June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
3-D style transfer between structure and flow channels in OCT angiography
Author Affiliations & Notes
  • Gwenole Quellec
    LaTIM, UMR 1101, INSERM, France
  • Yihao Li
    LaTIM, UMR 1101, INSERM, France
    Universite de Bretagne Occidentale, Brest, Bretagne, France
  • Hassan Al Hajj
    LaTIM, UMR 1101, INSERM, France
    Universite de Bretagne Occidentale, Brest, Bretagne, France
  • Sophie Bonnin
    Hopital Rothschild, Paris, Île-de-France, France
  • Hugang Ren
    Carl Zeiss Meditec Inc, Dublin, California, United States
  • Niranchana Manivannan
    Carl Zeiss Meditec Inc, Dublin, California, United States
  • Stephanie Magazzeni
    Carl Zeiss Meditec Inc, Dublin, California, United States
  • Ramin Tadayoni
    Hopital Lariboisiere, Paris, Île-de-France, France
  • Pierre-Henri Conze
    IMT Atlantique Bretagne-Pays de la Loire - Campus de Brest, Brest, Bretagne, France
    LaTIM, UMR 1101, INSERM, France
  • Mathieu Lamard
    Universite de Bretagne Occidentale, Brest, Bretagne, France
    LaTIM, UMR 1101, INSERM, France
  • Footnotes
    Commercial Relationships   Gwenole Quellec None; Yihao Li None; Hassan Al Hajj None; Sophie Bonnin None; Hugang Ren Carl Zeiss Meditec Inc, Code E (Employment); Niranchana Manivannan Carl Zeiss Meditec Inc, Code E (Employment); Stephanie Magazzeni Carl Zeiss Meditec Inc, Code E (Employment); Ramin Tadayoni Carl Zeiss Meditec Inc, Code C (Consultant/Contractor); Pierre-Henri Conze None; Mathieu Lamard None
  • Footnotes
    Support  ANR grant ANR-18-RHU-008
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2989 – F0259. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Gwenole Quellec, Yihao Li, Hassan Al Hajj, Sophie Bonnin, Hugang Ren, Niranchana Manivannan, Stephanie Magazzeni, Ramin Tadayoni, Pierre-Henri Conze, Mathieu Lamard; 3-D style transfer between structure and flow channels in OCT angiography. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2989 – F0259.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : OCT angiography (OCTA) is a promising imaging modality for improving the management of diabetic retinopathy (DR). The EviRed project aims to develop artificial intelligence tools to automate DR diagnosis and prognosis evaluation using this imaging among others. Given the large amount of data to process, we investigate the potential redundancy of the two 3-D channels in OCTA scans, namely structure and flow. In particular, can any of these two channels be predicted by the other?

Methods : High-definition OCTA scans (6x6mm) from 112 eyes of 61 diabetic patients (one scan per eye) were acquired with a PLEX® Elite 9000 (Carl Zeiss Meditec Inc., Dublin, CA, USA). DR severity level in each eye was graded by a retina specialist using fundus photographs. Scans were randomly distributed between three subsets with equal distributions of DR severity: 84 scans (45 patients) to train, 12 scans (6 patients) to validate and 16 scans (10 patients) to test. Next, various 3-D U-Nets were trained to predict the flow channel from the structure channel. A similar set of 3-D U-Nets was trained to predict the structure channel from the flow channel.

Results : Using the ICDR scale, 29 eyes had no DR, 28 had mild NPDR, 26 had moderate NPDR and 29 had more advanced DR; the test set contained 4 scans from each severity category. Based on the validation set, the best neural architecture was 3-D U-Net with a ResNet-152 backbone. In the test set, the mean absolute error (MAE), in 3-D, between the true and predicted flow was 4.19% and the MAE between the true and predicted structure was 4.11%. The Pearson correlation coefficient (PCC) between the true and predicted flow was 0.684 and the PCC between the true and predicted structure was 0.961. No significant Spearman correlation was found between the PCC and DR severity (p-value = 0.721 for both transfer tasks).

Conclusions : Flow and structure channels in 6x6mm PLEX® Elite OCTA scans can be predicted from one another, which indicates potential redundancy between these channels. Prediction performance does not depend on DR severity. Visually, we observe that pathological features (e.g., exudates in the structure channel, avascular zones in the flow channel) can be also predicted. Further investigations are needed to quantify the prediction accuracy, at voxel level, and evaluate the possibility of using one single OCTA 3D channel for DR management in the future.

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

 

Example of real and predicted OCTA scan.

Example of real and predicted OCTA scan.

×
×

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.

×