Automated keratoconus screening with corneal topography analysis.

Automated Keratoconus Screening With CornealTopography Analysis

Naoyuki Maeda, Stephen D. Klyce, Michael K. Smolek,and Hilary W. Thompson

Purpose. Although visual inspection of corneal topography maps by trained experts can bepowerful, this method is inherently subjective. Quantitative classification methods that candetect and classify abnormal topographic patterns would be useful. An automated system wasdeveloped to differentiate keratoconus patterns from other conditions using computer-as-sisted videokerat oscopy.

Methods. This system combined a classification tree with a linear discriminant function derivedfrom discriminant analysis of eight indices obtained from TMS-1 videokeratoscope data. Onehundred corneas with a variety of diagnoses (keratoconus, normal, keratoplasty, epikeratopha-kia, excimer laser photorefractive keratectomy, radial keratotomy, contact lens-induced war-page, and others) were used for training, and a validation set of 100 additional corneas wasused to evaluate the results.

Results. In the training set, all 22 cases of clinically diagnosed keratoconus were detected withthree false-positive cases (sensitivity 100%, specificity 96%, and accuracy 97%). With the valida-tion set, 25 out of 28 keratoconus cases were detected with one false-positive case, which was atransplanted cornea (sensitivity 89%, specificity 99%, and accuracy 96%).

Conclusions. This system can be used as a screening procedure to distinguish clinical kerato-conus from other corneal topographies. This quantitative classification method may also aid inrefining the clinical interpretation of topographic maps. Invest Ophthalmol Vis Sci.1994; 35:2749-2757.

JVeratoconus is one of the noninflammatory cornealthinning disorders characterized by anterior protru-sion of the cornea and stromal thinning.1 Detection ofkeratoconus is important to avoid unpredictable re-sults in refractive surgery and difficulties in contactlens fitting. Additionally, screening for the early formof keratoconus will lead to a better understanding ofthe natural course of keratoconus, as well as the ge-netic basis of this disease.

Although advanced keratoconus is easily diag-nosed by slit lamp findings and keratometry readings,it is difficult to detect early keratoconus with theseinstruments.'2 Photokeratoscopy1:< and computer-as-

t'nmi the Lmm Eye Research Laboratories, I.SU Eye Center, Louisiana StateUniversity Medical Canter School «j Medicine, New (Means, Louisiana.Supported in part by National Institutes of Health grants EY03311 and EYO2377and by Computed Anatomy, Inc. and Menicon, Co., Ltd.Presented in part at the annual meeting of the Association for Research in Visionand Ophthalmology, Sarasota, Florida, May 1993.Submitted for publication September 7, 1993; revised December 17, 1993; acceptedJanuary 5, 1994.Proprietary interest category: Cl 567/i.Reprint requests: Stephen D. Klyce, LSU Eye Center, 2020 Cravier Street. Suite H,New Orleans, LA 70112.

sisted videokeratography2''1' allow early keratoconusto be detected by the trained observer. In particular,computer-assisted videokeratoscopy has been used foranalyzing irregular astigmatism seen in the familymembers of patients with keratoconus,6' following upthe progression of subclinical keratoconus,8 and as-sisting in contact lens selection for keratoconus.9

Interpretation of videokeratographs requires ex-aminers to have prior training in discerning the com-plex patterns or subtle features contained in the con-tour map. Topographic maps of eyes with keratoconusdisplay a variety of different patterns5 that may be mis-interpreted by an untrained examiner. Therefore, au-tomatic quantitative image analysis of color-codedmaps and objective criteria for interpreting a true ker-atoconus pattern would be useful.

Although numerical methods have been devel-oped to distinguish the keratoconus cornea from thenormal cornea,4 there have been no reports of amethod to differentiate between keratoconus and anumber of other clinical entities whose topographicpresentation may share features similar to those of

lnvcsii^iiiivi- Ophthalmology & Visual S( ii'iicc. May 1904, Vol. ;$">. No. 6Copyright © Association lor Research in Vision and Ophthalmology 2749

Downloaded From: http://iovs.arvojournals.org/pdfaccess.ashx?url=/data/journals/iovs/933404/ on 04/12/2018

2750 Investigative Ophthalmology & Visual Science, May 1994, Vol. 35, No. 6

keraioconus. Therefore, we used eight quantitative in-dices derived from videokeratoscopy data (five new,three preexisting) as input to an automated keraio-conus pattern detection algorithm that uses discrimi-nant analysis embedded in an artificial intelligencemesh.

METHODS

Subjects

Color-coded maps and data files generated by a com-puter-assisted videokeratoscope (TMS-1, ComputedAnatomy, NY) were used. Institutional Review Boardapproval was obtained, all subjects signed informedconsent forms, and the tenets of the Declaration ofHelsinki were followed in taking videokeratographs.

Maps were categorized by three corneal topogra-phy researchers based on clinical records and topo-graphic appearances with the fixed Klyce/Wilson scale(28.0 to 65.5 D in 1.5 D steps). Videokeratographswere drawn from the Louisiana State University EyeCenter patient population and were divided randomlyby category into two sets. Corneas with mixed diag-noses, as indicated in the medical records, and poorlycentered videokeratographs were excluded. The kera-toconus detection program was developed using atraining set of 100 corneas and evaluated with a valida-tion set of an additional 100 corneas (Table 1). Eachset. comprised eight categories: normal, keratoconus,keratoplasty, epikeratophakia, excimer laser photore-fractive keratectomy, radial keratotomy, contact lens-induced warpage, and other. The distribution of diag-noses was similar in the two sets (Table 1). The normalcategory included corneas with regular astigmatismranging from 1.5 D to 3.75 D (seven maps in the train-ing set and nine maps in the validation set). The kera-toconus category encompassed a range of severity

TABLE l. Subjects in Each Set

KeraioconusMildMode raleAdvanced

Normal corneaRegular astigmatism

KeratoplastyEpikeratophakiaPhoiorcfractivc keratectomyRadial keratotomyContact lens-induced warpageOther

Total

Training Set

22(7)(7)(8)30(7)157554

12

100 Eyes

Validation Set

28(8)

(10)(10)30(0)14

74458

100 Lyes

from mild to advanced clinically diagnosed kerato-conus; 1 1 of the 50 eyes had centrally located cones.Corneas with localized areas of steepness on videoker-atography without conventional clinical findings (ker-atoconus suspect) were excluded. The categorymarked other consisted of 12 eyes in the training set(four relaxing incisions, four scarred corneas, twopost-retinal detachment surgery, one post-cataractsurgery, and one keratomileusis) and eight eyes in thevalidation set (seven scarred corneas and one ptery-gium).

Quantitative Indices of the Topographic Map

The statistical indices calculated for every topographicmap included three existing indices—Simulated Kl(SimKl), Simulated K2 (SimK2), and Surface Asym-metry Index (SAI)10"—and five new indices—theDifferential Sector Index (DSI), the Opposite Sectorlndex (OSI), the Center/Surround Index (CSI), theIrregular Astigmatism Index (I A I), and the AnalyzedArea (AA). Except for AA, all the statistical and nu-merical indices used in the study were derived fromthe spatial distribution of dioptric power in the video-keratographs.

For the calculation of DSI, OSI, CSI, and IAI,area-corrected power was used in place of the video-keratoscope-derived power. In videokeratography,more samples are commonly collected per unit areafrom the central cornea than from the periphery. Toequalize or unbias the calculation of the indices, eachpower was multiplied by the area of the cornea fromwhich it was derived, and the sum total was divided bythe total corneal area analyzed.

Keratoconus is typified in videokeratographs as anarea of significant localized steepening. To detect suchfeatures in videokeratographs, DSI, OSI, and CSIwere developed. The corneal surface was divided intoeight pie-shaped sectors, each subtending 45°. Thisreference pattern was rotated up to 45° in 32 incre-ments relative to the central axis of the contour map(Fig. la) to find the corneal sector with the greatestpower. Average power in each sector was calculatedfrom the area-corrected refractive power.

DSI reports the greatest difference in averagepower between any two sectors, and OSI representsthe greatest difference of the average power in oppo-site sectors (Fig. lb). CSI is the difference in the aver-age area-corrected corneal power between the centralarea and an annulus surrounding the central area (Fig.1c). These three indices help to differentiate amongnormal corneas, regular astigmatism, peripheral steep-ening keratoconus, and central steepening kerato-conus (Fig. 2).

The IAI describes the short-range seiniineridionalfluctuation of power distribution. IAI is the average


Keratoconus Screening 2751

rotate

OSI - 8DDSI = 10D

CSI-5D

FIGURE 1. Calculation method of Differential Sector Index(DSI), Opposite Sector Index (OSI), and Center/SurroundIndex (CSI) (a) The corneal surface was divided into a pat-tern of eight pie-shaped sectors, each subtending an angle of45°. The average power was calculated for each sector basedon area-corrected refractive power. New sector regionswere defined by rotating the pattern relative to the contourmap in 1.41° steps (360°/256 values per mire), up to a maxi-mum of 45°, and calculating average power in each sector ateach step, (b) DSI reports the greatest difference of theaverage power between any two sectors obtained during therotation (53 D — 43 D = 10 D). OSI represents the maxi-mum difference between average powers in opposite sectorsobtained during the rotation (53 D - 45 D = 8 D). (c) CSI isthe difference in the average area-corrected power betweenthe central area (3.0 mm diameter) and an annulussurrounding the central area (3.0 to 6.0 mm) (48 D — 43 D =5D). '

summation of inter-ring area-corrected power varia-tions along every semimeridian for the entire analyzedsurface and is normalized by the average cornealpower and number of data points (Fig. 3).

The last index used was the AA, the ratio of the

interpolated data area to the area circumscribed by thelast mire found in a videokeratoscope image.

Discriminant Analysis Classifier

Two-group discriminant analysis was used as a multi-variate statistical technique for the eight indices toscreen for keratoconus patterns. A linear discriminantfunction of the multiple independent variables isfound that allows one to discriminate between the twoclassifications (for example, keratoconus and nonker-atoconus).12 The linear discriminant function was ob-tained by discriminant analysis of the training set witha commercially available statistics software package(Statistical Analysis Systems Ver 6.06, SAS Institute,Cary, NC). The linear discriminant function yields asingle composite discriminant value for each map,which was designated the Keratoconus Prediction In-dex (KPI). The division between keratoconus and non-keratoconus patterns is the cutoff value (Fig. 4). Mapsthat had a KPI value greater than the optimum cutoffvalue were classified as keratoconus, whereas mapswith a KPI value less than the optimum cutoff valuewere classified as nonkeratoconus.

The efficacy of the discriminant analysis classifierwas tested using the validation set, and the results, weredescribed in terms of sensitivity (True positive/[Truepositive + False negative]), specificity (True negative/[True negative + False positive]), and accuracy.^[Truepositive + True negative]/Total number of maps).

Expert System Classifier

An expert system is a form of artificial intelligence thatcomprises an extensive set of decision rules. Decisionsby the expert system are made deductively with step-by-step logical operation.13 The discriminant analysisClassifier was embedded into a rule-based expert sys-tem to enhance keratoconus pattern screening abilityand to differentiate between peripheral and central

NormalRegular

astigmatismPeripheral keratoconus Central keratoconus

Differential Sector Index

Opposite Sector Index

Center/ Surround Index

low

low

low

high

low

low

high

high

low - middle

low - middle

low - middle

high

FIGURE 2. The principal differences in pattern of three indices seen in normal, regular astig-matism, and two types of keratoconus. The difference pattern of these three indices is helpfulto differentiate normal, regular astigmatism, and keratoconus.



IAI = B* In

j= 2. 30

h. 1,256

1=2,30 -D

b 1,256

i : semimeridional positionj : ring numberPs I: corneal power on the point (i, j)AA : area which corresponds to Power P^B : normalization by powerC : normalization by number of pointsD: scaling constant

low IAI high IAI

FIGURE 3. Definition of the Irregular Astigmatism Index (IAI). IAI was acquired by calculat-ing the average summation of inter-ring area-corrected power variations along ever)7 semi-meridian for the entire analyzed surface and normalized by the average corneal power andnumbers of all measuring points.

keratoconus patterns. The flow chart of the expertsystem used in this study is shown in Figure 5 and wasimplemented with the Pascal program language. Inaddition to KPI, which was obtained from discrimi-nant analysis, four indices (DSI, OSI, CSI, and SimK2)were used in the binary decision tree. Maps were firstclassified as either keratoconus, borderline, or non-keratoconus using KPI and SimK2 values. The border-

unweighted cutoff valuecutoff value

optimum cutoff value (0.30)

KPI

KPI < cutoff value

Non-keratoconuspattern

cutoff value < KPI

Keratoconuspattern

FIGURE 4. The cutoff value in Keratoconus Prediction Index(KPI). The cutoff value is the score against which each KPI isjudged to determine into which group the individual mapshould be classified. The optimum cutoff value is that whichpermits the least error in classifying the category and wasused as the cutoff value in the discriminant analysis. In theexpert system classifier, the efficacy of the expert system wasevaluated for all cutoff values between the unweighted andthe optimum cutoff values.

line maps were then divided into keratoconus or non-keratoconus by DSI, OSI, and CSI. Next, all kerato-conus patterns were classified into either peripheral orcentral keratoconus using a threshold combination ofthese indices. Thresholds for DSI, OSI, CSI andSimK2 were developed during the pilot study (unpub-lished data, 1992). Final output of the system was thedisplay of the certainty of keratoconus relative to theKPI value.

The most adequate cutoff value was determinedfrom the training set by calculating sensitivity, specific-ity, and accuracy for several values between the opti-mum cutoff value and the unweighted cutoff value(Fig. 4). Efficacy of the expert system classifier wasevaluated with sensitivity, specificity, and accuracy ob-tained for the validation set.

RESULTS

Table 2 shows the mean and standard deviation foreach of the eight indices used in the keratoconus andnonkeratoconus groups in the training set. Using theunpaired Student's /-test, it was possible to show signif-icant differences in six quantitative indices, but reli-able case-by-case classification of maps by any singleindex was not possible because of the large, overlap-ping ranges. For example, the Opposite Sector Indexis similar to the I-S value reported previously.4 If weset the OSI to distinguish keratoconus at a 95%



Input

Yes

non-keratoconus keratoconus

non-keratoconus pattern central steepenineg keratoconus peripheral steepening keratoconus

FIGURE 5. Binary decision tree in the expert system. Four indices (DSI, OSI, CSI, SimK2)combined with KIM were used in this decision tree. All cases were finally divided into threegroups (nonkeratoconus pattern, peripheral keratoconus pattern, and central keratoconuspattern).

correct level, then there would he 24% false-positivecases in the training set.

Discriminant analysis was performed on the train-ing set and the function obtained was:

KPI = 0.30 + 0.01 (-41.23 - 0.15 DSI

+ 1.18 OSI + 1.49 CSI + 4.13 SAl - 0.56 SimKl

+ 1.08SimK2 - 3.74 IAI + 0.10 AA).

In this equation, KPI has a value greater than theoptimum cutoff value (0.3) when a color-coded maphas a keratoconus pattern, and it has a value less than0.3 when a color-coded map shows nonkeratoconusfeatures. Table 3 shows the mean and standard devia-tion of KPT for each group in the training and thevalidation sets. The KPI for keratoconus in each set.was significantly higher than the KPIs for any othercategories with the unpaired Student's /-test.

TABLE 2. Results of All Indices in the Training Set

SimKlSimK2Surface Asymmetry Index (SAI)Differential Sector Index (DSI)Opposite Sector Index (OSI)Center/Surrounding Index (CSI)Irregular Astigmatism Index (IAI)Analyzed Area (AA)

Keratoconus(n = 22)(Mean ±

53.62 ±47.54 ±

3.18 ±11.16 +9.37 ±0.35 ±0.59 ±

71.79 ±

SD)

5.723.932.346.505.473.000.24

21.85

Nonkeratoconus(n = 78)(Mean ±

45.89 ±42.80 ±

0.60 ±3.99 ±1.84 ±

-0.28 ±0.47 ±

78.15 ±

SD)

4.272.870.623.712.020.900.15

15.16

P Value*

P= 0.0001P= 0.0001P= 0.0001P= 0.0001P= 0.0001P = 0.111 1P= 0.0039P= 0.1203

Unpaired Student's /-test.



TABLE 3. Keratoconus Prediction Index (KP1)

KeratoconusNormal corneaKeratoplastyKpikeralophakiaPhotorcfractivc

keratectoniyRadial keratotomyContact lens-induced

warpageOther

Training Set(Mean ± SD)

0.38 ± 0.070.20 ± 0 . 0 I t0.20 ± 0.04f0.10 ± O.OSf

0.18 ± 0.02f0.17 ± 0.04f

0.21 ± 0 . 0 If0.20 ± 0.02t

Validation Set(Mean ± SD)

0.34 ± 0.090.20 ± 0.02+0.17 ±0.06+0.17 ± 0.1 It

0.20 ± 0.03*0.1(5 ± 0.03t

0.10 ±0.02+0.10±0.03t

Unpaired /-lesi with kcratocoiiiii. */J < 0.05; \P < 0.01.

Table 4 shows the classification determined by dis-criminant analysis, compared with the actual diagno-sis. Nineteen out of 22 cases of clinically diagnosedkeratoconus were detected, with no false-positivecases in the training set. In the validation set, 19 out of28 keratoconus cases were detected with one false-po-sitive case. Sensitivity, specificity, and accuracy were86%, 100%, and 07% in the training set, and 68%,99%, and 90% in the validation set, respectively.

In Table 5, classification predicted by the expertsystem versus actual classification is shown. Sensitivity,specificity, and accuracy in the training set and in thevalidation set are shown in Figures 6 and 7, respec-tively. The optimum cutoff value in the expert systemwas found to be 0.28 for the most sensitive and accu-rate results in the training set. Here, sensitivity, speci-ficity, and accuracy were 100%, 96%, and 97% in thetraining set, and 89%, 99%, and 96% in the validationset, respectively (Table 6).

Using the expert system classifier at a 0.23 KPIcutoff"value, all false-positive maps (three in the train-ing set. and one in the validation set) were post-kerato-plasty eyes. It is important to note that localized ab-normal steepening was seen in these maps, and with-out referring to the patient's history, interpretation ofthe color-coded map alone would have been similarly

inconclusive, even to a well-trained observer. Threefalse-negative maps in the validation set had typicalkeratoconus in the contralateral eye. Maps of thesefalse-negative cases showed either a mild inferiorsteepening, a pellucid marginal degeneration-like pat-tern, or with-the-rule astigmatism with some irregular-ity. Thus, it is not surprising that these cases were in-correctly interpreted.

In the validation set, the expert system classifierdetected six more keratoconus maps compared to thediscriminant analysis classifier alone without increas-ing false-positive maps. The sensitivity with the expertsystem classifier (89%) was significantly better than thesensitivity with discriminant analysis alone (68%)(McNemar test with correction for small expected fre-(juencies"1; P = 0.016).

DISCUSSION

Computer-assisted videokeratography and the color-coded map provide an abundance of informationabout corneal surface characteristics. However, hu-man visual interpretation is essentially subjective,whereas contour information is difficult to analyzequantitatively. An objective assessment of videokera-tography is essential for statistical studies of the pro-gression of keratoconus, genetic studies, or screeningprocedures used for refractive surgery practice. There-fore, the thousands of data points in a color-codedmap must be reduced in some fashion to a series ofstatistically manageable indices.

Rabinowitz and McDonnell4 reported the first nu-merical method to differentiate between keratoconuspatterns and normals based on videokeratoscopy.They used central corneal power, difference in centralcorneal power between fellow eyes, and the Inferior-Superior (I-S) value. The I-S value was defined as anaverage refractive power difference between five infe-rior points and five superior points 3 mm from thecenter at 30° intervals. These three parameters weresignificantly different in patients with keratoconusthan in normal controls.

TABLE 4. Classification by Discriminant Analysis Alone

Predicted Category

Actual Category

Training setKeratoconusNonkeraloconus

Validation setKeraioconusNonkeratoconus

Kemtoconus

100

101

Nonkeraloconus

:\78

071

Sensitivity Specificity Accuracy

86%

68%

100%

00%

07%

00%



TABLE 5. Results of Categorization WithExpert System

Training Set Validation SetLutofl Value(KPI)

0.30*

0.29

0.28

0.27

0.26

0.25

0.24

0.23

0.22

0.21

Group

KCNonKCKCNonKCKCNonKCKCNonKCKCNonKCKCNonKCKCNonKCKCNonKCKCNonKCKCNonKC

KC

190

190

191

201

202

212

22• \

223

224

225

NonKC

378

378

377

277

276

1760

7f>0

750

740

73

KC

191

190

210

211

211

231

231

251

251

251

NonKC

971

972

772

771

771

571

571

371

371

371

KC, kcr;it(u:<>nus: NonKC, nnnkuruiorotuis.* Discriminant analvsis onlv.

However, central corneal power by itself is an in-adequate criterion; we have observed families of em-metropes whose central corneal power averages 48 to50 D. Furthermore, other corneal pathologies (cor-neal grafts, refractive surgeries, extracapsular cataractextraction, and so on) show asymmetric topographicpatterns between the superior and inferior cornea sim-ilar to those of keratoconus. In addition, steepening inkeratoconus is not limited to the inferior periphery.'

o Sensitivity• SpecificityA Accuracy

.24 .26KPI

FIGURE 6. Sensitivity, specificity, and accuracy of kerato-conus pattern detection in the training set with the expertsystem. The lower the cutoff value was set, the higher thesensitivity and the lower the specificity. The most efficientcutoff value was found to be 0.23.

o Sensitivity• SpecificityA Accuracy

.24 .26KPI

FIGURE 7. Sensitivity, specificity, and accuracy of kerato-conus pattern detection in the validation set with the expertsystem. The lower the cutoff" value was set, the higher thesensitivity and the lower the specificity. Note that the cutoffvalue of 0.23 also showed higher sensitivity with highest accu-racy in the validation set.

Comparison between fellow eyes is generally useful fordetecting the keratoconus pattern but may be prob-lematic for detecting cone progression because conedevelopment in the fellow eye may actually lag behindthe eye in question, or the contralateral cornea mayhave undergone surgery. Contralateral eye informa-tion is not routinely available in videokeratographyand should not be relied upon for quantitative analy-sis. Thus, dioptric maps cannot be reliably and univer-sally distinguished as keratoconus or nonkeratoconusbased on the method of Rabinowitz and McDonnell.

To detect topographic characteristics of kerato-conus quantitatively, the use of multiple parameters,each of which represents distinctive characteristics ofthe map, is desirable. Keratoconus patterns in video-keratoscopy can be characterized by an area of local-ized, abnormal steepening. Localized steepening is of-ten observed in the inferior quadrant, but sometimesit is seen in the center or superior portion of the cor-nea.:> This results in asymmetry and a large refractivepower difference across the corneal surface. In thisstudy, we used eight indices to extract these character-istics. The DS1 and OSI are sensitive to a localizedabnormal steepening in the periphery, and the CS1 issensitive to a centrally located steepening (Fig. 2). SAIis also sensitive to topographic asymmetry of decen-tered cones. IAI and AA describe the characteristicirregularity of the corneal power distribution oftenassociated with moderate to severe keratoconus.SimKl and SimK2 were used to detect corneal steepen-ing, and the amount of cylinder also associated nonex-clusively with keratoconus.

Discriminant analysis was used as a multivariateanalysis of these indices. When dependent variablesare categorical and the independent variables are di-



TABLE 6. Classification by Expert System (Cutoff Value 0.23)

Actual Category

Training setKeratoconusNonkeratoconus

Validation setKeratoconusNonkeraioconus

Predicted Category

Keratoconus

223

251

Nonkeratoconus

075

371

Sensitivity

100%

80%

Specificity

06%

00%

Accuracy

07%

06%

mensional, discriminant analysis is considered one ofthe most appropriate statistical techniques.12 KPI isthe index that is proportional to the discriminant cut-off value obtained from the discriminant function.Average KPI values for keratoconus were significantlyhigher than those of any other category. This meansthat KPI is able to differentiate keratoconus not onlyfrom normal corneas but also from keratoplasty, epi-keratophakia, photorefractive keratectomy, radial ker-atotomy, and contact lens-induced warpage. Althoughone can obtain a system that can perform keratoconusscreening with high specificity and high accuracy, dis-criminant analysis by itself is not sufficient for clinicalscreening of the keratoconus pattern because of rela-tively low sensitivity (68% in the validation set).

An expert system is an artificial intelligence ap-proach that contains a modifiable knowledge base(rules and facts), an interface to users, and an interfaceengine that makes logic-based decisions. Expert sys-tems have been applied previously to medical instru-ments as a tool of computer-assisted diagnosis.l:<1) Wecreated such an expert system through the combina-tion of discriminant analysis with a classification treeto achieve a system for keratoconus pattern screeningthat was better than the discriminant classifier alone.The current program with the cutoff value 0.23 couldcorrectly differentiate keratoconus and nonkerato-conus maps in 96% of cases. Although three false-ne-gative cases were clinically diagnosed as keratoconusand the contralateral eyes of these patients showed thetypical topographic appearance of keratoconus, wewere unable to confirm the diagnoses in this retrospec-tive study. In fact, the three false-negative maps re-sembled contact lens-induced corneal warpage in twoeyes and pellucid marginal degeneration in one eye.All the false-positive cases involved eyes that had un-dergone keratoplasty, and areas of localized, abnor-mal steepening similar in topographic appearance tokeratoconus were seen in these maps. These false-negative and false-positive cases are reasonable underthe current conditions in which there is no additionalinformation besides the topographic maps.

In this study, the new methodology enabled theinterpretation of topographic patterns quantitativelyand objectively, an advantage over the subjective vi-sual inspection of the color-coded map. Because thedetection scheme presented here is keyed to detect thekeratoconus pattern from videokeratographs, contactlens-induced corneal warpage,u> post-keratoplastycorneas, or other pathologic corneal conditions canbecome false positive. In fact, such false positives of-ten cannot be detected with visual inspection of thevideokeratograph alone. The clinical diagnosis of ker-atoconus must be confirmed with more traditionalclinical methods. To compare classifiers with well-es-tablished tests or findings that unequivocally definethe disease, keratoconus suspect patients were ex-cluded because there are no other accepted standardsexcept corneal topography itself. Although they showthe keratoconic topographic pattern, some of the ker-atoconus suspect cases become false negative with theclassifier because it was trained intentionally with clin-ically diagnosed keratoconus.

The ability to screen automatically for kerato-conus patterns will be a beneficial tool in the clini-cian's armamentarium. With experience, the examinercan interpret topographic abnormalities based on theindices and can gain new insight into the relationshipbetween the appearance of the color-coded map andspecific measures of corneal power distribution. Addi-tionally, this system can be adapted to detecting otherpatterns in corneal topography by developing appro-priate quantitative indices sensitive to the uniquecharacteristics of those patterns.

Key Words

corneal topography, keratoconus, diagnostic screening, dis-criminant analysis, expert system

References

1. Krachmer JH, Feder RS, Belin MVV. Keratoconus andrelated noninflammatory corneal thinning disorders.Surv Ophthalmol. 1084;28:203-322.



2. Maguire LJ, Bourne VVM. Corneal topography ofearly keratoconus. Am J Ophthalmol. 1989; 108:107-112.

3. Rowsey JJ, Reynolds A, Brown R. Corneal topogra-phy, comeascope. Arch Ophthalmol. 1981 ;99:1093-1100.

4. Rabinowiiz YS, McDonnell PJ. Computer-assistedcorneal topography in keratoconus. Refract CornealSurg. 1989; 5:4*00-408.

5. Wilson SE, Lin DTC, Klycc SD. Corneal topographyof keratoconus. Cornea. 1991; 10:2-8.

6. Rabinowiiz YS, Garbus J, McDonnell PJ. Computer-assisted corneal lopography in family members of pa-tients with keratoconus. Arch Ophthalmol. 1990; 108:365-371.

7. Gonzalez V, McDonnell PJ. Computer-assisted cor-neal topography in parents of patients with kerato-conus. Arch Ophthalmol. 1992; 110:1412-1414.

8. Maguire LJ, Lowry JC. Identifying progression ofsubclinical keratoconus by serial topography analysis.Am J Ophthalmol. 1991; 1 12:41-45.

9. Rabinowiiz Y, Garbus JJ, Garbus C, McDonnell PJ.Contact lens selection for keratoconus using a com-

puter-assisted videokeratoscope. CLAO J. 1991; 17:88-93.

10. Dingeldein SA, Klycc SD, Wilson SE. Quantitative de-scriptors of corneal shape derived from computer-as-sisted analysis of photokeratographs. Refract CornealSurg. 1989; 5:372-378.

1 1. Wilson SE, KJyce SD. Quantitative descriptors of cor-neal topography: A clinical study. Arch Ophthalmol.1991; 109:349-353.

12. Hair JF Jr, Anderson RE, Tatham RL, Grablowsky BJ.Multivariate Data Analysis luith Readings. Tulsa, OK:The Petroleum Publishing Company; 1979:82-122.

13. Marchevsky, AM. Expert systems for efficient han-dling of medical information. I. Lung cancer. AnalytQuant Cytol Histol. 1991;! 3:89-92.

14. Sicgel, S. Nonparavw.tric Statistics for the BehavioralSciences. New York: McGraw-Hill Book Company;1956:63-67.

15. Rubin, A. Design of an expert system and its applica-tion to dermatopalhology. Histopathology. 1992;21:269-274.

16. Wilson SE, Lin DTC, Klycc SD, Rcidy JJ, Inslcr MS.Topographic changes in contact lens-induced cornealwarpage. Ophthalmology. 1990;97:734-744.


Automated keratoconus screening with corneal topography analysis.

Documents