June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Artificial intelligence based analysis of retinal color photo and angiographic changes in retinopathy of prematurity
Author Affiliations & Notes
  • Jason Patrick Day
    Ophthalmology, University of Florida, Gainesville, Florida, United States
  • Amarjeet Kumar
    University of Florida, Gainesville, Florida, United States
  • Chelsea-Jane Arcales
    University of Florida, Gainesville, Florida, United States
  • Deja Blunt
    University of Florida, Gainesville, Florida, United States
  • Nazanin Ebrahimiadib
    Ophthalmology, University of Florida, Gainesville, Florida, United States
  • Wei Shao
    University of Florida, Gainesville, Florida, United States
  • Jinghua Chen
    Ophthalmology, University of Florida, Gainesville, Florida, United States
  • Footnotes
    Commercial Relationships   Jason Day None; Amarjeet Kumar None; Chelsea-Jane Arcales None; Deja Blunt None; Nazanin Ebrahimiadib None; Wei Shao None; Jinghua Chen None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 4930. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jason Patrick Day, Amarjeet Kumar, Chelsea-Jane Arcales, Deja Blunt, Nazanin Ebrahimiadib, Wei Shao, Jinghua Chen; Artificial intelligence based analysis of retinal color photo and angiographic changes in retinopathy of prematurity. Invest. Ophthalmol. Vis. Sci. 2023;64(8):4930.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : The goal of this study is to develop an artificial intelligence (AI) algorithm based on fluorescein angiography (FA) and color fundus images to predict the need for treatment in infants with retinopathy of prematurity (ROP).

Methods : All available color fundus photos, FA images and clinical data of the babies with ROP at the University of Florida (UF) Shands Hospital between 2012 and 2022 will be reviewed. The AI model was developed using convolutional neural network (CNN - based on VGG16 architecture) for feature extraction of images. The results were concatenated with the corresponding medical data, e.g., gestational age and birth weight. Finally, a dense neural network was utilized to predict the patient's ROP zone, stage and plus status. See Image 1.

Results : The AI model was trained with images and data from eight infants (16 eyes) and then tested on 3 infants (6 eyes).

The model has high accuracy to predict the zone, stage and plus status. The prediction accuracy is best when using fundus photographs and fluorescein angiographies together compared to only using one type of image input. The model can predict the zone and stage of ROP with 83.33% accuracy and the plus status with 100% accuracy. See Table 1.

Next, study staff will include color fundus photos, FA images and clinical data from 62 additional infants into the AI prediction model for further training. The AI prediction model will then undergo validation noting statistical outcomes for the prediction of ROP zone, stage and plus status.

Conclusions : Our results have demonstrated that we have developed an AI algorithm based on fundus imaging, fluorescein angiography and demographic data which has high diagnostic performance for detection of infants who require treatment for ROP. The accuracy of the model will continue to improve as trains and learns from more subjects. This AI model can be used to help improve diagnostic accuracy in neonatal intensive care units around the world. Further versions of the algorithm will be able to detect progression or regression of disease following ROP treatment.

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.

×