Future Promise of and Potential Pitfalls for Automated Detection of Diabetic Retinopathy.

Abstract

The effective and efficient evaluation and triage of patients at risk for vision loss fromdiabetic eyedisease remains a criticalhealthcarechallenge.Theglobalepidemicofdiabetesmellitus will put more than 590 million individuals at risk for vision-threatening complications of diabetic retinopathy (DR) over the next few decades. Advances in treatment have greatly improvedvisual outcomes formanypatientswithdiabetesmellitus.Nonetheless, optimal results fromthese therapies cannot be obtained unless individuals with visionthreatening diabetic pathology are identified early enough to be treated in an appropriate and timely fashion. Telemedicine programs for DR have the potential to address this issue by loweringbarriers to screeningandallowingaccess to a large population of patients with diabetes mellitus who would not otherwise be evaluated through traditional means.1 Ascamera technologieshaveadvanced,nonmydriatic retinal imagingobtainedwithmoreportable, durable, andaffordable equipmenthas allowedgreater access to teleophthalmology programs across a wide range of geographical and socioeconomic barriers. The use of remote retinal imaging to expand the reach of existing health care to those who might not otherwise have access to such care can dramatically improve visual outcomes in patients with diabetes mellitus. In theUnitedKingdom, the institutionofanationwideDRscreening program and improved glycemic control has led to dramatic reductions in blindness caused by diabetes mellitus.2 Nonetheless, limitations to this approach include the substantial resources required to ensure each image is evaluated by a trained individual. Given the growing number of individuals withdiabetesmellitus,manual retinal image evaluation is not scalable for a global population. Thedevelopmentofautomatedalgorithmsthat reliablydetect retinalpathologycouldeffectivelyaddress theneedtoprovide standardized and certified grading of retinal images for DRonaworldwide scale. Computerizedmethodsenable rapid throughput of large retinal image data setswhilemaintaining uniform grading standards across multiple patient populations.Anautomatedgradingalgorithmmightbeespeciallybeneficial for DR, given the complex nature of the Early TreatmentDiabeticRetinopathyStudyseverityclassificationsystem, which is thecurrentgold standard for clinical and researchuse. Walton and colleagues3 have evaluated an automated algorithm, IntelligentRetinal ImagingSystem(IRIS), to gradeDR presence and severity compared with manual grading of the same images. This study was performed in a large, 15 015personcohortofalldiabeticpatients imagedconsecutivelyover an 11-month period in the Harris Health System in Houston, Texas.When comparedwith eye care specialist interpretation of the 45° retinal images, the IRIS algorithm had a sensitivity of 66.4% (95%CI, 62.8%-69.9%), specificity of 72.8% (95%CI, 72.0%-73.5%),positivepredictivevalueof 10.8%(95%CI,9.6%11.9%), andnegativepredictivevalueof97.8%(95%CI, 96.8%98.6%) fordetectingsight-threateningdiseaseasdefinedbythe presence of either severe nonproliferative or proliferative DR. Although there are no set universal guidelines, the IRIS metrics are lower thanminimum thresholds (sensitivity, 80% and specificity, 95%) suggested by the British Diabetic RetinopathyWorkingGroup for detection of anyDR.4As the goal of disease detection changes from any retinopathy to visionthreateningretinopathy, it isevenmorecritical forpatientsafety that the sensitivityof analgorithmapproaches 100%toensure thatallpatientswithdiseaseareaccurately identified.Otherpreviously published automated algorithms in different populations have reported high sensitivities between 0.96 and 0.98, but lower specificities of 0.41 to 0.59.5 This article highlights an additional challenge faced in the current era of automated screening for DR, in that automated analysis of retinal images is necessarily limited by the quality of the camera systems used and the resulting images. Even with use of dilating drops to improve photographic quality, nearly 1 in every 6 patients in this study had images that were not acceptable for analysis. The large number of ungradable images decreases the generalizability of these results and potentially compromises the system’s sensitivity for detecting disease because ungradable images may be due to intraocular pathology, necessitating the need for a comprehensive eye examination. Perhaps themost important aspectof this report is that the authorsmethodically evaluated the algorithm in a large study of eyeswith a full range ofDR severity, clearly recognizing the need tovalidate any telemedicineprogramforDR.TheAmerican Telemedicine Association has defined validation criteria forDR telemedicineprograms, ranging from identifyingpresenceor absenceofDR (category 1) to the ability tomatchEarly TreatmentDiabetic Retinopathy Study protocol photography in any clinical or research arena (category 4).6 In this study,3 the IRIS programwasmeasured againstmanual grading of an unspecified number of 45° images from each eye. This does not ensure that IRIS identifies retinopathy accurately vis-avis the accepted standard of Early Treatment Diabetic Retinopathy Study protocol 7-field stereoscopic photography. Nevertheless, this study is an important early step in the development of this algorithm and its path to American Telemedicine Association validation. While thevalidationof individual automatedprograms for assessment of diabetic retinal disease is critical, what will ultimately be needed is the ability to compare multiple algorithms across different imaging platforms and diverse coRelated article page 204 Research Original Investigation Automated Teleretinal Screening Program for Diabetic Retinopathy