July 2020
Volume 61, Issue 9
Free
ARVO Imaging in the Eye Conference Abstract  |   July 2020
Bringing Deep Learning Models to the Data: An Application in Recognizing Intra-Retinal Fluid on Optical Coherence Tomography (OCT) Images
Author Affiliations & Notes
  • Khadija Raza
    Tufts Medical Center, Massachusetts, United States
    Boston Latin , Boston, Massachusetts, United States
  • Nihaal Mehta
    Tufts Medical Center, Massachusetts, United States
  • Cecilia Lee
    University Washington, Washington, United States
  • Luisa Mendonca
    Tufts Medical Center, Massachusetts, United States
  • Phillip Braun
    Tufts Medical Center, Massachusetts, United States
  • Jay duker
    Tufts Medical Center, Massachusetts, United States
  • Aaron Lee
    Tufts Medical Center, Massachusetts, United States
  • Footnotes
    Commercial Relationships   Khadija Raza, None; Nihaal Mehta, None; Cecilia Lee, None; Luisa Mendonca, None; Phillip Braun, None; Jay duker, Aldeyra (C), Allegran (C), Aura biosciences (C), Bausch Health (C), Beyeonics (C), Carl Zeiss Meditec (S), Eleven Biotherapeutics (C), Eye Point Pharma (C), Hemera Biosciences (I), Merck (C), Novartis (C), Optovue (S), Roche (C); Aaron Lee, Carl Zeiss Meditec (F), Genentech (C), Microsoft (F), Topcon (R), Verana Health (C)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2020, Vol.61, PB0019. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Khadija Raza, Nihaal Mehta, Cecilia Lee, Luisa Mendonca, Phillip Braun, Jay duker, Aaron Lee; Bringing Deep Learning Models to the Data: An Application in Recognizing Intra-Retinal Fluid on Optical Coherence Tomography (OCT) Images. Invest. Ophthalmol. Vis. Sci. 2020;61(9):PB0019.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Amidst intense interest in deep learning in medicine, concerns regarding data privacy, security, and sharing are of increasing importance. A model-to-data approach, in which the model itself is transferred rather than data, can circumvent many of these challenges, but has not been previously tested in ophthalmology. The objective of our study was to determine whether a model-to-data deep learning approach (i.e. validation of the algorithm without any data transfer) can be successfully applied for the first time to deep learning in ophthalmology.

Methods :
This is a cross sectional study in which a deep learning algorithm model developed at the University of Washington was trained using a model to data approach, to identify intraretinal fluid on OCT scans at the New England Eye Center (NEEC). Patients with active exudative age-related macular degeneration(AMD) imaged at the NEEC between August 2018 and February 2019 were enrolled. OCT images were extracted and labeled. This dataset was then randomly divided into a training dataset (400 eyes) and a testing dataset (70 eyes). The deep learning model was trained in recognizing intra-retinal fluid (IRF) on OCT B-scans. It was then tested using the testing data set and subsequently compared to human grading of intraretinal fluid

Results : The model was successfully trained (learning curve Dice coefficient > 80%) using 400 OCT B-scans. In comparing the model to manual human grading of IRF pockets, with the exception of one comparison, there was no statistically significant difference in Dice or intersection over union scores (P ≥ 0.05).

Conclusions : A model-to-data approach to deep learning was successfully applied for the first time in ophthalmology. Without any data transfer, we successfully trained, validated and tested a previously reported DL model that quantifies the area of intraretinal fluid in OCT. This proof-of-concept suggests that such a paradigm should be further examined in larger-scale, multi-center deep learning studies.

This is a 2020 Imaging in the Eye Conference abstract.

 

Schematic description of the model-to-data approach. The centralized DL model is transferred to the institution housing a data set, allowing for the data to remain within the regulatory bounds of the institution. The trained model can then be transferred to another institution

Schematic description of the model-to-data approach. The centralized DL model is transferred to the institution housing a data set, allowing for the data to remain within the regulatory bounds of the institution. The trained model can then be transferred to another institution

 


Example segmentation of intra-retinal fluid (IRF) by human grader (above) and by the DL algorithm (below)


Example segmentation of intra-retinal fluid (IRF) by human grader (above) and by the DL algorithm (below)

×
×

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.

×