July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Clinical Validation of a Deep Learning Automated Algorithm for the Detection of Diabetic Retinopathy and Macular Edema
Author Affiliations & Notes
  • Kristen Stebbins
    Welch Allyn, Baldwinsville, New York, United States
  • Ynjiun Wang
    Welch Allyn, Baldwinsville, New York, United States
  • Lilian Tang
    Welch Allyn, Baldwinsville, New York, United States
  • Naren Suri
    Welch Allyn, Baldwinsville, New York, United States
  • Su Wang
    Welch Allyn, Baldwinsville, New York, United States
  • Amish Purohit
    Welch Allyn, Baldwinsville, New York, United States
  • Max Johnson
    Welch Allyn, Baldwinsville, New York, United States
  • Edward Chaum
    Welch Allyn, Baldwinsville, New York, United States
  • Footnotes
    Commercial Relationships   Kristen Stebbins, Welch Allyn (E); Ynjiun Wang, Welch Allyn (E); Lilian Tang, Welch Allyn (E); Naren Suri, Welch Allyn (E); Su Wang, Welch Allyn (F), Welch Allyn (C); Amish Purohit, Welch Allyn (C); Max Johnson, Welch Allyn (C); Edward Chaum, Welch Allyn (C)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 1440. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Kristen Stebbins, Ynjiun Wang, Lilian Tang, Naren Suri, Su Wang, Amish Purohit, Max Johnson, Edward Chaum; Clinical Validation of a Deep Learning Automated Algorithm for the Detection of Diabetic Retinopathy and Macular Edema. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1440.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : The purpose of this study was to determine the accuracy of a deep learning based automated algorithm, AutoDx-DR, for the detection of diabetic retinopathy (DR) and macular edema (DME) from fundus photographs. It was hypothesized that AutoDx-DR could identify more than mild DR or any DME with high sensitivity and specificity, proving it to be an effective tool for use in clinical practice.

Methods : In a retrospective study we utilized a set of fundus images from the RetinaVue® Network (Welch Allyn®, Skaneateles Falls, NY) database for analysis. Images were each sent to 4 retina specialists for independent interpretation and grading, followed by adjudication if needed. Images included in the final analysis set had agreement of at least 3 readers using the International Clinical Diabetic Retinopathy Disease Severity (ICDRS) scale including the presence or absence of exudates, a marker for DME in the ICDRS. AutoDx-DR provided a “refer” result for images with more than mild DR or any DME and a “pass” result for no DR or mild DR.

Results : There were 565 images with 3 reader agreement included in the final analysis. The results of the manual reads were as follows: 407 (72%) had no DR, 48 (8%) had mild nonproliferative DR, 98 (17%) had moderate nonproliferative DR, 3 (1%) had severe nonproliferative DR, and 9 (2%) had proliferative DR. Twenty (4%) images showed evidence of DME. AutoDx-DR sensitivity was 97.1% (95% CI: 91.1%-99.2%) and specificity was 96.5% (95% CI: 94.1%-97.9%) compared to the pass/refer grade by the adjudicated reader agreement. There were 39 images (6.9%) that were considered too low of quality for interpretation by AutoDx-DR. Individual readers had substantial agreement with the final analysis (κ = 0.72 to κ = 0.90). AutoDx-DR had very strong agreement with the final analysis by the readers (κ = 0.89).

Conclusions : Deep learning-based algorithms like AutoDx-DR can detect referable DR and DME with a high level of accuracy. AutoDx-DR identified referable DR and DME with sensitivity and specificity exceeding 95%, suggesting that it is as accurate as a retina specialist grader. With the prevalence of diabetes increasing dramatically and diabetic eye exams becoming more common in non-specialty care settings, tools like AutoDx-DR can provide real-time, automated, accurate assessment and referral of patients at risk for vision-threatening diabetic retinopathy.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×