June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Deep-learning model to localize biological markers on OCT volumes from weak annotations
Author Affiliations & Notes
  • Javier Gamazo Tejero
    AIMI, Universitat Bern, Bern, Bern, Switzerland
  • Pablo Marquez-Neila
    AIMI, Universitat Bern, Bern, Bern, Switzerland
  • Thomas Kurmann
    AIMI, Universitat Bern, Bern, Bern, Switzerland
    Intuitive Surgical Inc, Sunnyvale, California, United States
  • Mathias Gallardo
    AIMI, Universitat Bern, Bern, Bern, Switzerland
  • Martin Sebastian Zinkernagel
    Inselspital Universitatsspital Bern, Bern, Bern, Switzerland
  • Sebastian Wolf
    Inselspital Universitatsspital Bern, Bern, Bern, Switzerland
  • Raphael Sznitman
    AIMI, Universitat Bern, Bern, Bern, Switzerland
  • Footnotes
    Commercial Relationships   Javier Gamazo Tejero None; Pablo Marquez-Neila None; Thomas Kurmann Intuitive Surgical, Code E (Employment); Mathias Gallardo None; Martin Zinkernagel Bayer, Novartis, Roche, Zeiss, Code C (Consultant/Contractor), Bayer, Boehringer Ingelheim, Code F (Financial Support); Sebastian Wolf Bayer, Boehringer Ingelheim, EarlySigh, Novartis, Roche, Zeiss, Code C (Consultant/Contractor), Bayer, Novartis, Roche, Code F (Financial Support); Raphael Sznitman RetinAI Medical AG, Code I (Personal Financial Interest), RetinAI Medical AG, Code O (Owner)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 1082. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Javier Gamazo Tejero, Pablo Marquez-Neila, Thomas Kurmann, Mathias Gallardo, Martin Sebastian Zinkernagel, Sebastian Wolf, Raphael Sznitman; Deep-learning model to localize biological markers on OCT volumes from weak annotations. Invest. Ophthalmol. Vis. Sci. 2023;64(8):1082.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Recent developments in deep learning have shown success in accurately predicting the location of biological markers in OCT volumes of patients with Age-Related Macular Degeneration (AMD) and Diabetic Retinopathy (DR). This approach has the potential to improve clinical practices and advance medical research. However, producing fine annotations to train these algorithms is burdensome for experts. We propose a method that automatically identifies and assigns biological markers to the ETDRS rings, only requiring B-scan-level presence annotations.

Methods : A neural network was trained using 22,723 OCT B-Scans of 460 eyes (433 patients) with AMD and DR annotated with slice-level labels for Intraretinal Fluid (IRF) and Subretinal Fluid (SRF). The neural network split the B-Scans into 16 columns and predicted the presence probability of the markers in each column. These outputs were mapped into the 1mm, 3mm, and 6mm ETDRS rings. We incorporated the class annotations and domain knowledge into a loss function to constrain the network output with biologically plausible solutions. The method was tested on a set of OCT volumes with 322 eyes (189 patients) with Diabetic Macular Edema, with slice-level SRF and IRF presence annotations for the ETDRS rings. We used the area under the ROC curve and the average precision as evaluation metrics.

Results : Our method accurately predicted the presence of IRF and SRF in each ETDRS ring, outperforming previous baselines even in the most challenging scenarios (Fig. 1). This included rare occurrences of SRF in the outer ETDRS ring. Our model was also successfully applied to en-face marker segmentation and showed consistency within C-scans, despite not incorporating volume information in the training process (Fig. 2). We achieved a correlation coefficient of 0.946 for the prediction of the IRF area.

Conclusions : Weak supervision leads to solid performance in localizing IRF and SRF in OCT volumes, leveraging new perspectives for other biomarkers known to be difficult to segment.

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

 

ROC and precision-recall curves for both markers and rings on the testing dataset. Plots compare our method (solid lines) to partial convolutions (dashed lines).

ROC and precision-recall curves for both markers and rings on the testing dataset. Plots compare our method (solid lines) to partial convolutions (dashed lines).

 

En-face projection of predictions for the specified markers. (a) B-Scan at the location in green. (b) Prediction with our method. (c) Expert pixelwise ground-truth. (d) Expert segmentation rescaled into 16 columns.

En-face projection of predictions for the specified markers. (a) B-Scan at the location in green. (b) Prediction with our method. (c) Expert pixelwise ground-truth. (d) Expert segmentation rescaled into 16 columns.

×
×

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.

×