Validity of Automated Software Supported Diabetic Retinopathy Scr
Journal of Clinical and Experimental Ophthalmology

Journal of Clinical and Experimental Ophthalmology
Open Access

ISSN: 2155-9570

+44 1223 790975

Research Article - (2018) Volume 9, Issue 4

Validity of Automated Software Supported Diabetic Retinopathy Screening Compared to Digital Retinal Photograph Evaluation by Retina Subspecialist

Ziaul Haq Yasir*, Hassan Al-Dhibi and Rajiv Khandekar
Department of Ophthalmology, King Khaled Eye Specialist Hospital, Riyadh, Saudi Arabia
*Corresponding Author: Ziaul Haq Yasir, King Khaled Eye Specialist Hospital, Al oruba street, Um Al-Hamam West, Riyadh, Riyadh 11462, Saudi Arabia, Tel: + 966 547081988 Email:

Keywords: Diabetic retinopathy; Sight threatening diabetic retinopathy; Diabetic macular edema; RetinaLyze


Diabetes Mellitus (DM) Type II is a major public health problem worldwide [1]. Diabetic retinopathy (DR) is a leading cause of vision loss in the working-age population [2]. DR affects 126.6 million diabetics globally and 37.3 million diabetics have sight threatening diabetic retinopathy (STDR) [3]. Early detection of DR and timely management of STDR will decrease the public health burden of this disease. The exponential rise in the prevalence of diabetes implies that the current yearly screening efforts for DR will be inadequate [4].

The evaluation of DR using fundus photographs is a reliable and the most commonly used method [5]. Changes in DR are noted and documented using ophthalmoscopy, slit lamp bio-microscopy, fundus cameras and smart phone apps. The images are then sent to the reading centers via tele-ophthalmology or through a secure online website [6].

Automated screening software linked to digital cameras was introduced to decrease the workload of specialists and for a faster diagnosis. The technical enhancements for automated image analysis require accurate algorithms. Automated screening software could be useful in developing countries where the burden of DR is high and resources are limited.

Various automated screening programs are available including ARIA, retmarker (Portugal), EyeArt (USA), IDP (USA), iGrading and RetinaLyze [7]. The diagnostic accuracy of these automated programs is debatable especially in the presence of other ocular comorbidities [8,9]. However, the automated screening programs may assist the primary care physicians in timely referral of the majority of STDR cases [10]. A Dutch study reported high validity of a commercially available automated fundus image analysis software (RetinaLyze, Netherland) [11]. However, to the best of our knowledge, there is no published study that evaluated this software on a diabetic Arab population. Hence, prior to broad scale application, it is essential to test this software in an Arab population.

Saudi Arabia is facing an epidemic of DM and DR. In 2012, the prevalence of DM in adult Saudis was 29.7%. Among registered diabetics, the prevalence of DR and STDR was 36.8% and 17.5% respectively [12]. In 2016, more than 452,200 diabetic patients attended government health institutions in Saudi Arabia [13]. However, there are only 700 ophthalmologists in these institutions, posing a major challenge for annual DR screening. In developing countries, task shifting has been used as a strategy for effective DR screening [14]. In Saudi Arabia, mid-level eye care professionals are limited yet the geographic spread of patients with DM is vast. Therefore, using the services of mid-level eye care professionals for DR screening remains a major challenge. A software system for automated fundus image analysis could therefore benefit the national health program. This study compares the grading of DR from digital retina images by RetinaLyze and a retina specialist in Saudi Arabia.

Subjects and Methods

The institutional ethics and research board approved this study (P-1309). Diabetics registered at two Primary Health Centers (PHCs) in the Riyadh region of Saudi Arabia and patients who presented at the screening unit of a tertiary eye care hospital were invited to participate in this study. Informed verbal consent was obtained from all patients. This cross-sectional validity study was performed between December 2016 and June 2017. Diabetics with media opacity and hazy digital images as per retina specialist’s evaluation were excluded. Those declining to participate were also excluded.

To calculate the sample size for the present study, we assumed that the sensitivity of the software assisted grading of DR was 90% [11]. To achieve a 95% confidence interval (CI), with a 5% acceptable margin of error and a clustering effect of 2, at least 277 eyes of diabetic patients were required for evaluation by the software and by the retina specialist [15].

Medical retina specialist, ophthalmic technician and epidemiologist were the study investigators. Diabetics were defined as individuals who were registered in the diabetes registry of PHCs or referred to the eye hospital for management of DR.

The digital fundus images were obtained using TRC-NW-300 (Topcon Corp., Tokyo, Japan) non-mydriatic retina camera. One central fundus image covering approximately 45° of retina from the fovea was captured [16]. The fundus images were uploaded from the laptop attached to the retina camera to the website of RetinaLyze. For maintaining confidentiality of patient images, a designated login ID and password were used. A specific client reference number was added for each photo. After uploading the image, the option of running analysis for DR and for age related macular degeneration (AMD) was selected. The results were displayed on the monitor within a few seconds. The software detected ‘red lesions’ (micro-aneurysms and/or hemorrhages) and used this information to grade DR. RetinaLyze can also detect hard exudates and/or cotton-wool spots designated as a ‘bright lesion’. The drusen could also be detected by software and only included in the algorithm for age related macular degeneration (AMD) screening. The steps to software interpretation of the image are described in the manual [17]. The software converts each retina image into a gradient representation. Automated lesion detection is based on the advanced mathematical analysis of the gray-level intensity of the images, where the periphery of potential lesions is established from each of a number of seed points. The optic nerve head (ONH) and the arcades are automatically identified and are excluded to define them as a lesion. A measure of visibility was assigned to each potential red and bright lesion, and lesions exceeding user-supplied visibility thresholds were automatically detected and displayed by the system [18]. The RetinaLyze software gave color code results; ‘no immediate alteration’ (green), ‘few alterations’ (yellow) or ‘severe alterations’ (red). This software however, does not indicate diabetic macular edema (DME). DR changes in fundus image can be located by clicking the ‘Toggle DR’ overlay icon. The software encircles the micro-aneurysms and/or hemorrhages present in that particular photo with a black ring. Based on the automated analysis, the software then recommends if a visit to an ophthalmologist is warranted earlier than scheduled. Software identification of a single ‘red lesion’ of any type in any image of a diabetic patient labels the patient as having DR and recommends an ophthalmic referral. Image quality is measured by the variation in the gradients in the image and according to a designated cutoff level of the image quality threshold. Images with small or no gradients are rejected and defined as ungradable. The rejection of one image only from a specific patient will classify the patient as having images of insufficient quality and will recommend referral to an ophthalmologist [17]. The analyzed image with overlay can be downloaded and printed.

The digital retina image was physically transferred using high quality external hard disc to the retina specialist to grade DR. An information technology (IT) expert ensured that the image quality was not negatively affected during the image transfer. The retina status of DR and DME was graded separately. Macular edema was defined as the presence of hard exudates or localized retina thickening within 500 μm of the fovea. The severity of DR was defined according to the Early Treatment Diabetic Retinopathy Study (ETDRS) as No DR, Nonproliferative DR (Mild-Moderate-Severe), and Proliferative DR (PDR) [19]. STDR was defined as PDR with or without DME [20]. Feedback from field staff was collected on image capture with the retina camera and image upload to the software.

Data were collected on a pretested data collection form and then transferred to an Excel® spreadsheet (Microsoft Corp., Redmond, WA, USA). For univariate analysis, a parametric method was used with Statistical Package for Social Sciences (SPSS-24) (IBM Corp., Armonk, NY, USA). The agreement rate was estimated for software-assisted outcomes versus a retina specialist’s grading of DR. The retina specialist interpreted the digital images and graded them as: (1) presence of DR, (2) presence of STDR (warranting an ophthalmologist intervention). The percentage proportions were calculated. In cases with more than 2 response options the Kappa value was calculated. The sensitivity was defined as the ability of the software to correctly diagnose the presence of DR as compared to the retina specialist’s report. The specificity was defined as the ability of the software to determine an eye without DR correctly compared to the gold standard. A false positive indicated that the software wrongly diagnosed that DR was present when the retina specialist declared the eye did not have DR. A false negative indicated that the software indicated that no DR when the retina specialist had indicated the presence of DR. The 95% confidence intervals (CI) of validity parameters were also calculated. The observations of the retina specialist and RetinaLyze were collected separately to ensure masking of the data and outcome.


We included 476 images of 239 diabetic patients (two patients were monocular). There were 160 (67%) males and 79 (33.0%) females. The mean age was 56.7 ± 11.5 years.

The prevalence of DR in study population was 52.2% (95% CI: 47.7-56.7). The prevalence of STDR was 22.4% (95% CI: 18.6-26.2). The prevalence of DME in the study population was 21.5% (95% CI: 17.7-25.3).

For validation of software based DR grading, 49 images were considered blurred by the software. The retina specialist found 16 images were blurred and the remaining 33 were adequate for grading of DR. Thus 460 images were reviewed by both methods to determine the validity parameters (Table 1). In 281 (61.1%) images there was agreement between the retina specialist and automated software for grading DR.

  Retina specialists grading of DR
DR staging by RetinaLyze System Severe alterations 5 65 5 75
Few alterations 5 92 21 118
No immediate alteration 4 46 184 234
Blurred images 4 19 10 33
  Total 18 222 220 460
Note: Agreement: 281, 61.1%; False negative: 55, 12.0%; False positives: 124, 27.0%; No agreement: 179, 38.9%; Kappa value is 0.424. Total 476 eyes; 16 blurred images according to the retina specialist were not included in the calculation. Thus final sample for comparison was 476-16=460. DR: Diabetic Retinopathy; PDR: Proliferative Diabetic Retinopathy; NPDR: Non-proliferative Diabetic Retinopathy.

Table 1: Validity of RetinaLyze software for DR screening compared to evaluation by a retina specialist.

Total of 47 (10.3%) eyes with STDR were not detected by RetinaLyze as a severe grade warranting referral (Table 2).

  Retina specialists grading of STDR
RetinaLyze Software Severe alterations 5 70 75
Not severe 9 343 352
  Total 14 413 427
Note: Sensitivity: 35.7%; Specificity: 83.3%; False negative: 2.6%; False positives: 93.3%; STDR: Sight Threatening Diabetic Retinopathy; *STDR include proliferative diabetic retinopathy (PDR) and/or diabetic macular edema (DME); Not severe alterations include no immediate alterations or few alterations; Of the 476 eyes of diabetics, 49 blurred images were declared by the software and were not included in the calculation. 476-49=427 eyes were used as gold standard.

Table 2: Validity of RetinaLyze software in Sight Threatening Diabetic Retinopathy (STDR) screening compared to evaluation by a retina specialist.

Field staff feedback suggested that the RetinaLyze software was easy to use and required very little training on the digital camera and the software.


The outcomes of this study indicate that RetinaLyze automated software was easy to use and agreed with a retina specialists evaluation on the stage of DR in approximately 60% of the images. Agreement between only 6 of 10 images indicates that the automated software needs to be more robust to be considered a good DR screening utility. The large number of false positives indicated that it is less likely to reduce the workload of ophthalmologists for timely and accurate DR grading. Additionally, the false positives generated by the software will negatively affect the acceptance of DR screening by caregivers and patients. In the current study, nearly one in ten screened diabetics was false negative. These were the patients with DR that required ophthalmic evaluation. By declaring these patients normal, referrals of genuine DR cases for prompt management may be missed.

The software could not interpret 7% of the images that a retina specialist could review. If an ophthalmologist has to reassess all these patients then the workload will increase considerably.

The automated RetinaLyze software does not directly provide feedback on macular status of retina in diabetics. Vision loss in diabetes is mainly due to DME. DME responds well with timely treatment [21]. It is surprising that this software claims to successfully grade AMD but not DME [22]. DME is an important component of STDR. In the current study, patients with STDR (10.3% of diabetics) who required urgent referral would not have been evaluated by an ophthalmologist if software based action had been taken.

The method based on red lesion identification by this software is likely to miss DR changes at the ONH and vascular arcade near ONH. This could explain lower level of validity in our study. The lack of bright lesion detection by this software is another disadvantage. Exudative retina lesions, which are common in diabetic nephropathy and hyperlipidaemia, would be missed by the lack of bright lesion detection. The presence of these lesions could allow physicians to predict the risk of rapid progression of diabetic complications [23]. The validity of this software can be increased by combining red and bright lesion detection [18].

The validity in our study was lower than that reported by Larsen et al., Hansen et al. and Bouhaimed et al. [11,17,18,24]. Our sample size was much larger than the previous studies [11,17,18,24]. The sensitivity in the previous studies ranges from 82% to 96.7% and the specificity from 71.4% to 100% [11,17,18,24]. The wide difference in the validity parameters among previous studies and ours is difficult to explain. A study regarding glaucoma medication had highlighted outcome differences among researches that were industry sponsored and nonsponsored [25]. It should be noted that our study was not funded however, the software was provided free of cost for independent testing.

We used an ETDRS grading system for diabetic retinopathy [19]. In UK National Health Service (NHS), a more practical grading system is applied [26,27]. Future studies are required to validate this software compared to the NHS grading system.

There are some limitations to this study. The sample size was calculated based on the rate of DR published by Larsen et al. [11] and not for STDR. The prevalence of STDR was lower than DR in our study. Therefore the sample size for validating STDR was not ideal.


More than half of the diabetics in the study population had DR and nearly one fourth of diabetics had sight-threatening stages of DR. RetinaLyze automated screening software was perceived as a useful tool in the field for DR screening. However, its action oriented recommendation based on the image evaluation matched to the advice of retina specialists in 62% of cases only. Further refinement of software to grade DR is needed before it is applied on a wider scale especially for the diabetic population with low prevalence of DR.

Conflict of Interest

Authors declare that there is no conflict of interest.


  1. American Diabetes Association (2010) Diagnosis and classification of diabetes mellitus. Diabetes Care 33: S62-S69.
  2. Tufail A, Rudisill C, Egan C, Kapetanakis VV, Salas-Vega S, et al. (2017) Automated diabetic retinopathy image assessment software: diagnostic accuracy and cost-effectiveness compared with human graders. Ophthalmology 124: 343-351.
  3. Zheng Y, He M, Congdon N (2012) The worldwide epidemic of diabetic retinopathy. Indian J Ophthalmol 60: 428-431.
  4. Shaw JE, Sicree RA, Zimmet PZ (2010) Global estimates of the prevalence of diabetes for 2010 and 2030. Diabetes Res Clin Pract 87: 4-14.
  5. Perumalsamy N, Prasad NM, Sathya S, Ramasamy K (2007) Software for reading and grading diabetic retinopathy. Diabetes Care 30: 2302-2306.
  6. Scott RE (2016) Transforming healthcare with information technology and 21st ISfTeH international conference. J Int Soc Telemed eHealth. 4: 2-1.
  7. Sim DA, Keane PA, Tufail A, Egan CA, Aiello LP, et al. (2015) Automated retina image analysis for diabetic retinopathy in telemedicine. Curr Diab Rep 15: 14.
  8. Abràmoff MD, Reinhardt JM, Russell SR, Folk JC, Mahajan VB, et al. (2010) Automated early detection of diabetic retinopathy. Ophthalmology 117: 1147-1154.
  9. American Academy of Ophthalmology (2018) Screening of diabetic retinopathy-2014. Am Acad Opthamol.
  10. Azrak C, Palazón-Bru A, Baeza-Díaz MV, Folgado-De la Rosa DM, Hernández-Martínez C, et al. (2015) A predictive screening tool to detect diabetic retinopathy or macular edema in primary health care: construction, validation and implementation on a mobile application. PeerJ 3: e1404.
  11. Larsen N, Godt J, Grunkin M, Lund-Andersen H, Larsen M, et al. (2003) Automated detection of diabetic retinopathy in a fundus photographic screening population. Invest Ophthalmol Visual Sci 44: 767-771.
  12. Al Ghamdi AH, Rabiu M, Hajar S, Yorston D, Kuper H, et al. (2012) Rapid assessment of avoidable blindness and diabetic retinopathy in Taif, Saudi Arabia. Br J Ophthalmol 96: 1168-1172.
  13. Praveen D, Patel A, Raghu A, Clifford GD, Maulik PK, et al. (2014) SMARTHealth India: Development and field evaluation of a mobile clinical decision support system for cardiovascular diseases in rural India. JMIR Mhealth Uhealth 2: e54.
  14. Al-Rubeaan K, Abu El-Asrar AM, Youssef AM, Subhani SN, Ahmad NA, et al. (2015) Diabetic retinopathy and its risk factors in a society with a type 2 diabetes epidemic: a Saudi National Diabetes Registry-based study. Acta Ophthalmol 93: 140-147.
  15. Srihatrai P, Hlowchitsieng T (2018) The diagnostic accuracy of single-and five-field fundus photography in diabetic retinopathy screening by primary care physicians. Indian J Ophthalmol 66: 94.
  16. Hansen AB, Hartvig NV, Jensen MS, Borch-Johnsen K, Lund-Andersen H, et al. (2004) Diabetic retinopathy screening using digital non-mydriatic fundus photography and automated image analysis. Acta Ophthalmol Scand 82: 666-672.
  17. Bouhaimed M, Gibbins R, Owens D (2018) Automated detection of diabetic retinopathy: results of a screening study. Diabetes Technol Ther 10: 142-148.
  18. Wu L, Fernandez-Loaiza P, Sauma J, Hernandez-Bogantes E, Masis M. Classification of diabetic retinopathy and diabetic macular edema. World J Diabetes 2013; 4: 290.
  19. Yau JW, Rogers SL, Kawasaki R, Lamoureux EL, Kowalski JW, et al. (2012) Global prevalence and major risk factors of diabetic retinopathy. Diabetes Care 30: 556-564.
  20. Lee R, Wong TY, Sabanayagam C (2015) Epidemiology of diabetic retinopathy, diabetic macular edema and related vision loss. Eye Vision 2: 17.
  21. Rosario RF, Prabhakar S (2006) Lipids and diabetic nephropathy. Curr Diab Rep 6: 455-462.
  22. Larsen M, Godt J, Larsen N, Lund-Andersen H, Sjølie AK, et al. (2003) Automated detection of fundus photographic red lesions in diabetic retinopathy. Invest Ophthalmol Visual Sci 44: 761-766.
  23. Alasbali T, Smith M, Geffen N, Trope GE, Flanagan JG, et al. (2009) Discrepancy between results and abstract conclusions in industry-vs nonindustry-funded studies comparing topical prostaglandins. Am J Ophthalmol 147: 33-38.
  24. Eszes DJ, Szabó DJ, Russell G, Kirby P, PauliK E, et al. (2016) Diabetic retinopathy screening using telemedicine tools: pilot study in Hungary. J Diabet Res 2016.
  25. Hansen AB, Sander B, Larsen M, Kleener J, Borch-Johnsen K, et al. (2004) Screening for diabetic retinopathy using a digital non‐mydriatic camera compared with standard 35‐mm stereo colour transparencies. Acta Ophthalmologica 82: 656-665.
Citation: Yasir ZH, Al-Dhibi H, Khandekar R (2018) Validity of Automated Software Supported Diabetic Retinopathy Screening Compared to Digital Retinal Photograph Evaluation by Retina Subspecialist. J Clin Exp Opthamol 9: 734.

Copyright: © 2018 Yasir ZH, et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.