June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Optimal registration of noisy retinal autofluoresence images with cloud computing
Author Affiliations & Notes
  • Jonathan Young
    Novai ltd, United Kingdom
  • James Owler
    Novai ltd, United Kingdom
  • Richard E. Daws
    Novai ltd, United Kingdom
  • Natalie Pankova
    Novai ltd, United Kingdom
  • Francesca Cordeiro
    Novai ltd, United Kingdom
    Institute of Opthalmology, University College London, London, London, United Kingdom
  • Footnotes
    Commercial Relationships   Jonathan Young Novai Ltd, Code E (Employment); James Owler Novai Ltd, Code E (Employment); Richard Daws Novai Ltd, Code E (Employment); Natalie Pankova Novai Ltd, Code E (Employment); Francesca Cordeiro Novai Ltd, Code E (Employment), Novai Ltd, Code P (Patent)
  • Footnotes
    Support  Innovate UK smart grant 10028366
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 2375. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jonathan Young, James Owler, Richard E. Daws, Natalie Pankova, Francesca Cordeiro; Optimal registration of noisy retinal autofluoresence images with cloud computing. Invest. Ophthalmol. Vis. Sci. 2023;64(8):2375.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Registration of retinal images is a key step both in longitudinal studies and for the detection of apoptosing retinal cells (DARC) technology. However, DARC requires near infrared autofluoresence (NIRAF) images. NIRAF images can be challenging to register due to high levels of noise or poor contrast. Here we performed a large grid search to compare denoising methods and find optimal registration parameters to maximize the accuracy of NIRAF image registration.

Methods : 193 human NIRAF image pairs acquired from the same eyes and sessions were registered together. For each pair, a rigid plus affine transformation was computed using a 3-stage coarse to fine-grained scheme. A grid search was performed via Amazon Web Services's (AWS) Batch feature to tractably search combinations of the learning rate, shrink factor, and smoothing radius parameters for the registration process and find the optimum. This process was repeated with 3 image denoising methods (none, total variation (TV) regularisation and bilateral filtering (BF) ) as preprocessing steps. The normalized cross correlation (NCC) between registered images was used as a metric for registration accuracy and compared between denoising methods with a one-way analysis of variance (ANOVA).

Results : A significant effect on NCC was found for the denoising type (one-way ANOVA: F(2, 573)=150.49, p<0.001 ). A post hoc pairwise Tukey test showed that all pairs of denosing types differed significantly (p<0.001). Denoising clearly improved registration, with TV denoising provided the best results (figure 1). Visualising the grid search (figure 2) suggests that the following parameters are optimal for NIRAF registration (TV denoising, shrink factors [12,8,4], smoothing [5,2,2] pixels, learning rate = 0.1) producing an average of approximately 0.8 NCC.

Conclusions : Precise alignment between images is critical for accurate analyses of NIRAF images. The large registration parameter space makes local processing challenging. We demonstrate that registration of retinal NIRAF images can be improved by TV regularisation and that fine-tuning a large set of registration parameters is tractable with AWS.

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

 

Figure 1: boxplots of variation in registration accuracy by denoising method. All differences are statistically significant (****p<0.001).

Figure 1: boxplots of variation in registration accuracy by denoising method. All differences are statistically significant (****p<0.001).

 

Figure 2: Grid search results for mean NCC and success rate. Denoising is with TV and learning rate = 0.1.

Figure 2: Grid search results for mean NCC and success rate. Denoising is with TV and learning rate = 0.1.

×
×

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.

×