Professional Documents
Culture Documents
Frank D. Verbraak,1
Diagnostic Accuracy of a Device Michael D. Abramoff,2,3,4
Gonny C.F. Bausch,5 Caroline Klaver,6,7,8
for the Automated Detection of Giel Nijpels,9 Reinier O. Schlingemann,10
and Amber A. van der Heijden9
Diabetic Retinopathy in a Primary
Care Setting
Diabetes Care 2019;42:651–656 | https://doi.org/10.2337/dc18-0148
Multiple diagnostic algorithms for the and reflected the mixed multiethnicity (22–24), independently graded each
detection of DR are now commercially of the general population of Rotterdam, exam per ICDR grading system. Graders
available for which the performance has with around 15% non-Caucasian inhab- were masked to any algorithm outputs.
been independently evaluated (9–12). itants. Disagreements between the two readers
One of these, the IDx-DR-EU-2.1 device, were adjudicated by an experienced ret-
has been enhanced with deep learn- Imaging inal specialist (F.D.V.) for the final grade.
ing. Deep learning, a machine learning Participants underwent fundus imaging For analysis, the final ICDR grades were
technique that uses multilayer neural according to a strict standardized pro- combined into no or mild DR (and no
networks, has allowed substantial im- tocol (two per eye: one macula centered DME) and moderate DR (mtmDR and not
provements in artificial intelligence and one disc centered [45° field of view]) vtDR) or vtDR (see Supplementary Table
(AI)-based diagnostic systems (13–17). using Topcon TRC-NW200 cameras op- 1). The presence of exudates, retinal
Because deep learning is used to build its erated by experienced Star-SHL techni- thickening (if visible on nonstereo photo-
explicit retinopathy lesion (biomarker) cians. The images were made in eight graphs), within 1 disc diameter of the
detectors, the IDx-DR-EU-2.1 is a lesion- different sites, and settings of the cameras fovea, was taken as evidence of DME (19).
Between 1 January 2015 and 31 Decem- Reading center reference standard No or mild Insufficient
ber 2015, 1,616 participants were imaged. (ICDR grading system) DR mtmDR vtDR quality Total
Mean age was 63 years (SD 11.3), and No DR 1,050 19 6 112 1,187
53% of the participants were male (see Mild DR 104 40 11 12 167
STARD diagrams [Figs. 1 and 2]). Moderate DR 13 23 11 8 55
Of these 1,616 participants, the im- vtDR 0 0 16 0 16
ages of 191 (11.7%) were graded as of All 1,167 82 44 132 1,425
insufficient quality by the reference
Data are n.
standard. Of the 1,425 participants
654 Diagnostic Accuracy of Automated Detection of DR Diabetes Care Volume 42, April 2019
There were 13 false negative exams and an independent reference standard. patients with diabetes are regularly
for the enhanced device’s mtmDR out- These results confirm corresponding re- seen, could improve the percentage of
put according to the ICDR reference sults in an earlier study of essentially the patients screened when indicated. In ad-
standard, and all images for these par- same algorithm in a laboratory setting dition, such a device would lead to im-
ticipants are shown in Fig. 3. Review (13). Specifically, the device achieved proved accuracy compared with present
of the images of the 13 false negative high sensitivity (100%) in people with standard of care and will lead to a higher
cases in Fig. 3 indicated that these par- vtDR, as the device did not miss any number of patients with images with
ticipants had a single isolated hemor- vtDR, or DME, according to the ICDR sufficient quality owing to the direct
rhage or cotton wool spot and had no grading system. It also achieved high feedback of the device regarding the
microaneurysms. specificity (97.8%). However, the number image quality. Nongradable images can
of vtDR cases, although representative either be seen by a human grader or
CONCLUSIONS for the studied patient population, was directly referred to an eye care provider,
The results show that a hybrid lesion- low and prevents definite conclusions. implying that no diagnoses of DR were
based device, with deep learning The device also had a high sensitivity to missed as a result of images of insuffi-
Figure 3—Right and left eye images (two images per eye, with one disc and one fovea centered) of the 13 participants who were false negative for
the mtmDR output of the device, according to the ICDR reference standard. None had vtDR or macular edema. The vtDR output did not have any
false negatives.
care.diabetesjournals.org Verbraak and Associates 655
Clinicians increasingly deviate from the color images, which lack stereo, and no University of Iowa; Research to Prevent Blind-
methods used by reading centers, as defined macular optical coherence tomography ness, New York, NY. This material is the result of
work supported with resources and the use of
in the original standards (31). For example, was availablednow a widely used facilities at the Iowa City VA Medical Center.
whether a single red lesion is a microaneur- method for determining the presence Contents are solely the responsibility of the
ysm or a hemorrhage can make the differ- of DME. Isolated retinal thickening may authors and do not necessarily represent the
ence between a mild versus moderate level be underappreciated (37), though human official views of the Department of Veterans
of DR. These levels were used in the primary expert detection of DME from exudates Affairs or the U.S. government.
Duality of Interest. This study was funded by
outcome studies that to a great degree still only, in nonstereo images, was shown to be IDx. M.D.A. is listed as inventor on patents and
determine the management of DR, such as almost as sensitive as clinical stereo bio- patent applications related to the study subject.
the Diabetic Retinopathy Study (DRS) (32), microscopic analysis of retinal thickening M.D.A. is director of and shareholder in IDx. All
Early Treatment of Diabetic Retinopathy (38,39). DME prevalence and severity may authors, with the exception of G.N., received
financial support from IDx. No other potential
Study (ETDRS) (33), and DCCT/EDIC be underestimated in this data set, and a
conflicts of interest relevant to this article were
(Diabetes Control and Complications reference standard including optical co- reported.
Trial/Epidemiology of Diabetes Interven- herence tomography could lead to differ- Author Contributions. F.D.V. drafted the man-
retinopathy from the Nakuru Study, Kenya. PLoS referable diabetic retinopathy. JAMA Ophthal- and the temporal trends in diagnostic angi-
One 2015;10:e0139148 mol 2013;131:351–357 ography in stable coronary artery disease:
10. van der Heijden AA, Abramoff MD, Verbraak 20. Wilkinson CP, Ferris FL III, Klein RE, et al.; a population-based study. Am Heart J 2016;
F, van Hecke MV, Liem A, Nijpels G. Validation of Global Diabetic Retinopathy Project Group. 175:153–159
automated screening for referable diabetic retinop- Proposed international clinical diabetic reti- 32. Photocoagulation treatment of proliferative
athy with the IDx-DR device in the Hoorn Diabetes nopathy and diabetic macular edema disease diabetic retinopathy: the second report of di-
Care System. Acta Ophthalmol 2018;96:63–68 severity scales. Ophthalmology 2003;110:1677–1682 abetic retinopathy study findings. Ophthalmol-
11. Maguire MG, Daniel E, Niemeijer M, Pistilli 21. Retinopathie D. Concept richtlijn (02-05- ogy 1978;85:82–106
M, Folk JC, Abramoff MD. Identifying diabetic eye 2017) [Internet], 2017. Available from https:// 33. Early Treatment Diabetic Retinopathy Study
disease: comparison of clinical examination by www.oogheelkunde.org/richtlijn/diabetische- Research Group. Fundus photographic risk fac-
ophthalmologists to automated detection from retinopathie-multidisciplinaire-richtlijn-geautoriseerd- tors for progression of diabetic retinopathy.
retinal color images (Abstract). ARVO Meeting november-2017. Accessed 5 May 2017 ETDRS report number 12. Ophthalmology 1991;
Abstracts 2015;56 22. de Voogd S, Ikram MK, Wolfs RC, et al. Is 98(Suppl. ):823–833
12. Tufail A, Kapetanakis VV, Salas-Vega S, et al. diabetes mellitus a risk factor for open-angle 34. American Diabetes Association. Executive
An observational study to assess if automated glaucoma? The Rotterdam Study. Ophthalmol- summary: standards of medical care in diabe-
diabetic retinopathy image assessment software ogy 2006;113:1827–1831 tes–2012. DiabetesCare 2012;35(Suppl. 1):S4–S10
can replace one or more steps of manual imaging 23. Klaver CCW, Assink JJM, van Leeuwen R, et al. 35. Hutchinson A, McIntosh A, Peters J, et al.