June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Deep learning algorithms for detection of diabetic macular edema requiring treatment from color fundus photographs
Author Affiliations & Notes
  • Tien-En Tan
    Singapore National Eye Centre, Singapore
    Singapore Eye Research Institute, Singapore
  • Yi Pin Ng
    Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR), Singapore
  • Claire Calhoun
    Jaeb Center for Health Research, Tampa, Florida, United States
  • Xinxing Xu
    Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR), Singapore
  • Liu Yong
    Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR), Singapore
  • Rick SM Goh
    Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR), Singapore
  • Gavin Tan
    Singapore National Eye Centre, Singapore
    Singapore Eye Research Institute, Singapore
  • Jennifer K Sun
    Joslin Diabetes Center, Boston, Massachusetts, United States
    Department of Ophthalmology, Harvard Medical School, Boston, Massachusetts, United States
  • Daniel SW Ting
    Singapore National Eye Centre, Singapore
    Singapore Eye Research Institute, Singapore
  • Footnotes
    Commercial Relationships   Tien-En Tan None; Yi Pin Ng None; Claire Calhoun None; Xinxing Xu None; Liu Yong None; Rick Goh None; Gavin Tan None; Jennifer Sun Adaptive Sensory Technologies, Boehringer Ingelheim, Genentech/Roche, Janssen, Physical Sciences, Inc, Novo Nordisk, Optovue, Code F (Financial Support); Daniel Ting None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 2648. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Tien-En Tan, Yi Pin Ng, Claire Calhoun, Xinxing Xu, Liu Yong, Rick SM Goh, Gavin Tan, Jennifer K Sun, Daniel SW Ting; Deep learning algorithms for detection of diabetic macular edema requiring treatment from color fundus photographs. Invest. Ophthalmol. Vis. Sci. 2023;64(8):2648.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Diabetic retinopathy (DR) screening, whether by human graders or deep learning algorithms (DLAs), is generally performed with color fundus photographs (CFP). However, this modality is suboptimal for detection of diabetic macular edema (DME), and has poor specificity with a high false positive referral rate. Furthermore, not all DME requires treatment. Guidelines recommend treatment only for center-involved DME (CI-DME), with best-corrected visual acuity (BCVA) of 20/32 or worse. We aimed, therefore to develop DLAs for detection of DME requiring treatment from CFPs.

Methods : We used a dataset consisting of diabetic patients with or without DME from DRCR Retina Network clinical trials, who had optical coherence tomography (OCT), CFP, and BCVA from the same visit, with only 1 eye per participant included. Separate DLAs were trained to detect CI-DME (based on OCT with machine- and sex-specific thresholds), and eyes with BCVA of 20/32 or worse. The dataset was split into training, validation, and test sets in a 60:20:20 ratio. Images from 2,763 eyes and 2,694 eyes across 8 protocols were used for the CI-DME and BCVA DLAs respectively. Output from these 2 DLAs were then combined to detect “DME requiring treatment”, which was defined as CI-DME with BCVA of 20/32 or worse.

Results : Overall, 46% of eyes had CI-DME, 35% had BCVA of 20/32 or worse, and 24% had DME requiring treatment. For CI-DME detection, areas under the receiver operating characteristic curves (AUCs) in the validation and test sets were 0.850 (95% CI 0.814-0.882) and 0.839 (0.803-0.874) respectively. For BCVA classification, AUCs in the validation and test sets were 0.809 (95% CI 0.764-0.851) and 0.790 (0.743-0.834) respectively. When the models were combined for detection of DME requiring treatment, the AUCs were 0.861 (95% CI 0.823-0.896) and 0.831 (0.792-0.864) in the validation and test sets respectively. At a threshold of 0.0577, sensitivity and specificity for detection of DME requiring treatment were 84.6% and 73.5% in the validation dataset, and 83.0% and 67.1% in the test dataset.

Conclusions : These CFP-based DLAs are potentially useful, not only for detecting DME patients requiring treatment in DR screening programs, but also for identifying those who may be eligible for future DME clinical trials. Future studies will be required to prospectively validate these AI models.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

×
×

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.

×