June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Variance and bias of artificial intelligence based denoising and conventional averaging in optical coherence tomography angiography
Author Affiliations & Notes
  • Koenraad Arndt Vermeer
    Rotterdam Ophthalmic Institute, Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Lisette Smid
    Rotterdam Ophthalmic Institute, Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Mirjam EJ Van Velthoven
    Rotterdam Eye Hospital, Rotterdam, Netherlands
  • Footnotes
    Commercial Relationships   Koenraad Vermeer, Canon (F); Lisette Smid, Canon (F); Mirjam Van Velthoven, Canon (F)
  • Footnotes
    Support  Rotterdamse Stichting Blindenbelangen
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 1773. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Koenraad Arndt Vermeer, Lisette Smid, Mirjam EJ Van Velthoven; Variance and bias of artificial intelligence based denoising and conventional averaging in optical coherence tomography angiography. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1773.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To examine the effect of artificial intelligence (AI) based denoising and conventional averaging on the reproducibility and bias of quantitative measurements, i.e., vessel density (VD) and foveal avascular zone (FAZ), derived from optical coherence tomography angiography (OCT-A) en face images.

Methods : In this retrospective cohort study, the non-pathological fellow eyes of patients with unilateral non-systemic disorders were included. OCT-A scans of two visits were analyzed on a Canon OCT-HS100 workstation. The VD and FAZ were measured on the first single scan (SS), best single scan (BSS), averaged scan (AS), and those scans after AI-denoising (SS-AI, BSS-AI, AS-AI). The reproducibility was calculated as the mean absolute difference (MAD) between the two visits, and the bias as the mean value (MV). A two-factor repeated-measures ANOVA for both the MAD and MV of the FAZ and VD was performed introduced by the factors AI-denoising and averaging.

Results : We analyzed 16 non-pathological eyes of 16 patients (61.9 ± 17.5 years). For VD, the reproducibility of AI-denoised scans (SS-AI, AS-AI) was significantly worse than scans without AI-denoising (SS, AS), but significantly better in BSS-AI than in BSS (Figure 1). AI-denoising also negatively affected the reproducibility of the FAZ measurement (Figure 2). AI-denoising and averaging both introduced a significant bias for the VD; for the FAZ, a significant bias was only introduced by AI-denoising. The marginal effects and results of the pairwise comparisons are presented in Table 1.

Conclusions : AI-denoising negatively affect the reproducibility and introduces a bias for quantitative OCT-A measurements, while the reproducibility of averaged scans was not significantly different from single scans. Only the VD measurement was biased after averaging. When the original single image is of high quality, the AI-denoising technology has a good reproducibility for VD. Otherwise, we recommend using conventional averaging for quantitative comparisons.

This is a 2021 ARVO Annual Meeting abstract.

 

 

×
×

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.

×