June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
A robust, flexible retinal segmentation algorithm designed to handle neuro-degenerative disease pathology (NDD-SEG)
Author Affiliations & Notes
  • Rahele Kafieh
    Centre for Transformative Neuroscience and Institute of Biosciences, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
    School of Advanced Technologies in Medicine, Medical Image and Signal Processing Research Center, Isfahan University of Medical Sciences, Isfahan, Isfahan, Iran (the Islamic Republic of)
  • Sajed Rakhshani
    School of Advanced Technologies in Medicine, Medical Image and Signal Processing Research Center, Isfahan University of Medical Sciences, Isfahan, Isfahan, Iran (the Islamic Republic of)
  • Jeffry Hogg
    Royal Victoria Infirmary Eye Department, Newcastle Upon Tyne Hospitals NHS Foundation Trust, Newcastle Upon Tyne, Newcastle upon Tyne, United Kingdom
    Translational and Clinical Research Institute, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Rachael A Lawson
    Clinical Ageing Research Unit, Newcastle upon Tyne Hospitals NHS Trust and Newcastle University, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
    Translational and Clinical Research Institute, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Nicola Pavese
    Clinical Ageing Research Unit, Newcastle upon Tyne Hospitals NHS Trust and Newcastle University, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
    Translational and Clinical Research Institute, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Will Innes
    Royal Victoria Infirmary Eye Department, Newcastle Upon Tyne Hospitals NHS Foundation Trust, Newcastle Upon Tyne, Newcastle upon Tyne, United Kingdom
    Translational and Clinical Research Institute, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • David H Steel
    Sunderland Eye Infirmary, Sunderland, Tyne and Wear, United Kingdom
    Centre for Transformative Neuroscience and Institute of Biosciences, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Jaume Bacardit
    School of Computing, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Jenny Read
    Centre for Transformative Neuroscience and Institute of Biosciences, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Anya Hurlbert
    Centre for Transformative Neuroscience and Institute of Biosciences, Newcastle University, Newcastle upon Tyne, Tyne and Wear, United Kingdom
  • Footnotes
    Commercial Relationships   Rahele Kafieh None; Sajed Rakhshani None; Jeffry Hogg None; Rachael A Lawson None; Nicola Pavese None; Will Innes None; David Steel None; Jaume Bacardit None; Jenny Read None; Anya Hurlbert None
  • Footnotes
    Support  UK NIHR AI_AWARD01976 / Newcastle BRC
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2080 – F0069. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Rahele Kafieh, Sajed Rakhshani, Jeffry Hogg, Rachael A Lawson, Nicola Pavese, Will Innes, David H Steel, Jaume Bacardit, Jenny Read, Anya Hurlbert; A robust, flexible retinal segmentation algorithm designed to handle neuro-degenerative disease pathology (NDD-SEG). Invest. Ophthalmol. Vis. Sci. 2022;63(7):2080 – F0069.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Retinal Optical Coherence Tomography (OCT) has potential for early detection of neurodegenerative disease (NDD). This requires accurate, automatic segmentation of retinal boundaries despite challenges often pronounced in images from people with NDD, e.g. atrophic layers and indistinct boundaries, media opacity, and movement artefacts. An ideal algorithm would also handle different OCT devices and imaging protocols with diverse resolution, size, artefacts and noise. Our algorithm (NDD-SEG) gives a solution.

Methods : The basic structure is a U-net without pooling and upsampling layers. An initial Retinal Tissue Segmentation Network (RTSN) identifies the retina within the image, even against a noisy background. The Retinal Layer Segmentation Network (RLSN) focuses on fine segmentation of the intra-retinal layers, using the probability map from RTSN cascaded with the original input image. This probability map and texture features are injected into decoder levels of RLSN. The Self-Attention Transformer Block represents higher-level descriptions of the merged data instead of simple concatenation. A new Boundary Preservation Loss (BPL) function is introduced which enables high-precision layer edges to feed into a classification loss function. Together, texture awareness and precision targeting of edges make the network robust to noise.

Results : From 111 OCT volumes from healthy controls and patients with multiple sclerosis (doi:10.1016/j.dib.2018.12.073 and doi:10.1155/2015/259123), 23 are randomly selected for testing and the remainder for training. Reported results are based on two metrics: the mean absolute difference (MAD) for 9 boundaries (in pixels) and Dice's coefficient for 8 layers as distance/similarity between the predicted boundaries/layers and ground truth. The MAD values and Dice scores are presented in Table 1. Size-independence is demonstrated by a < 1% variation in Dice score between segmentation outputs for the original vs. 5 size-distorted images (mean Dice 93.4%). For challenging Parkinson’s OCT images (N=44) from an untrained dataset (Newcastle PDD-SRC), NDD-SEG outputs were expert-rated as useable on 54% vs. 1% for a comparator U-net algorithm (Doi:978-3-319-24574-4_28), and more accurate on 100%. See Figure 1 for examples.

Conclusions : Our new method achieves size-independence and robust segmentation even in the presence of image artefacts and pathologies.

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

 

 

×
×

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.

×