diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image...

7

Click here to load reader

Upload: anja-b-hansen

Post on 20-Jul-2016

217 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

Diabetic retinopathy screeningusing digital non-mydriaticfundus photography andautomated image analysis

Anja B. Hansen,1 Niels V. Hartvig,1 Maja S. Jensen,1 Knut

Borch-Johnsen,2 Henrik Lund-Andersen1,2 and Michael Larsen1

1Department of Ophthalmology, Herlev Hospital, University of Copenhagen, Denmark2Steno Diabetes Centre, Gentofte, Denmark

ABSTRACT.

Purpose: To investigate the use of automated image analysis for the detection of

diabetic retinopathy (DR) in fundus photographs captured with and withoutpharmacological pupil dilation using a digital non-mydriatic camera.

Methods: A total of 83 patients (165 eyes) with type 1 or type 2 diabetes, represent-

ing the full spectrum of DR, were photographed with and without pharmacologicalpupil dilation using a digital non-mydriatic camera. Two sets of five overlapping, non-

stereoscopic, 45-degree field images of each eye were obtained. All images were

graded in a masked fashion by two readers according to ETDRS standards and

disagreements were settled by an independent adjudicator. Automated detection of

red lesions as well as image quality control was made: detection of a single red lesion

or insufficient image quality was categorized as possible DR.

Results: At patient level, the automated red lesion detection and image quality

control combined demonstrated a sensitivity of 89.9% and specificity of 85.7% in

detecting DR when used on images captured without pupil dilation, and a

sensitivity of 97.0% and specificity of 75.0% when used on images captured

with pupil dilation. For moderate non-proliferative or more severe DR the

sensitivity was 100% for images captured both with and without pupil dilation.Conclusion: Our results demonstrate that the described automated image ana-

lysis system, which detects the presence or absence of DR, can be used as a first-

step screening tool in DR screening with considerable effectiveness.

Key words: diabetic retinopathy – non-mydriatic – automated image analysis – screening

Acta Ophthalmol. Scand. 2004: 82: 666–672Copyright # Acta Ophthalmol Scand 2004.

doi: 10.1111/j.1600-0420.2004.00350.x

Introduction

Photographic screening of patients withdiabetes mellitus entails classification

of diabetic retinopathy (DR) based onthe type, number and location of anymicrovascular lesions in the fundus of

the eye, combining a count of the num-ber of lesions and a comparison withstandard photographs of various stagesof retinopathy (Early Treatment Dia-betic Retinopathy Study ResearchGroup 1991). This method requirestrained personnel, is time-consumingand costly, and also tedious, semi-subjective and prone to error.

The application of automated imageanalysis to digital fundus images may,however, reduce the workload and costsby minimizing the number of photo-graphs that need to be manually graded,while at the same time providing anobjective, repetitive system that reducesintra- and interobserver variability.

In the present study we tested theability of an automated system, whichincludes both red lesion detection andimage quality control, to correctly iden-tify patients as having or not havingDR, based on images captured withand without pharmacological pupildilation using a digital non-mydriaticcamera. The main end-points were: formethodological purposes, a per-eyeanalysis of the system’s performancein relation to variations in lesiondetection sensitivity; and for clinicalpurposes, a per-patient analysis of thesystem’s ability to identify patientswith any DR, as in clinical practicethe decision to refer a patient to anophthalmologist is based on thepatient’s worst eye.

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

666

Page 2: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

Material and Methods

Patients

The digital images of all 83 patientsused in this study were obtained aspart of a photographic validationstudy of the digital non-mydriaticcamera carried out in the Departmentof Ophthalmology, Herlev Hospital.

All patients were recruited at theSteno Diabetes Centre based on arecord review of the most recently diag-nosed level of retinopathy accordingto four morphological groups: noretinopathy, background retinopathy(microaneurysms, haemorrhages, hardexudates), preproliferative retinopathy(which can include intraretinalmicrovascular abnormalities (IRMA)cotton wool spots and venous abnor-malities) and proliferative retinopathy.The recruitment was weighted towardless retinopathy in order to pose agreater challenge in the gradingprocess.

Inclusion criteria were: age 18 yearsor older, diagnosed type 1 or type 2diabetes mellitus. Exclusion criteriawere: pregnancy; previous retinal lasertreatment in either eye; history ofconditions in either eye that mightpreclude pupil dilatation, and the useof eye drops (mydriatic or miotic) thatmight alter pupil size or reactivity.

The study was approved by the localmedical ethics committee. All partici-pants gave written informed consentafter receiving full information accord-ing to the Helsinki Declaration.

Photography

A non-mydriatic digital fundus camera,a TRC-NW6S (TOPCON, Tokyo,Japan) interfaced with a 3CCD colourcamera (KY-F70B; JVC, Tokyo,Japan) with a pixel resolution of1450� 1026 and IMAGEnet 2000computer system were used to capture:

(1) five overlapping, non-stereoscopic45-degree photographs of each eyewithout pharmacological pupil dilation(posterior pole, nasal, temporal, sup-erior and inferior), covering up to94 degrees of the fundus includingboth the macula and the optic disc(Shiba et al. 1999), and(2) five non-stereoscopic images ofeach eye with pharmacological pupildilation using 2.5% phenylephrine and1% tropicamide, covering the samearea.

All digital images were captured intrue colour (24 bits), labelled and storedin TIF format on the IMAGEnet compu-ter system. For manual reading, the fiveimages from each eye were arranged asa single mosaic image and storedusing the IMAGEnet mosaic software.All retinal images were viewed on a24-inch Trinitron colour graphic display(GDM-FW900; Sony, Tokyo, Japan),true colour 1920� 1200 resolution.

Manual grading

All images were graded according to amodification of the ETDRS extensionof the modified Airlie House classifica-tion of DR (Diabetic RetinopathyStudy Research Group 1981; EarlyTreatment Diabetic RetinopathyStudy Research Group 1991) asdescribed elsewhere (Hansen et al.2004). Note that macula oedema wasnot graded in this study, and that‘Cannot grade’ was used when theimage quality made it impossible todetermine whether a characteristic waspresent in a field: if an area of three ormore than three disc areas of the retinawas visible, and that area was withoutthe characteristic being graded, thefield was then graded as ‘no evidence’.If the characteristic was present in thevisible area, the field was then gradedas if the characteristic was present withthe same degree of severity in the wholefield.

Two independent readers graded alldigital images in a masked fashion andthe results were compared. In cases ofdisagreement, an independent adju-dicator performed a final overallgrading. The readers were also askedto evaluate the image quality.

Automated lesion detection

The fundus images were analysed asindividual images and not as the mosaicimage, using commercial fundus imageanalysis software (Larsen et al. 2003a,2003b). The system uses advancedmodelling of the grey-level image func-tion of digital images, primarily thegreen colour channel, and providesautomated red microaneurysm andhaemorrhage lesion detection as wellas image quality measurement.

Each image is converted into a gra-dient representation. The vessel treeand the optic nerve head are identifiedand extracted from the image. Seedpoints of candidate lesion areas (i.e.

dark fundus areas) are located andgrown with custom-developed algo-rithms. A visibility parameter describ-ing the densitometric steepness of theedge of each candidate lesion and itscontrast relative to the surroundingfundus is calculated; and according toa present cutoff level: the visibilitythreshold, it is determined whether thecandidate lesion is accepted or notaccording to a preset cut-off levelknown as the visibility threshold. Notethat identification of a single red lesionof any type in any image from a specificpatient will classify the patient ashaving DR and will lead to referral.

Image quality is measured by the vari-ation in the gradients in the image and;and according to a present cutoff level:the image quality threshold, imageswith small or no gradients are rejected.Note that rejection of one image onlyfrom a specific patient will classify thepatient as having images of insufficientquality and will lead to referral.

Besides an image-scale parameter,the visibility threshold and the imagequality threshold are the only user-supplied parameters in the system. Thevisibility threshold controls the lesiondetection sensitivity and thus the bal-ance between sensitivity and specificityof the patient classification. In a prac-tical screening scenario the thresholdshould be adjusted to the specificimage acquisition and grading proto-col, preferably by a calibration study.For the present study, the adjustmentwas based on the study data, using thesame threshold (visibility threshold 2.1and image quality threshold 0.57)for images captured with and withoutpharmacological pupil dilation.

Statistical analysis

Patients were classified as having auto-matically detected DR if the algorithmidentified a single red lesion of any typein any of the images from that patientor if one or more of the images wereidentified as being of insufficient imagequality. The clinical relevance of thedetected lesions was characterized bythe sensitivity and specificity againstthe manual grading results (Altman1999). The receiver operating charac-teristic (ROC) of the automated redlesion detection was used to character-ize the relationship between sensitivityand specificity, with the area under thecurve (AUC) serving as a general

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

667

Page 3: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

measure of performance (van Erkel &Pattynama 1998; Greiner et al. 2000).All quoted AUC values are presentedwith 95% confidence intervals.

Results

A total of 83 patients (165 eyes) werephotographed and their DR statusassessed by visual grading (the goldstandard) for images captured withoutand with pharmacological pupildilation.

Red lesion detection in images captured

without pharmacological pupil dilation

One patient was found to have verysevere age-related macular degenera-tion (AMD) in both eyes and wasexcluded from this part of the analysisbased on the following considerations:AMD is known to share many featureswith DR, including haemorrhages andhard exudates, making it impossible forthe automated image analysis algo-rithms to differentiate between the twodiseases. A further 11 eyes were alsoexcluded from this part of the analysisdue to images being classified asungradable by the visual grading(including one patient [two eyes] inwhom it was not possible to obtainany images due to small pupils). Thustotal number of eyes was 152 (76patients).

The per-eye ROC curve (Fig. 1A) ofthe automated red lesion detectionillustrates the balance between sensiti-vity and specificity for images capturedwithout pupil dilation at eye level. TheAUC of the ROC curve was 84.1%(95% CI 77.3–90.9%). When the visibil-ity threshold was set to 2.1 for the auto-mated algorithm, the sensitivity fordetecting eyes with DR was 81.1%and the specificity for detecting eyeswithout DR was 80.5%.

The per-patient ROC curve (Fig. 1B)of the automated red lesion detectionillustrates the balance between sensitiv-ity and specificity for images capturedwithout pupil dilation at patient level.The AUC of the ROC curve was 91.8%(95% CI 85.7–98.0%). When the visibil-ity threshold was set at 2.1 for theautomated algorithm, a sensitivity of88.7% and a specificity of 85.7% fordetecting patients with and withoutDR, respectively, were obtained.

Red lesion detection in images captured

with pharmacological pupil dilation

Images from all 83 patients were classi-fied as gradable. The patient with verysevere AMD in both eyes was alsoexcluded from this part of the analysisbased on the considerations statedpreviously. Thus the total number ofeyes was 163 (82 patients).

The per-eye ROC curve (Fig. 1C) ofthe automated red lesion detectiondemonstrates the range between sensi-tivity and specificity for images cap-tured with pupil dilation at eye level.The AUC of the ROC curve was90.8% (95% CI 86.3–95.3%). Whenthe visibility threshold was set at 2.1for the automated algorithm, the sensi-tivity for detecting eyes with DR was90.1% and the specificity for detectingeyes without DR was 73.8%.

The per-patient ROC curve (Fig. 1D)of the automated red lesion detectionillustrates the balance between sensitiv-ity and specificity for images capturedwith pupil dilation at patient level. TheAUC of the ROC curve was 94.0%(95%CI 88.9–99.2%).When the visibil-ity threshold was set to 2.1 for theautomated algorithm, a sensitivity of93.9% and specificity of 75.0% fordetecting patients with and withoutDR, respectively, were obtained.

Image quality control and red lesion

detection in images captured without

pharmacological pupil dilation

All 83 patients were included in thispart of the analysis. The visibilitythreshold was set at 2.1 and the imagequality threshold at 0.57, i.e. patientswith one or more images below thisthreshold were referred owing toinsufficient image quality.

Table 1 summarizes the results of thepatient classification with respect toautomatically detected red lesions andthe image quality measure of theimages captured without pupil dilationand compares them with the results ofthe visual grading. The sensitivityand specificity of the automatedalgorithm with respect to determiningwhether a patient was in need ofreferral were 89.9% and 85.7%, respecti-vely (Table 1). Note that all patients(22 in total) evaluated as havingmoderate non-proliferative or moresevere DR by visual grading werecorrectly identified for referral to anophthalmologist.

False-negative classification (i.e.patients with retinopathy not identifiedby the automated algorithm) amountedto 8.4% (7/83); false-positive classifica-tion (i.e. patients without retinopathyidentified as having retinopathy by theautomated algorithm) amounted to2.4% (2/83). Of the seven false-negatives, five had microaneurysmsonly, one had microaneurysms, haemor-rhages and hard exudates, and one wasjudged as ungradable by the visualgrading and thus referred. The twofalse-positives both had a small hyperpigmentation, which the algorithm mis-took for a red lesion.

In total six patients were evaluated asbeing in need of referral by the visualgrading due to ungradable images,including the one patient where it wasnot possible to obtain any images with-out pupil dilation due to small pupils(Table 1). The automated algorithmslikewise identified six patients for refer-ral due to insufficient image quality, ofwhom five (including the patient withno images) were among those identifiedby the visual grading.

Image quality control and red lesion

detection in images captured with

pharmacological pupil dilation

Images from all 83 patients wereincluded in this part of the analysis.The visibility threshold was set at 2.1and the image quality threshold at 0.57.

Table 2 summarizes the results of thepatient classification with respect toautomatically detected red lesions andthe image quality measure of theimages captured with pupil dilationand compares them with the results ofthe visual grading; The sensitivity of theautomated algorithm with respect todetermining whether a patient was inneed of referral was 97.0%, and thespecificity 75.0% (Table 2). Note thatall patients (in total 27) evaluated ashaving moderate non-proliferative ormore severe DR by the visual gradingwere correctly identified for referral toan ophthalmologist.

The proportion of false-negativeclassification was 2.4% (2/83) andthat of false-positive classificationwas 4.8% (4/83). The two false-negatives both had microaneurysmsonly, whereas the four false-positiveseach had a small hyper pigmentation,which the algorithm mistook for a redlesion.

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

668

Page 4: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

All images were evaluated as grad-able by the visual retinal grading,whereas the automated algorithm iden-tified two patients for referral solelydue to insufficient image quality.

Discussion

Our study demonstrates that anautomated image analysis system as

described in this article can be usedin a screening scenario to detect thepresence or absence of DR with consi-derable effectiveness.

At eye level, when analysing imagescaptured without pharmacologicalpupil dilation using red lesion detectiononly, we achieved a sensitivity of 81.1%and a specificity of 80.5%. Pharmaco-logical dilation of the pupil resulted inan increased sensitivity of 90.1% but at

the cost of a lower specificity of 73.8%.Thus dilation of the pupil increased thenumber of correctly identified eyes withDR but also the number of false-positives. This is likely due to a differencein image quality, as we have previouslyfound that images captured with pupildilation tend to be of better image qual-ity (increased brightness) compared tothose captured without (Hansen et al.2004). The better image quality

Fig. 1. Receiver operating characteristic (ROC) of the automated diabetic retinopathy red lesion detection algorithm. Sensitivity is depicted as a

function of 1 minus specificity. The square denotes the current setting of the algorithm with a visibility threshold of 2.1. With this setting the cited

sensitivities and specificities were achieved. (A) Per-eye analysis on images captured without pupil dilation (152 eyes). (B) Per-patient analysis on

images captured without pupil dilation (76 patients). (C) Per-eye analysis on images captured with pupil dilation (163 eyes). (D) Per-patient analysis on

images captured with pupil dilation (82 patients).

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

669

Page 5: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

facilitates red lesion detection but alsoincreases the number of false-positives;as previous studies using the automatedalgorithm (Larsen et al. 2003a, 2003b)have shown that the false-positiveautomatically detected lesions weremainly found in well defined bright

areas of visible nerve fibre layers or inareas with bright posterior hyaloidreflexes. In these areas small, well cir-cumscribed features of normal yellow-red fundus pigmentation, and smallvessel segments, not identified asbranchings off the main vascular

trunks, appeared as isolated highcontrast elements that mimicked smallhaemorrhages and microaneurysms.

When automated image quality con-trol as well as red lesion detection wasapplied to the images, the sensitivity atpatient level increased from 88.7% to89.9% (without pupil dilation) andfrom 93.9% to 97.0% (with pupil dila-tion). The specificity remainedunchanged. For images captured with-out pupil dilation the increased sensitiv-ity was due to the image quality controlcorrectly identifying for referral five ofthe six patients whose images had beenevaluated as ungradable by the visualgrading (i.e. excluded from the initialanalysis). For images captured withpupil dilation the increase was due tothe image quality control correctlyidentifying for referral two patientswho were assessed as having DR bythe visual grading, but who werefound to have no detectable lesions bythe red lesion detection alone. Thisdemonstrates that the addition of animage quality control to an automatedsystem will not only increase the sensi-tivity of the system, but will also reducethe workload, as previous studies testingautomated algorithms have found itnecessary to manually grade retinalimages for image quality before proces-sing them automatically (Hipwell et al.2000; Sinthanayothin et al. 2002; Larsenet al. 2003b; Olson et al. 2003).

All patients with false-negative grad-ings in this study (seven patients basedon the images captured without pupildilation and two patients based on theimages captured with pupil dilation)had minimal/mild non-proliferativeretinopathy. The potential risk of mis-classifying a patient with that level ofretinopathy can be estimated based ondata published by the Liverpool DiabeticEye Study. Here, the 1-year incidencefor development of sight-threateningDR (defined as ETDRS level � 40)for patients who, at baseline, had noDR or background retinopathy was0.3% and 3.6%, respectively, forpatients with type 1 diabetes (Youniset al. 2003a), and 0.3% and 5.0%,respectively, for patients with type 2diabetes (Younis et al. 2003b). Thesenumbers indicate that the risk of over-looking a single or few microaneurysmsmay be inconsequential in a screeningsituation where the follow-up intervalis 1 year. However, if disease in thesame patient is missed on subsequent

Table 1. Choice of action (referral/non-referral) based on visual grading (gold standard) versus

automated detection of red lesions and/or low image quality per-patient level (n¼ 83), based on

images captured without pupil dilation using a non-mydriatic digital camera, with a visibility

threshold of 2.1 for red lesion detection and image quality threshold of 0.57.

Visual classification Automated classification

No referral Referral Total

Reasons for referral

DR Low image DR and low

quality image quality

No referral (NDR) 12 2 2 0 0 14

Referral 7 62 69

Minimal/mild NPDR 6 32 1 1

Moderate/severe NPDR 0 18 0 0

PDR 0 3 0 0

Other non-DR eye disease 0 1 0 0

Ungradable 1 0 5* 83

Total 19 64

Sensitivity 89.9%

Specificity 85.7%

NDR¼ no diabetic retinopathy; NPDR¼non-proliferative diabetic retinopathy; PDR¼proliferative

diabetic retinopathy.

* Including one patient who had no images due to small pupils, and thus was recorded as

ungradable.

Table 2. Choice of action (referral/non-referral) based on visual grading (gold standard) versus

automated detection of red lesions and/or low image quality per-patient level (n¼ 83). Based on

images captured with pupil dilation using a non-mydriatic digital camera, visibility threshold of 2.1

for red lesion detection and image quality threshold of 0.57.

Visual classification Automated classification

No referral Referral Total

Reasons for referral

DR Low image DR and low

quality image quality

No referral (NDR) 12 4 4 0 0 16

Referral 2 65 67

Minimal/mild NPDR 2 33 2 2

Moderate/severe NPDR 0 23 0 0

PDR 0 2 0 0

Other non-DR eye disease 0 1 0 0

Ungradable 0 0 0 0

Total 14 69 83

Sensitivity 97.0%

Specificity 75.0%

NDR¼no diabetic retinopathy; NPDR¼ non-proliferative diabetic retinopathy; PDR¼ proliferative

diabetic retinopathy.

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

670

Page 6: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

examinations the risk obviouslyincreases, necessitating future studiesto address whether the missing oflesions is a random function or anevent that is repeated systematicallyin a subgroup of eyes or patients; ifthe latter is shown to be the case,possible markers of poor detectabilityshould be identified so they can betaken into account, if possible. Inthis study all patients evaluated ashaving moderate non-proliferative ormore severe DR by the visual gradingwere correctly identified for referral.

In our study we investigated the useof the automated algorithm on imagescaptured with and without pharmacol-ogical pupil dilation.Which of these twoformats is preferable will ultimatelydepend upon the screening setting, thescreening population and the preferredstandards of sensitivity and specificity.The British Diabetic Association (1997)recommended that, when performingretinal photographic screening forDR, the standard for sensitivity shouldbe 80% and for specificity 95%. It ispossible to almost comply with thisrecommendation by changing thealgorithm’s visibility threshold to 4.5for images captured with pupil dilation,thereby obtaining sensitivity of 80.3%and specificity of 93.4%, and to 3.0 forimages captured without pupil dilation,thereby obtaining sensitivity of 80.6%and specificity of 92.9%. However, itshould be noted that these are recom-mendations and are not based on anyscientific evidence; whether such accu-racy and precision are clinically neces-sary has yet to be demonstrated. Wefound that in a screening situation it isimportant to have a system with highsensitivity. We therefore recommendthat patients have their pupils pharma-cologically dilated prior to capturingthe images, making it possible to obtaina sensitivity of 97%. Noting that incertain situations it may not be practi-cal or desirable to dilate the pupilspharmacologically, in which case dataindicate that the experience of thephotographer may influence the imagequality of images captured withoutpharmacologically dilated pupils,which may potentially increase thesensitivity of these images (Hansenet al. 2004).

In light of the increasing number ofpatients with diabetes mellitus (Amoset al. 1997; King et al. 1998; Glumeret al. 2003) and the resultant increase

in pressure on the health care system,we propose the following first-linescreening system for DR as representa-tive of a cost-effective telemedicinetool: patients would be screened yearlyusing images obtained with a digitalnon-mydriatic camera graded by anautomated image analysis algorithm.If the automated algorithm detectedno retinopathy, the patient would berecalled 1 year later for repeated screen-ing; if retinopathy were detected, thepatient’s images would be referred forhuman inspection. As the majority ofpatients with diabetes in most screeningpopulations are likely to have no reti-nopathy the application of such anautomated system would substantiallyreduce the burden of manual grading.

In conclusion, this study showed thatautomated detection of red lesionscombined with image quality controlwas successful in identifying alldiabetes patients with moderate non-proliferative or more severe DR, and97.0% all of patients with anyretinopathy using images capturedwith pupil dilation. It was also demon-strated that the addition of an imagequality control to an automated systemwill not only increase the sensitivity ofthe system, in our study by little over3%, but also reduce the workload,as there will be no need to manuallyassess image quality before images areprocessed automatically.

AcknowledgementsThis study was funded by the DanishDiabetes Association, the Danish EyeHealth Society (Værn om Synet), the EyeFoundation (Øjenfonden), the T. ElmquistFoundation and Synoptik A/S.

Commercial InterestNone of the authors have any commercialinterest in the described automated imageanalysis system.

ReferencesAltman DG (1999): Diagnostic tests. In Pract-

ical Statistics for Medical Research. 1st edn.

London: Chapman & Hall/CRC 409–419.

Amos AF, McCarty DJ & Zimmet P (1997):

The rising global burden of diabetes and its

complications: estimates and projections to

the year 2010. Diabet Med 14: S1–S85.

British Diabetic Association (1997): Retinal

Photography Screening for Diabetic Eye

Disease. London: BDA.

Diabetic Retinopathy Study Research Group

(1981): A modification of the Airlie House

classification of diabetic retinopathy. DRS

Report No. 7. Invest Ophthalmol Vis Sci

21: 210–226.

van Erkel AR&PattynamaPM (1998): Receiver

operating characteristic (ROC) analysis:

basic principles and applications in radiol-

ogy. Eur J Radiol 27: 88–94.

Glumer C, Jorgensen T & Borch-Johnsen K

(2003): Prevalences of diabetes and impaired

glucose regulation in a Danish population:

the Inter99 study. Diabetes Care 26:

2335–2340.

Early Treatment Diabetic Retinopathy

Study Research Group (1991): Grading dia-

betic retinopathy from stereoscopic colour

fundus photographs � an extension of the

modified Airlie House classification.

ETDRS Report No. 10. Ophthalmol 98:

786–806.

Greiner M, Pfeiffer D & Smith RD (2000):

Principles and practical application of the

receiver operating characteristic analysis for

diagnostic tests. Prev Vet Med 45: 23–41.

Hansen AB, Sander B, Larsen M, Kleener J,

Borch-Johnsen, Klein R & Lund-Anderson

H (2004): Screening for diabetic retinopathy

using a non-mydriatic camera compared

with standard 35mm stereo colour transpar-

encies. Acta Opthalmol Scand

Hipwell JH, StrachanF,Olson JA,McHardyKC,

Sharp PF & Forrester JV (2000): Auto-

mated detection of microaneurysms in

digital red-free photographs: a diabetic

retinopathy screening tool. Diabet Med 17:

588–594.

King H, Aubert RE & Herman WH (1998):

Global burden of diabetes, 1995–2025: pre-

valence, numerical estimates and projec-

tions. Diabetes Care 21: 1414–1431.

LarsenN,Godt J,GrunkinM,Lund-AndersenH

& Larsen M (2003b): Automated detection

ofdiabetic retinopathy ina fundusphotographic

screening population. Invest Ophthalmol Vis

Sci 44: 767–771.

Larsen M, Godt J, Larsen N et al. (2003a):

Automated detection of fundus photo-

graphic red lesions in diabetic retinopathy.

Invest Ophthalmol Vis Sci 44: 761–766.

Olson JA, Strachan FM, Hipwell JH,

Goatman KA, McHardy KC, Forrester JV

& Sharp PF (2003): A comparative evalu-

ation of digital imaging, retinal photography

and optometrist examination in screening

for diabetic retinopathy. Diabet Med 20:

528–534.

Shiba T, Maruo K & Akahoshi T (1999):

Development of a multi-field fundus photo-

graphing system using a non-mydriatic

camera for diabetic retinopathy. Diabetes

Res Clin Pract 45: 1–8.

Sinthanayothin C, Boyce JF, Williamson TH,

Cook HL, Mensah E, Lal S & Usher D

(2002): Automated detection of diabetic retino-

pathy on digital fundus images. Diabet Med

19: 105–112.

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

671

Page 7: Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis

Younis N, Broadbent DM, Harding SP & Vora

JP (2003a): Incidence of sight-threatening

retinopathy in type 1 diabetes in a

systematic screening programme. Diabet

Med 20: 758–765.

Younis N, Broadbent DM, Vora JP &

Harding SP (2003b): Incidence of sight-

threatening retinopathy in patients with

type 2 diabetes in the Liverpool Diabetic

Eye Study: a cohort study. Lancet 361:

195–200.

Received on March 31st, 2004.

Accepted on August 17th, 2004.

Correspondence:

Anja Bech Hansen

Department of Ophthalmology

Herlev Hospital

DK-2730 Herlev

Denmark

Tel:þ 45 44 88 46 57

Fax:þ 45 44 53 86 69

Email: [email protected]

ACTA OPHTHALMOLOGICA SCANDINAVICA 2004

672