June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
A novel deep learning approach for identifying outer retinal tubulations in retinal degenerations
Author Affiliations & Notes
  • Amandeep Josan
    Nuffield Department of Clinical Neurosciences, University of Oxford, OXFORD, United Kingdom
  • Footnotes
    Commercial Relationships   Amandeep Josan None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 284. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Amandeep Josan; A novel deep learning approach for identifying outer retinal tubulations in retinal degenerations. Invest. Ophthalmol. Vis. Sci. 2023;64(8):284.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Outer retinal tubulations (ORT) are round hypo-reflective voids with hyper-reflective borders seen in the retinal outer nuclear layer on optical coherence tomography (OCT) b-scans. They are associated with choroideremia and dry age-related macular degeneration (AMD). It has been hypothesised that the hyper-reflective borders of ORT’s represent mitochondria within degenerating rods which have undergone remodelling after loss of outer segments. Here we present a deep learning framework which is able to reliably identify and segment ORT.

Methods : 164 b-scans from a range of choroideremia and AMD patients were manually segmented. An image processing pipeline was created in order to optimize the training images prior to training via a modified U-net deep learning algorithm. Each b-scan was automatically cropped to remove irrelevant regions such as the vitreous and below the choroid. Due to the small size of ORT, the commonly employed process of scaling down of images to allow processing via limited GPU memory for training may lead to loss of precision. Thus we retain high resolution images and avoid GPU memory limitations by splitting images into manageable sized patches. Training is carried out using a U-net architecture with modified skip connections incorporating Bidirectional Long Short-Term Memory (Bi-LSTM).

Results : 100 images were used for training with 64 reserved for testing (32 choroideremia and 32 AMD). A mean dice score of 0.942 was achieved with the patch-based modified U-net architecture compared to 0.871 when using a basic U-net architecture, thus demonstrating a reliable segmentation algorithm with precision comparable to that performed manually by clinicians.

Conclusions : Patch-based modified architecture improves the overall precision in the task of segmentation of ORT when compared to standard U-net approach. Our algorithm is able to segment high resolution OCT images whilst retaining information regarding the global positions of the ORTs. This deep learning framework can be used to rapidly identify ORTs within OCT which may prove to be a useful diagnostic biomarker for RPE loss and photoreceptor stress. Alternatively, the system may serve in interventional clinical trials to monitor disease progression around the edges of outer retinal atrophy.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

×
×

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.

×