June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Applying nnU-Net to the segmentation of microaneurysms in OCTA data in patients with non-proliferative diabetic retinopathy
Author Affiliations & Notes
  • Lennart Husvogt
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany
    Research Laboratory for Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
  • Omar Abu-Qamar
    Tufts Medical Center, Boston, Massachusetts, United States
  • Emily Levine
    Tufts Medical Center, Boston, Massachusetts, United States
  • Julia Schottenhamml
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany
  • Stefan B Ploner
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany
  • Katharina Breininger
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany
  • Eric Moult
    Research Laboratory for Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
  • Nadia K Waheed
    Tufts Medical Center, Boston, Massachusetts, United States
  • James G Fujimoto
    Research Laboratory for Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
  • Andreas K Maier
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Germany
  • Footnotes
    Commercial Relationships   Lennart Husvogt, None; Omar Abu-Qamar, None; Emily Levine, None; Julia Schottenhamml, None; Stefan Ploner, VISTA-OCTA (P); Katharina Breininger, None; Eric Moult, VISTA-OCTA (P); Nadia Waheed, ;Nidek Medical Products, Inc (F), Carl Zeiss Meditec (F), Macula Vision Research Foundation (F), Optovue Inc (C), Topcon Medical Systems, Inc (F); James Fujimoto, Carl Zeiss Meditec Inc (P), Optovue Inc (I), Optovue Inc (P), Optovue Inc (C), Topcon Medical Systems (R), VISTA-OCTA (P); Andreas Maier, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2152. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Lennart Husvogt, Omar Abu-Qamar, Emily Levine, Julia Schottenhamml, Stefan B Ploner, Katharina Breininger, Eric Moult, Nadia K Waheed, James G Fujimoto, Andreas K Maier; Applying nnU-Net to the segmentation of microaneurysms in OCTA data in patients with non-proliferative diabetic retinopathy. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2152.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Diabetic retinopathy (DR) is a leading cause of vision loss in working age populations worldwide. Patients may not notice any symptoms during the early non-proliferative stage (NPDR) while microaneurysms (MA) can still be detected using fluorescein angiography or optical coherence tomography angiography (OCTA). The purpose of this contribution is to automatically detect MA from OCTA en face images to aid in the early detection and diagnosis of NPDR.

Methods : Patients were enrolled at the New England Eye Center at Tufts Medical Center in Boston. We collected data from 90 eyes (70 patients) with NPDR who were imaged using the OptoVue Avanti system using a field size of 3x3 mm. The system software automatically performs a segmentation of the retinal layers. The OCTA en face images of the superficial capillary layer were used. Two expert graders at the Boston Imaging Reading Center manually labeled MA in these images using custom software. The labeled data were divided into 73 training and 16 testing images. We then trained a 2D nnU-Net to segment the MA. nnU-Net is a self-configuring deep learning tool that can automatically generate 2D and 3D U-nets and uses heuristic and data-based rules to choose suitable hyper-parameters. nnU-Net runs a five-fold cross-validation on the training data, which allows to use the resulting five trained nets as an ensemble.

Results : 12 of the 16 test images contained MA. The number of expert-labeled MA was 21. 11 of those MA were correctly found by nnU-Net. There were 4 false positives, but no false MA were detected on the 4 test images without labeled MA.
Figure 1 shows four of the test images. The top row shows the images without, the bottom row with marked aneurysms. The images show undetected MA (false negatives) marked in red, correctly detected areas in green and false positives in pink. Sub-figure A shows one detected aneurysm and one false negative. B shows two correctly detected MA. C shows a false positive in the center and a not correctly detected aneurysm in the upper image. Both are located directly near a saccade. D shows a correctly identified MA.

Conclusions : nnU-Net detected 11 out of 21 labeled MA. Future work will focus on improving these results by optimizing loss, data augmentation and training of the generated U-Net.

This is a 2021 ARVO Annual Meeting abstract.

 

Four test images. True positives are shown in green, false negatives in red and false positives in pink.

Four test images. True positives are shown in green, false negatives in red and false positives in pink.

×
×

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.

×