June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Deep learning (DL) to segment retinal layer disruption on optical coherence tomography (OCT) in neovascular age-related macular degeneration (nAMD)
Author Affiliations & Notes
  • Yun Li
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Huanxiang Lu
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Thomas Albrecht
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Andreas Maunz
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Ales Neubert
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Fethallah Benmansour
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Alvaro Gomariz
    F. Hoffmann-La Roche Ltd., Basel, Switzerland
  • Jennifer Luu
    Genentech Inc, South San Francisco, California, United States
  • Daniela Ferrara
    Genentech Inc, South San Francisco, California, United States
  • Footnotes
    Commercial Relationships   Yun Li Roche, Inc. , Code E (Employment); Huanxiang Lu Roche, Inc. , Code E (Employment); Thomas Albrecht Roche, Inc. , Code E (Employment); Andreas Maunz Roche, Inc. , Code E (Employment); Ales Neubert Roche, Inc. , Code E (Employment); Fethallah Benmansour Roche, Inc. , Code E (Employment); Alvaro Gomariz Roche, Inc. , Code E (Employment); Jennifer Luu Genentech, Inc., Code E (Employment); Daniela Ferrara Genentech, Inc., Code E (Employment), Roche, Inc., Code I (Personal Financial Interest)
  • Footnotes
    Support  Yes, F. Hoffmann-La Roche Ltd., Basel, Switzerland, provided support for the study and participated in the study design; conducting the study; and data collection, management, and interpretation.
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2079 – F0068. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Yun Li, Huanxiang Lu, Thomas Albrecht, Andreas Maunz, Ales Neubert, Fethallah Benmansour, Alvaro Gomariz, Jennifer Luu, Daniela Ferrara; Deep learning (DL) to segment retinal layer disruption on optical coherence tomography (OCT) in neovascular age-related macular degeneration (nAMD). Invest. Ophthalmol. Vis. Sci. 2022;63(7):2079 – F0068.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Accurate segmentation of retinal layers offers a wealth of information to clinicians and scientists in terms of diagnostic and predictive biomarkers. Disruption of certain retinal layers is key to determining visual outcomes for patients with nAMD. Accurate quantification of such features remains a time-consuming task for retinal specialists and is challenging to automate. This work proposes an effective DL solution to segment the external limiting membrane (ELM) and ellipsoid zone (EZ) with and without disruptions on OCT scans in eyes with nAMD.

Methods : Our dataset contains OCT images acquired with Cirrus HD-OCT (Carl Zeiss Meditec) from patients enrolled in the HARBOR clinical trial (NCT00891735). Manual annotations denoting ELM, EZ, retinal pigment epithelium (RPE), and Bruch's membrane (BM) as layer boundary elevation maps were performed on a subset of the dataset (989 B-scans from 170 patients).

A DL model with UNet architecture is trained on 90% of the dataset using generalized Dice loss. Schematic of the end-to-end pipeline is shown in Figure 1. Performance is validated on the test set (10% of dataset) with (1) Chamfer distance (CD) for measuring distance in layer boundary between the ground truth and segmented result, (2) layer Dice coefficients of retinal layers using argmax of the segmented output, and (3) 1-dimensional (1D) disruption Dice coefficients for measuring overlap of the linearized layer disruption area between the ground truth and segmentation results. An example of how 1D disruption Dice score is calculated on a single B-scan is shown in Figure 2. Results are reported in mean ± SD.

Results : Evaluated on the test set, model performances to segment ELM and EZ, respectively, are as follows: CD, 17.67 ± 13.22, 13.00 ± 9.35; layer Dice coefficient, 0.56 ± 0.16, 0.60 ± 0.15; 1D disruption Dice coefficient, 0.62 ± 0.38, 0.62 ± 0.37.

Conclusions : Our proposed end-to-end DL solution segments disruptions in the ELM and EZ with reasonable accuracy. This provides clinicians and scientists with an automated tool to quickly generate layer disruption features, which may benefit patient management. The proposed 1D disruption Dice score presents a metric designed to capture the model’s ability to accurately predict retinal layers with disruption.

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

 

Figure 1. Schematic of End-to-End DL Pipeline

Figure 1. Schematic of End-to-End DL Pipeline

 

Figure 2. Schematic of 1D Disruption Dice Score

Figure 2. Schematic of 1D Disruption Dice Score

×
×

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.

×