso very signifint tsk in DR ignosis. II. RELEVANT WORK Sophrk et l. [10] propose n lgorithm se on the enhnement of ontrst in the histogrm equliztion of the 978-1-4799-699-7/14/$1.00 2014 IEEE
RGB funus imge. Vessels were eliminte y using losing opertor. Afterwrs the opti is is remove y pplying n entropy feture metho. A morphologil reonstrution using iltion opertion ws use to segment exute res. Welfer et l. [5] pplie mthemtil morphology using the pereptully uniform Luv of olour spe for lightness L, where the vritions of intensity in the L n were slightly less thn in the retinl funus imge. They lso pplie morphologil opertions se on top-ht tehniques, lol minim n thresholing tehniques to etet the exute regions. Eghi et l. [11] lso propose metho se on morphologil opertions. The green omponent ws firstly tken from retinl imges in pre-proessing stge. Then, the opti is n loo vessels were eliminte. Finlly, morphologil opertions, for instne top-ht\ottom-ht n reonstrution opertions were omine to segment the exute regions. Youssef et l. [12] propose metho to etet ll res of yellow olour, the highest intensities, high ontrst n the ontours. Firstly, the opti is n vessels were eliminte from the imge otine from ege etetion, whih gve primry estimtion of the exute regions. Then metho of morphologil reonstrution ws use to ientify the exutes in the finl stge. Snhez et l. [1], mske out n lolize the opti is using regionl mxim n Hough trnsform lgorithms. Then the exutes were segmente y pplying morphologil opertions. Kumri et l. Rvishnkr et l. [14] evelope n pproh to extrt the loo vessels of ifferent thikness using opening n losing opertions. Exutes were etete using opening n losing opertions with ifferent sizes of struturing elements (SEs). The loo vessel imge is sutrte from the imge of etete exutes to otin the opti is region. Menwhile the opti is n the exutes hve similr strutures in term of shpes n high gry levels; therefore, the opti is is remove in most methos in orer to reue flse positives. Retinl imge quisition using low-ost funus mer is the most wiely use metho. However, the segmenttion of suh right lesions n susequent follow up opertions re not esy. This is ue to: 1) the presene of ntomil strutures with highly orrelte pixels from the lesion; 2) illumintion vriility; n ) the movement of the eye uring exmintions of the ptient. Therefore, this pper proposes n urte n effiient lgorithm to extrt ll of the exutes in RGB imges. The proposl is se on morphologil n reonstrution opertions n it hs two setions: (1) loo vessel elimintion n exute extrtion; n (2) opti is etetion, whih is use to ifferentite it from the exute re. Although the tehniques previously presente [11-14] hve provie methos to ientify exutes, the uthors i not lerly stte how the opti is is eliminte when in ifferent lotions. In our ontrst, suggeste metho inites how opti iss re isolte in ifferent lotions. III. MATERIALS The RGB retinl imges whih were use in this pper were quire from the DIAREDB1 tse [15], whih ontins 89 RGB imges of pixel size 1500x1152. The RGB funus imges onsist of ifferent lesions, for instne exute regions, mironeursysms n hemorrhges. The tse lso ontins the groun truth imges whih were use to test our lgorithm. IV. METHODOLOGY A omintion of morphologil opertions is use to etet the exutes in this propose metho. At this urrent stge of the work, the exutes re etete using mthemtil morphology n reonstrution tehniques. An importnt step in the extrtion proess is to remove prominent strutures in the retin, suh s loo vessels n the opti is. This stge n e seprte into three steps: pre-proessing, opti is removl n exute extrtion. In pre-proessing, the RGB funus imge is trnsforme into the HSI imge spe. Afterwrs, the tehnique of mein filtering is performe on the intensity n (I-n) in orer to suppress noise [16]. Next, the ontrst limite ptive histogrm equliztion tehnique is use so s to enhne the ontrst in orer to voi over-sturtion in the ientil regions in the imge of the retin [17]. Noise is remove y pplying Gussin funtion tehnique. The results of these pre-proessing opertions re shown in Fig 2. After tht, the RGB imge is trnsforme into the HSI spe n the intensity n is use for further proessing [18]. Exutes n the opti is hve similrities n show high intensity vlues in the intensity n, n s result opti is elimintion is neee. Also it is well known tht the ojet with the lrgest irulr or lrgest ovl shpe in the RGB imge will e the opti is (OD) [10]. A. Opti Dis Isoltion Firstly, the tehnique of losing opertion with struturing element for flt is is use with the preproesse imge I1 [19]. A thresholing tehnique is pplie to the imge otine to onvert it to inry imge. All of the onnete ojets from the new inry imge Ω of Ci re generte y Eqution 1: Ω = Ck, Ci Cj = 0, i, j m, k m i where k is the numer of onnete ojets n m = 1, 2,,,. If the ojet hs irulr shpe n the lrgest numer of pixels mong Ci it ontins the opti is, n for tht reson the opti is hs to e extrte to seprte it from other regions in the RGB imge. If the lrgest onnete ojet is Ri, the omptness C of Ri is mesure using Eqution 2: C A = 4π (2) P 2 j (1)
where A(Ri) is the totl numer of pixels in re i n P (Ri) is the totl numer of pixels roun Ri. The thresholing tehnique is use to mesure omptness y pplying Nillk s metho n the regionprops funtion seprtely. With this tehnique, the opti is is isolte efore exute etetion, sine oth types of ojets re of similr intensity n olour. Figure () isplys the imge otine fter the losing opertion tehnique ws use with the struturing element of flt is in orer to isolte the loo vessels. The lrgest irulr ojet in the RGB imge is the opti is, n for tht reson it is extrte. For thresholing, 1.6 is given s weight using the Nillk s tehnique n the region of the opti is is righter thn other ojets in the RGB imge. The inry imges re otine s shown in Figure () n () fter the Nillk s metho n the regionprops funtion hve een use. Therefore, the lrgest onnete ojet is onsiere to e the opti is, ue to it hving the high vlue of ensity. The results of opti is isoltion in the RGB imges re not lwys preise, n the Hough trnsformtion tehnique is therefore use when the shpe of the opti is is not irulr, for exmple, n ovl. Thus the tehnique of the Hough trnsform is not lwys epenle in eteting the opti is (OD) sine not ll RGB imges inlue n opti is with irulr shpe. Therefore, the opti is is lolize n etete using the moifie regionprops funtion. B. Exute Detetion After tht, exutes hve to e extrte from the imge I2 fter the opti is hs een isolte. The loo vessels re remove fter pplying losing opertion tehnique with struturing element of flt is to the I2 imge, s oth loo vessels n exutes show high vlues of ontrst (see Figure 4()). Fig.. Imges otine uring opti is elimintion (I2).. Closing opertion tehnique. Nillk s tehnique is use for imge thresholing. Lrge irulr ojet is etete using the moifie regionprops funtion. Results otine from the pre-proesse imge fter the opti is is isolte e Fig. 2. Imges from the pre-proessing phse (I1).. RGB imge. HSI spe. Imge of intensity n. Imge fter mein filter is pplie e. Aptive histogrm equlistion tehnique f. Gussin funtion In the exmple shown in Fig. (), fter pplying Nillk s tehnique, the resultnt imge isplys high omptness in the lrge irulr ojet. f We pply the ove losing opertion metho with struturing element of flt is of rius of 10 pixels to isolte the loo vessels with high vlues of ontrst in the RGB imges efore the thresholing tehniques re pplie. The resulting imge I is otine fter pplying the stnr evition n is erive using Eqution : I 1 = ( I ( i) I ( ) 2 ( x) 2 x) N 1 i W ( x ) where x is set of ll pixels in su-winow W(x) with N s I ( ) the numer of pixels in W(x), n x is the men of I(x) (see Figure 4()). The threshole imge is otine using the tringle metho fter the lol ontrst of the imge I is enhne (see Figure 4()). ()
Sine the etetion of exutes n e onfuse y the eges of oth the imges of ertin ojets n the opti is, we use losing n opening opertors to remove the imge orer from the threshole imge in orer to otin losely istriute exutes (see Figure 4()). Afterwrs, floo filling is pplie to ll holes (see Figure 4(e)). Sine in some RGB imges the opti is is lose to the orers while in others this is not the se, we use our moifie regionprops funtion tehnique for the opti is to e isolte, s shown in Figure 4(f). Next, mrker imge msk is pplie s shown in Figure 4(g) for morphologil reonstrution [20]. During this proess, the peks in the mrker imge re ilte n eroe until the ontour of the imge of the mrker fits elow the imge of the msk, s inite in Figure 4(h). The ifferene etween the imge otine from the erlier stge n the imge of the intensity n is tken for the threshole imge. The imge otine from thresholing s shown in Figure 4(i) is superimpose on the originl retinl imge. Figure 4(j) shows the finl output. V. RESULTS AND DISCUSSION In this pper, the propose metho ws teste on the DIARETDB1 tse in orer to evlute the system. The imges use for testing hve suffiient qulity, illumintion n vrition in olour. Figure 4 shows the results of the whole proeure of exute etetion. Also, the results shown in Figure 5 revel the vntges of the propose lgorithm. For exmple, even though the ontrst in the imge n opti is re very wek, the lgorithm still hs the ility to effiiently etet the exutes. The RGB imges of size of 1500 1152 pixels re selete in orer to test our propose system. The lgorithm ws evlute y ompring our extrtion results with the groun truth imge tse. The sensitivity mesurement is use in orer to evlute lssifier performne, whih is the perentge of rel exute pixels etete. Sensitivity is efine using Eqution 4: Sensitivit y = TP TP + FN (4) e g f h where TP is the numer of exute pixels orretly etete (true positive), n the FN is the numer of exute pixels tht re not etete (flse negtive). The results hieve re summrize in Tle I ompring our propose tehnique n other reent methos reporte in the literture. All of the pprohes shown in the tle use morphologil opertions n were teste on RGB imges from the DIARETDB1 tse. TABLE I. COMPARISON OF EXUDATE DETECTION TECHNIQUES USING THE DIARETDB1 DATABASE i Fig. 4. Imges otine uring exute etetion (I).. Closing opertion tehnique. Tehnique of stnr evition. Tringle metho for threshole imge. Borers remove using losing opertor e. Holes re fille f. OD eliminte using moifie regionprops tehniques g. Msk of mrker imge h. Reonstrute morphologil imge i. Threshole imge j. Exutes otine superimpose on originl imge n mrke in green. j Metho Averge sensitivity Sophrk, et l [10] 4.48 % Welfer, et l [5] 70.48 % Eghi, et l [11] 78.28 % Propose metho 85.9 % The verge sensitivity hieve for our propose tehnique is 85.9% s shown in Tle I, wheres the verge sensitivities of the other methos rnge from 4.48% to 78.28%. This shows tht exute etetion using our metho hs signifintly improve upon the results hieve using other methos in terms of sensitivity.
VI. CONCLUSION This pper hs presente n urte n effiient lgorithm to extrt ll of the exutes in RGB imges. At this stge, the intensity n (I-n) of HSI spe ws use. Due to the noise present in RGB imges, severl tehniques for preproessing steps re use for the suppression of noise n the enhnement of strutures in orer to mth regions isplying high vlues of ontrst. The fully utomti tehnique ws pplie to tse of retinl imges ville in the puli omin, without neeing to hnge ny prmeters uring lgorithm exeution. The experimentl results using DIARETDB1 tse imges inite tht the propose tehnique n work with very low intensity vlues of imges n signifintly enhnes the results of exute etetion in terms of sensitivity when ompre with other pprohes ville in the literture. Also, the metho presente for opti is etetion using our moifie regionprops funtion tehnique enhnes the ury of exute etetion. In future work, we will thus fous on eteting soft exutes n soft rusen with very low intensity vlues n susequent feture extrtion n seletion, long with integrtion with neurl network lssifier in orer to hieve etter results for isese etetion. We will lso eploy the overll system to moile pltform to further promote instnt n erly isese ignosis. Furthermore, energy effiieny lgorithms will lso e onsiere for the eployment of our pplition to moile pltform [21]. Fig. 5. Originl imge (); n the exutes otine mrke in green superimpose on the originl low qulity imge (). REFERENCES [1] A. Bhuiyn, B. Nth, n K. Rmmohnro, "Detetion n lssifition of ifurtion n rnh points on retinl vsulr network,"interntionl Conferene on Digitl Imge Computing Tehniques n Applitions (DICTA),,8, -5 De. 2012. [2] D. Ning, n Y. Li, "Automte ientifition of ieti retinopthy stges using support vetor mhine," 2n Chinese Control Conferene (CCC), 882,886, 26-28 July 201. [] Worl Helth Orgnistion "WHO-Blinness-Prevention" [Online]. Aville: http://www.who.intllinness/prevention of Blinness from Dietes MeUitus-with-over-smll.pf.Aesse on 1 st of Mrh 2014. [4] M. Niemeijer, M. D. Aràmoff, n B. Vn Ginneken, "Segmenttion of the opti is, mul n vsulr rh in funus photogrphs." IEEE Trnstions on Meil Imging, 26, no. 1 (2007): 116-127. [5] D. Welfer, J. Shrnski, n D. Rushel Mrinho, "A orse-to-fine strtegy for utomtilly eteting exutes in olor eye funus imges." Computerize Meil Imging n Grphis 4, no. (2010): 228-25. [6] W. Srisukkhm, P. Leph, M. A. Hossin, L. Zhng, R. Jing, H. M. Lim, "A moile enle intelligent sheme to ientify loo ner for remote res - ell memrne segmenttion using mrker ontrolle wtershe segmenttion phse." 7th Interntionl Conferene on Softwre, Knowlege, Informtion Mngement n Applitions, (SKIMA 201), Ching Mi, Thiln, 104-114. [7] C. Snygm, E. S. Ti, A. Shnkr, J. Lee, C. Sun, n T. Y. Wong, "Retinl rteriolr nrrowing inreses the likelihoo of hroni kiney isese in hypertension." Journl of Hypertension 27, no. 11 (2009): 2209-2217. [8] H. Rui, C. M. Queen, n G. Zourikis, "Detetion of Buruli uler isese: preliminry results with ermosopi imges on smrt hnhel evies." Point-of-Cre Helthre Tehnologies (PHT), IEEE, 168-171. IEEE, 201. [9] C.-N. Lee, Y.-T. Chu, L. Cheng, C.-C. Wu, n C.-Y. Yng, "Usge of smrt moile evie t the telemeiine." 2011 Interntionl Conferene on Mhine Lerning n Cyernetis (ICMLC),2, 582-587. IEEE, 2011. [10] A. Sophrk, B. Uyynonvr, S. Brmn, n T. H. Willimson, "Automti etetion of ieti retinopthy exutes from non-ilte retinl imges using mthemtil morphology methos." Computerize Meil Imging n Grphis 2, no. 8 (2008): 720-727. [11] M. G. F. Eghi, n H. Pourrez, "Loliztion of hr exutes in retinl funus imge y mthemtil morphology opertions." 2n Interntionl econferene on Computer n Knowlege Engineering (ICCKE), 185-189. IEEE, 2012. [12] D. Youssef, N. Soloum, A. El-i, M. Mrouk, n A.-B. Youssef, "New feture-se etetion of loo vessels n exutes in olor funus imges." 2n Interntionl Conferene on Imge Proessing Theory Tools n Applitions (IPTA), 294-299. IEEE, 2010. [1] C. I. Sánhez, M. Niemeijer, M. S. A. Suttorp Shulten, M. Aràmoff, n B. Vn Ginneken, "Improving hr exute etetion in retinl imges through omintion of lol n ontextul informtion." IEEE Interntionl Symposium on Biomeil Imging: From Nno to Mro, 2010, 5-8. IEEE, 2010. [14] S. Rvishnkr, A. Jin, n A. Mittl, "Automte feture extrtion for erly etetion of ieti retinopthy in funus imges." IEEE Conferene on Computer Vision n Pttern Reognition (CVPR), 2009, 210-217. IEEE, 2009. [15] T. Kuppi, V. Klesnykiene, J.-K. Joni-Kristin Kmrinen, L. Lensu, I. Sorri, A. Rninen, R. Voutilinen, H. Uusitlo, H. Kälviäinen, n J. Pietilä, "The DIARETDB1 ieti retinopthy tse n evlution protool." BMVC, 1-10. 2007. [16] M. Kirhner, n J. Fririh, "On etetion of mein filtering in igitl imges." IS&T/SPIE Eletroni Imging, Interntionl Soiety for Optis n Photonis, 2010. [17] K. Vij, n Y. Singh, "Enhnement of imges using histogrm proessing tehniques", Interntionl Journl Computtionl Tehnology Applitions, (2011), no 2, 09 1. [18] A. Sophrk, B. Uyynonvr, n S. Brmn, "Automti exute etetion from non-ilte ieti retinopthy retinl imges using fuzzy -mens lustering." Sensors 9, no. (2009): 2148-2161. [19] P. Mrgos, n R. W. Shfer, "Morphologil systems for multiimensionl signl proessing." Proeeings of the IEEE 78, no. 4 (1990): 690-710. [20] R. C. Gonzlez, R. E. Woos, S. L. Eins "Morphologil reonstrution from igitl imge proessing using MATLAB", MATLAB, Digitl Imge Proessing Thir Eition, Lirry of Congress Ctloging-in-Pulition Dt on File, 2008. [21] A. Bourouis, M. Fehm, M. A. L. Hossin, Zhng, "An intelligent moile se eision support system for retinl isese ignosis. eision support systems". Elsevier, Mrh 2014, 41 50.