From: AAAI Technical Report SS-94-05. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Diagnosis of Diabetic Retinopathy by ComputerVision by Samuel C. Lee, Ph.D School of Electrical Engineering University of Oklahoma Norman, OK 73072 Vivian S. Lee, M.D., Ph.D Department of Radiology Duke University Durham, NC Elisa T. Lee, Ph.D Department of Biostatistics and Epidemiology University of Oklahoma Oklahoma City, OK Over one-third of adult Native Americans in Oklahomahave noninsulin-dependent diabetes mellitus, and hence this population is at significant risk for diabetic eye disease and subsequent visual impairment. One of the leading causes of the visual impairmentis diabetic retinopathy. A large-scale epidemiologic study of diabetes (including diabetic retinopathy through eye examinations and fundus photography) in AmericanIndians in Oklahomawas conducted by Kelly West [1] and Elisa Lee [2] over a period of 30 years (1972-1992) involving 1,012 Indians. 1.n this study 824 fundus photographs were taken from the subjects and manually examined and graded. The long-term objective of this project is to develop a low-cost, real-time computervision expert system to analyze and grade retinal images and to diagnose diabetic retinopathy. In achieving it, we propose to develop an expert system whichwill detect and quantitate retinal lesions and grade the extent of diabetic retinopathy based on fundus photographs (or images). Applications for this system include large-scale retinopathy screening, epidemiologicstudies, clinical trials, and in routine-clinical settings, it mayprovide a useful quantitative index of disease for ophthalmologists and primary care providers. The development of such a system will require achieving the following specific aims: (1) determine the gradability of the image, (2) to apply and develop image processing techniques for image enhancementand restoration such as for images with poor focus due to cataracts or vitreous hemorrhages, (3) to establish standard chromatic characteristics for fundus images and to develop standardization procedures, (4) to developcomputervision methodsfor the identification of essential retinal features, the optic disc, the macula, and the blood vessels, and for the detection and differentiation of specific retinal lesions and vascular abnormalities. Based on the detection of retinal lesions, to use computervision to grade the severity of diabetic retinopathy based on developed classification criteria and to compare computergrading results with grading results by retinal specialists. In addition, the technical feasibility of a low-cost, real-time computervision diagnostic system for diabetic retinopathy will be assessed. Chromaticand geometric analyses of retinal lesions/features and deterministic and statistical pattern recognition techniques will be applied to the development of the expert system. A readily available set of 824 fundus images will be used to develop the expert system as well as to test the reliability of the system. 50 1. Color Analysis of FundusImages Thefirst question whicharose waswhetherchromaticinformation of a fundusimagealone is sufficient to discriminatethe essential retinal features,e.g., optic disc, macula,vessels, etc., andlesions, e.g., hemorrhagesand microaneurysms, hard exudates, cotton-woolspots, etc. It is foundthat (1) the chromaticinformationof fundusimagesis approximatelyrepresented by the u-coordinate of the UCS color coordinatesystemand the retina is roughlyseparatedfrombloodvessels in the coordinate,and(2) arteries and veins cannothe completelydistinguishedin any coordinate,but if the area in the imageis limited, the V-values(indicating intensity) are useful for discriminatingarteries fromveins. Toverify their findings, weconducteda chromaticanalysis of a numberof fundusimagesusing 35mmcolor transparencies. Someof the transparencies were standard imagesused for comparisonin the ModifiedAirlie HouseClassification Systemfor DiabeticRetinopathy.Theslides wereprojected onto a screen and the images captured by a CCDcolor camera(Sony CCD-G5) which was connected to 80386personalcomputerthrough an imagedigitization board (Professional ImageBoardfrom Atronics). Imagesweredigitized at resolution of 512x 256 pixels. Eachpixei in the digitized imageoccupied15 hits which consisted of red (R), green (G), and blue (13) componentsand each assigned correspondingto 32 gray levels. Wesoughtto obtain chromaticsignal informationon the followingretinal features/lesions: (1) optic disc, (2) macula,(3) vessels (arteries and veins), (4) hemorrhages,(5) hard exudates,(6) woolspots, (7) drusen, (8) laser photoeoagulation scars, (9) fibrous proliferation, and (10) background retina. Thedigitized data for sampleareas of each of the abovefeatures wereplotted onto the UCScolor ¯ coordinatesystem(u,v,V) to see if these features could be distinguished. Wefoundthat, except for retinal imagesfromnormalyoungadults or personswith early stage nonproliferative retinopathy, the backgroundretina cannot be separated fromblood vessels by the u-coordinate. In fact, they cannot be separated from each other by any single or combinedcoordinates of the UCScolor coordinate system. Furthermore,becausethe color of the backgroundretina maydiffer substantially fromone fundusimage to another, noneof the essential features or lesions couldbe separated fromthe background retina by a single or multiple color threshold. In addition, wefoundthat there is a considerableamountof overlap amongthe features/lesions in the u-v and u-Vspaces; noneof the essential retinal features and lesions could be distinguished from one another by an ordinary global thresholding method. Wealso foundthat, amongthe three components,u, v, and V, the Vcomponent can best separate the retinal features and lesions. Theresults of using the V intensity values can be enhancedif an appropriate color filter is employed.Since the major color component of the fundusimageis red, a greencolor filter wouldproducean imagewith the highestcontrast in intensity, whichresults in sharper edges than those of the original image. By doing so, we found that not only the vessels could be separated fromthe other features/lesions muchmoreeasily than fromthe unfiltered original image,but also the arteries and veins in the entire imagecouldbe distinguished. Fromthese results, weconcludethat the chromaticcharacteristics of the samefeature/lesion within a givenimage,as well as betweenimages,differed significantly and color analysis alone is not sufficient to distinguish amongretinal features/lesions. However,the usefulness of color filters was encouraging.Morerefined procedureswereattemptedby using color filters and informationabout their geometry(shape, size, orientation, etc.). In other words,the principles of pattern recognitionshould applied. 51 2. Pattern Recognition Based on the optical and special properties of the features/lesions to be recognized, we constructeda filter with four different templatesto seek edgesin all four directions. This multi-template matchedfilter approachcombinesedgedetection and shape detection into a single computationalstep. In the actual implementationof the algorithm, twelve or possibly moretemplates should be considered dependingon howrefined the lines and shapes are desired. Anedge is detected whenthe convolution of the imagedata by the templatewhosedirection is the sameas the edgeline exceedsa certain threshold value. A roundspot of high (low) intensity is detected whenthe convolutionsof the imagedata by all of the templatesare greater (less) than a certain threshold value. For this reason, the multi-template matchedfilter techniquecan be designedto extract any line edges, for example,edgesof the optic disc and vessels, and any high or low intensity spots, such as hemorrhagesand mieroaneurysms (HMAs),and exudates. However,this methodcannot distinguish betweenhard and soft exudates. Anexampleof comparingthe detections of HMAs and exudates in a fundus image by an ophthalmologist and by our pattern recognition methodare shownin Figure 1. It should be noted that while detecting HMAs and exudatesby the matched-filter method,edges of vessels and the optic disc werealso detected. Since our interest is in the detectionsof HMAs andexudates,edgesof other features/lesionswereeliminatedin this display. Wesee that without using a green filter, only five out of eight HMAs and three out of seven exudatesweredetected. Thespots illuminatedin Figure l(b-2) and (e-2) indicated the lesions detected. Applyinga green filter to the feature extraction of the transparencies, we obtained the feature/lesion detection results shownin Figures 2. Figures on the left-hand side are results obtained froma fundusimagetaken withouta green filter whereasthe ones on the right-handside showthe results obtained with a green filter. Figure 2(b-2) and (c-3) showthat (1) the HMAs and exudates, previously were missed, are nowdetected and (2) vessel edges, whichwerebroken, are nowconnected. In fact, all the eight HMAs were correctly detected. Theninth HMA, located on the lower left side of the optic disc and which wasmissed by humaneye, was detected by this method.A comparisonof the vessel and optic disc edgestaken withoutand with a green filter is shownin Figure 3. It is seen that edgesin the imagetaken with the filter are muchclearer and display less disrupted lines. In summary,from the preliminary study welearned that: (1) a greenfilter enhancesthe detectability, (2) the matchedfilter pattern recognition methodprovidesgoodaccuracyin the detection of HMAsand exudates, (3) solely basedon the shape and size of the lesion, this methodcannotdistinguish between hard and soft exudates. 3. Knowledge-Based Expert System Theabovetwo studies indicated that due to various possible sources of noise and artifacts and the inherent close resemblancesand similarities of someretinal features and lesions, withoutthe use of biological/pathologicalinformation,the chromaticand geometricanalysesare not sufficient to distinguish them. Wemust incorporate the available and necessary biological/pathological knowledgenormally exercised by the ophthalmologistin every step of the decision makingprocess. Withthis in mind,we initiated a preliminarystudyto detect/locate the optic disc usinga knowledge-based color analysis/pattern recognition (CAPR)method.Becauseit is required as an essential elementby every existing grading criterion, the optic disc wasselected as our first target. Thetransparencies wehaveare single-field fundusphotographs,each includes both the optic disc and the macula,whichare roughlysymmetricalon 52 the twosides of the center line. Oncewecouldlocate the optic disc, the tasks of locatingthe maculaand the landmarkvessel reference points wouldbe easier. In our studies, weobservedthe followingtwodistinct features of the optic disc with respect to its color and intensity: (1) Thedisc containsan area of distinctly high intensity (very bright and white in color). (2) Theaverageintensity of the optic disc is usuallyhigherthan that of its surroundingarea. It mayhavesurroundingyellow-whiteatrophythat confoundsthe picture. Theshapeof the disc is always roundand its size varies sightly fromindividual to individual. Usingthese facts, a knowledge-based CAPR algorithmdescribed belowfor detecting and locating the optic disc wasdeveloped. . Definethe permissibleregionsof the disc. If a retinal imageincludingthe optic disc is divided into four regions by twomutuallyperpendiculardiagonallines, the tworegions on the right andleft are def’medas the permissiblere~ionsof the disc. . Findthe intensity histogramof the permissibleregions of the contrast-enhanced intensity witha greenfilter. . Findthe areas in these regionswhoseintensity correspondsto the first high-intensitypeak in the histogramand whosesizes are less than the size of the disc. 4. Findthe geometriccenter of eachof these areas. . For each center, create a square area with a size of 2Do,whereDodenotesthe diameter of the disc (see Figure 3Co)). Whenever the square areas exceedthe boundariesof disc permissibleregions, use the disc permissibleregion boundariesas their boundaries. Thesenewlydefined areas are referred to as Oisc-~earchareas. . Applythe multi-templatematchedfilter to each disc-search area to detect all the edge points in the region. . Applythe Houghtransformto the edgepoints to detect circles with their sizes close to that of the disc. J . If such circles are found, proceedwith the followingverification process. Compute the average intensity of the pixels inside the circle and the average intensities of the surroundingcircular areas of the samesize. If the formeris greater than the latter, the circle foundis confirmedto be the optic disc. If, in step 8, noneof the circles wasverified as the disc, go back to step 2, find additional disc search areas fromthe secondhigh-intensity peak in the histogramand repeat steps 3-7, and so on. Whenall the areas of the disc permissibleregions havebeen searchedand yet no disc is found, wethen concludethat the retinal imageunder study doesnot containthe disc or it is out of focus. A preliminary version of the algorithm has been implemented.Theresult is shownin figure 4. Figures1-4 will be presentedat the Poster Presentation. 53