June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
The Impact of Deep Learning Architecture on Fluid Segmentation and Classification on SD-OCT
Author Affiliations & Notes
  • Sydney Sterben
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Jon Whitney
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Duriye Damla Sevgi
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Jordan M. Bell
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Jenna Hach
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Sunil K. Srivastava
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Jamie Reese
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Justis P Ehlers
    Cleveland Clinic Cole Eye Institute, Cleveland, Ohio, United States
  • Footnotes
    Commercial Relationships   Sydney Sterben, None; Jon Whitney, None; Duriye Damla Sevgi, None; Jordan Bell, None; Jenna Hach, None; Sunil K. Srivastava, Abbvie (C), Allergan (F), Allergan (C), Eyepoint (F), Eyepoint (C), Eyevensys (F), Eyevensys (C), Gilead (C), Leica (P), Novartis (C), Regeneron (F), Regeneron (C), Santen (F), Zeiss (C); Jamie Reese, None; Justis Ehlers, Adverum (C), Aerpio (F), Aerpio (C), Alcon (F), Alcon (C), Allegro (C), Allergan (F), Allergan (C), Boehringer-Ingelheim (F), Genentech (F), Genentech/Roche (C), Leica (C), Leica (P), Novartis (F), Novartis (C), Regeneron (F), Regeneron (C), Santen (C), Stealth (C), Thrombogenics/Oxurion (F), Thrombogenics/Oxurion (C), Zeiss (C)
  • Footnotes
    Support  RPB Cole Eye Institutional Grant, NIH/NEI K23-EY022947
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2127. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Sydney Sterben, Jon Whitney, Duriye Damla Sevgi, Jordan M. Bell, Jenna Hach, Sunil K. Srivastava, Jamie Reese, Justis P Ehlers; The Impact of Deep Learning Architecture on Fluid Segmentation and Classification on SD-OCT. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2127.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose :
Manual fluid segmentation in SD-OCT is tedious and subject to variability. This study compared machine learning (ML) model architecture performance in fluid segmentation and classification to understand the optimal model architecture for use in automated segmentation.

Methods :
A deep convolutional neural network was tested modifying parameters such as batch normalization, kernel size (10x10 compared to a 5x5), and dilation (rate = 2 compared to rate = 0). The training set was composed of 1369 SD-OCT slices for classification, and 1593 slices for segmentation and included retinal disorders, such as diabetic macular edema and age-related macular degeneration. The segmentation model detected “all fluid” whereas the classification model categorized the type of fluid [e.g., intraretinal fluid (IRF), subretinal fluid (SRF)] present in ground truth segmented “generic” fluid. Performance was evaluated using F-scores.

Results :
Segmentation F-scores ranged from 0.00 (when batch normalization not performed) to 0.72. The optimal model architecture for segmentation was the 5x5-kernel, with batch normalization, and no dilation. This may be caused by reduced training data dilution when using smaller kernels. Classification F-scores varied from 0.78 to 0.87. The optimal classification architecture used 5-kernel with batch normalization and a dilation rate of 2. This 5x5-kernel model took the same amount of spatial context as a 10x10-kernel model, using fewer nodes. This efficiently incorporates contextual information from relative location and retinal features.

Conclusions :
Model architecture impacts fluid segmentation and classification performance. Optimizing architecture to incorporate contextual information efficiently is key to high performance in spatial problems with limited datasets.

This is a 2021 ARVO Annual Meeting abstract.

 


Figure 1: Dilation decreased model precision and small kernels increased confidence and segmentation sharpness.


Figure 1: Dilation decreased model precision and small kernels increased confidence and segmentation sharpness.

 

Figure 2: In this example the smaller kernel model had a higher degree of confidence. Blue is IRF and green is SRF.

Figure 2: In this example the smaller kernel model had a higher degree of confidence. Blue is IRF and green is SRF.

×
×

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.

×