June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Performance of a Diabetic Retinopathy Deep Learning Model for Ultra-widefield Imaging
Author Affiliations & Notes
  • Tunde Peto
    Centre for Public Health, Queen's University Belfast, Belfast, Belfast, United Kingdom
  • Lloyd P Aiello
    Department of Ophthalmology, Beetham Eye Institute, Joslin Diabetes Centre, Boston, Massachusetts, United States
  • Srinivas R Sadda
    Doheny Eye Institute, Los Angeles, California, United States
    UCLA, Department of Ophthalmology, Los Angeles, California, United States
  • Drew Lewis
    Estenda Solutions Inc, Conshohocken, Pennsylvania, United States
  • Anne Marie Cairns
    Optos plc, Dunfermline, Fife, United Kingdom
  • Dana Keane
    Optos plc, Dunfermline, Fife, United Kingdom
  • Sunny Virmani
    Google Inc, Mountain View, California, United States
  • Jerry Cavallerano
    Department of Ophthalmology, Beetham Eye Institute, Joslin Diabetes Centre, Boston, Massachusetts, United States
  • Barbra Hamill
    Centre for Public Health, Queen's University Belfast, Belfast, Belfast, United Kingdom
  • Lily Peng
    Google Inc, Mountain View, California, United States
  • Sara Ellen Godek
    Optos plc, Dunfermline, Fife, United Kingdom
  • Lu Yang
    Google Inc, Mountain View, California, United States
  • Naho Kitade
    Google Inc, Mountain View, California, United States
  • Jonathan Krause
    Google Inc, Mountain View, California, United States
  • Kira Whitehouse
    Google Inc, Mountain View, California, United States
  • Dale Webster
    Google Inc, Mountain View, California, United States
  • Footnotes
    Commercial Relationships   Tunde Peto Optos, Optomed, Code C (Consultant/Contractor), Allergan, Genentech/Roche, Oxurion, Novartis, Bayer, Heidelberg, Optos, Apellis, Alimera, Bayer, Code R (Recipient); Lloyd Aiello KalVista, NovoNordisk, Code C (Consultant/Contractor), KalVista, Code O (Owner); Srinivas Sadda Amgen, Allergan, Genentech/Roche, Iveric, Oxurion, Novartis, Regeneron, Bayer, 4DMT, Centervue, Heidelberg, Optos, Merck, Apellis, Gyroscope, Code C (Consultant/Contractor), Nidek, Topcon, Heidelberg, Carl Zeiss Meditec, Optos, Centervue, Code F (Financial Support), Carl Zeiss Meditec, Nidek, Optos, Novartis, Code R (Recipient); Drew Lewis None; Anne Marie Cairns OPTOS, Code E (Employment); Dana Keane OPTOS, Code E (Employment); Sunny Virmani Google, Code E (Employment); Jerry Cavallerano None; Barbra Hamill None; Lily Peng Google, Code E (Employment); Sara Godek OPTOS, Code E (Employment); Lu Yang Google, Code E (Employment); Naho Kitade Google, Code E (Employment); Jonathan Krause Google, Code E (Employment); Kira Whitehouse Google, Code E (Employment); Dale Webster Google, Code E (Employment)
  • Footnotes
    Support  OPTOS
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 587 – A0152. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Tunde Peto, Lloyd P Aiello, Srinivas R Sadda, Drew Lewis, Anne Marie Cairns, Dana Keane, Sunny Virmani, Jerry Cavallerano, Barbra Hamill, Lily Peng, Sara Ellen Godek, Lu Yang, Naho Kitade, Jonathan Krause, Kira Whitehouse, Dale Webster; Performance of a Diabetic Retinopathy Deep Learning Model for Ultra-widefield Imaging. Invest. Ophthalmol. Vis. Sci. 2022;63(7):587 – A0152.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To evaluate the performance of a deep learning model for diabetic retinopathy (DR) and diabetic macular edema screening when using ultra-widefield (UWF) imaging.

Methods : For model development, 67,200 UWF images were collected from DR programs and ophthalmology clinics worldwide. 30,836 images were double graded and adjudicated at 8 grading centres by 125 certified graders using ETDRS extension of the Modified Airlie House Classification of Diabetic Retinopathy following the JVN Clinical Trial Ultrawide Field Grading Manual v1.0. The grading system used traditional ETDRS 7-SF field definition as well as extended fields 3-7 to evaluate the retinal periphery. A further 36,364 UWF images were graded using a grading protocol based on the ICDR classification. The dataset was split into training, tuning and testing. The final DR model is an ensemble of 10 EfficientNet-b0 neural networks, independently trained with standard image augmentation techniques. For model validation, two independent sets of images were collected. Model performance was evaluated by comparing its predictions to the adjudicated ground truth for both sets of images.

Results : Prior to clinical validation, the model performance was internally evaluated on an independent set of 1967 images, of which 1050 were graded via adjudication as negative for more than mild diabetic retinopathy (mtmDR negative), and 917 as having referable diabetic retinopathy (mtmDR positive). The overall performance (Table 1) was weighted by target DR distribution. Clinical validation evaluated an independent data set of 420 images selected to achieve a target distribution that enabled appropriate confidence intervals for mtmDR sensitivity and specificity A panel of three graders adjudicated these 420 images and assessed 241 as mtmDR negative, 179 as mtmDR positive and 135 as vtDR positive. Model’s performance on the clinical validation set is shown in Table 2.

Conclusions : The deep learning model was developed with high quality graded UWF images and performed at a level that highly suggests usefulness in a clinical screening setting. A large, prospective multi-center clinical trial is currently evaluating the performance of a similar model in a real-world clinical setting.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

×
×

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.

×