June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Instance Segmentation of Reticular Pseudodrusen (RPD) in Eyes with Intermediate Age-related Macular Degeneration
Author Affiliations & Notes
  • Yelena Bagdasarova
    Ophthalmology, University of Washington, Seattle, Washington, United States
  • Himeesh Kumar
    Centre for Eye Research Australia, Royal Victoria Eye and Ear Hospital, University of Melbourne, East Melbourne, Victoria, Australia
    Ophthalmology, Department of Surgery, University of Melbourne, Melbourne, Victoria, Australia
  • Robyn H Guymer
    Centre for Eye Research Australia, Royal Victoria Eye and Ear Hospital, University of Melbourne, East Melbourne, Victoria, Australia
    Ophthalmology, Department of Surgery, University of Melbourne, Melbourne, Victoria, Australia
  • Cecilia S Lee
    Ophthalmology, University of Washington, Seattle, Washington, United States
  • Aaron Y Lee
    Ophthalmology, University of Washington, Seattle, Washington, United States
  • Zhichao Wu
    Centre for Eye Research Australia, Royal Victoria Eye and Ear Hospital, University of Melbourne, East Melbourne, Victoria, Australia
    Ophthalmology, Department of Surgery, University of Melbourne, Melbourne, Victoria, Australia
  • Footnotes
    Commercial Relationships   Yelena Bagdasarova None; Himeesh Kumar None; Robyn Guymer Roche, Genentech, Novartis, Bayer, Apellis, Code C (Consultant/Contractor); Cecilia Lee None; Aaron Lee Genentech, Johnson and Johnson, Verana Health, Gyroscope, Code C (Consultant/Contractor), US Food and Drug Administration, Code E (Employment), Santen, Microsoft, Carl Zeiss Meditec, Novartis, NVIDIA, Code F (Financial Support), Topcon, Code R (Recipient); Zhichao Wu None
  • Footnotes
    Support  NIH/NIA R01AG060942, NIH/NEI K23EY029246, NIH/NIA U19AG066567, Latham Vision Grant, Research to Prevent Blindness Unrestricted Core Grant, Karalis Johnson Retina Center
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2056 – F0045. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Yelena Bagdasarova, Himeesh Kumar, Robyn H Guymer, Cecilia S Lee, Aaron Y Lee, Zhichao Wu; Instance Segmentation of Reticular Pseudodrusen (RPD) in Eyes with Intermediate Age-related Macular Degeneration. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2056 – F0045.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Reticular Pseudodrusen (RPD) is a critical phenotype in age-related macular degeneration. We trained a deep learning model to reliably segment RPD on optical coherence tomography (OCT) B-scans in participants with intermediate AMD.

Methods : OCT volume scans from 120 participants from the Laser Intervention in the Early Stages of AMD Study were used for model development and internal testing. A total of 60 eyes with RPD and 180 eyes without RPD (with 49 scans per eye) were split at the participant level into 6 folds; 5 folds were used for cross-validation model training and fold 6 was reserved for testing. Each of 5 models consisted of a Mask-RCNN head with ResNeXt-101-32x8d-FPN backbone. The final model is a soft-voting ensemble of the 5 models.

Results : The instance-level Precision-Recall curve for the test fold, which contains 10 eyes with RPD and 30 eyes without RPD, is plotted in Figure 1a. At the instance-level the ensemble achieved precision of 0.64 [95% CI: 0.61,0.67], recall of 0.69 [0.66,0.72], and F1 of 0.66 [0.64,0.68] at the Intersection over Union (IoU) threshold of 0.2 and model probability threshold of 50%. The false positive rate (FPR), defined as the number of negative scans for which at least one RPD was predicted, was 4.4% [3.5%,5.5%]. The predicted vs actual RPD count per scan is plotted in Figure 1b.

For classifying RPD at the eye level, the ensemble achieved precision of 0.56 [0.33, 0.75], 0.83 [0.55,0.95], and 1.00 [0.72,1.00] when thresholding on 1 or more, 5 or more, and 10 or more RPD instances per eye, respectively, in the model prediction of the test fold. The ensemble achieved a recall of 1.00 [0.72,1.00] at those instance thresholds.

Conclusions : We successfully trained a model to segment RPD instances in B-scans and have validated the model on an internal test set.

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

 

A) Instance-level Precision vs Recall curve for the test fold at various IoU thresholds. B) Standard boxplot of number of predicted RPD instances per scan grouped by number of ground truth instances per scan. IoU and probability thresholds are set at 0.2 and 50% respectively.

A) Instance-level Precision vs Recall curve for the test fold at various IoU thresholds. B) Standard boxplot of number of predicted RPD instances per scan grouped by number of ground truth instances per scan. IoU and probability thresholds are set at 0.2 and 50% respectively.

 

Example of RPD instance segmentation. Top: original, Middle: ground truth annotation. Bottom: model prediction with probability.

Example of RPD instance segmentation. Top: original, Middle: ground truth annotation. Bottom: model prediction with probability.

×
×

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.

×