June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
A deep learning system for sickle cell retinopathy detection using retinal OCT images from children with sickle cell disease
Author Affiliations & Notes
  • Jing Jin
    Ophthalmology/Surgery, Nemours Children's Hospital Delaware, Wilmington, Delaware, United States
  • Ashuta Bhattarai
    Computer and Information Sciences, University of Delaware, Newark, Delaware, United States
  • Robin Miller
    Hematology/Oncology, Nemours Children's Hospital Delaware, Wilmington, Delaware, United States
  • Edward Kolb
    Hematology/Oncology, Nemours Children's Hospital Delaware, Wilmington, Delaware, United States
  • Chandra Kambhamettu
    Computer and Information Sciences, University of Delaware, Newark, Delaware, United States
  • Footnotes
    Commercial Relationships   Jing Jin None; Ashuta Bhattarai None; Robin Miller None; Edward Kolb None; Chandra Kambhamettu None
  • Footnotes
    Support  This work is supported by an Institutional Development Award (IDeA) under grant number U54-GM104941, award number 2P20GM109021 and 5P20 GM103446-21 from the National Institute of General Medical Sciences of the National Institutes of Health.
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 3280 – A0332. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jing Jin, Ashuta Bhattarai, Robin Miller, Edward Kolb, Chandra Kambhamettu; A deep learning system for sickle cell retinopathy detection using retinal OCT images from children with sickle cell disease. Invest. Ophthalmol. Vis. Sci. 2022;63(7):3280 – A0332.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Retinal damage in sickle cell disease (SCD) begins with vascular occlusion by sickled red blood cells. Inner retinal thinning due to tissue volume loss is the most common finding in OCT retinal images of SCD patients. Inner retinal thinning from neuronal migration is one of the characteristics of normal fovea. This project aims to develop and validate a deep learning system to detect retinal thickness changes due to sickle cell retinopathy (SCR) using retinal OCT from children with SCD.

Methods : We use a grid-based object detection algorithm, You Only Look Once (YOLO), to detect retinal thinning caused by SCR. Our dataset contains 3906 B-scans from 63 OCT studies of 33 SCD patients (14 male, 21 SS, 9 SC, 2 Sβ+ and 1 Sβ0, age 12.99±4.56, 5.35 to 20.26 years). Instances of SCR and fovea were annotated by an ophthalmologist using bounding boxes on each image. We divided the dataset into 5 sets, each containing approximately equal numbers of B-scans, to perform a 5-fold cross validation. The algorithm was trained 5 time, and each time on different 4 sets and tested on the remaining set using a Tesla V100 Graphical Processing Unit. We measured the training loss using Binary Cross-Entropy and Focal loss. Mean Average Precision (mAP) based on Intersection over Union between annotated and predicted bounding boxes was used as a performance metric.

Results : Our trained model achieved an average mAP of 96% and 72% on fovea and SCR detection, respectively. Figure 1 presents the corresponding confusion matrix. Low mAP on SCR detection can be attributed to the ambiguity and inconsistency in annotating SCR instances, as depicted by the Background column in Figure 1. Figure 2 compares SCR and fovea detection made by the trained model verses annotated images.

Conclusions : In this examination of retinal OCT images from children with SCD, a deep learning system has shown good precision in identifying SCR and differentiating disease injuries from similarly structured profiles of the normal fovea. Our future research combines the current approach with computer vision-based analysis of retinal layer thickness, extending our work towards predicting SCR progression by analyzing multiple OCT scans of re-visiting patients.

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

 

Figure 1: Confusion matrix illustrating summary of predicted results

Figure 1: Confusion matrix illustrating summary of predicted results

 

Figure 2: Left: Annotated images. Right: Predicted from the trained model with confidence scores

Figure 2: Left: Annotated images. Right: Predicted from the trained model with confidence scores

×
×

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.

×