August 2021
Volume 62, Issue 11
Open Access
ARVO Imaging in the Eye Conference Abstract  |   August 2021
3D deep learning algorithm for denoising OCTA volumes acquired at 1.68 MHz A-scan-rate
Author Affiliations & Notes
  • Michael Niederleithner
    Center for Medical Physics and Biomedical Engineering, Medizinische Universitat Wien, Wien, Wien, Austria
  • Anja Britten
    Center for Medical Physics and Biomedical Engineering, Medizinische Universitat Wien, Wien, Wien, Austria
  • Philipp Matten
    Center for Medical Physics and Biomedical Engineering, Medizinische Universitat Wien, Wien, Wien, Austria
  • Niranchana Manivannan
    Carl Zeiss Meditec, Inc., California, United States
  • Aditya Nair
    Carl Zeiss Meditec, Inc., California, United States
  • Lars Omlor
    Carl Zeiss Meditec, Inc., California, United States
  • Rainer Leitgeb
    Center for Medical Physics and Biomedical Engineering, Medizinische Universitat Wien, Wien, Wien, Austria
  • Wolfgang Drexler
    Center for Medical Physics and Biomedical Engineering, Medizinische Universitat Wien, Wien, Wien, Austria
  • Tilman Schmoll
    Carl Zeiss Meditec, Inc., California, United States
  • Footnotes
    Commercial Relationships   Michael Niederleithner, Carl Zeiss Meditec, Inc. (C); Anja Britten, Carl Zeiss Meditec, Inc. (C); Philipp Matten, Carl Zeiss Meditec, Inc. (C); Niranchana Manivannan, Carl Zeiss Meditec, Inc. (E); Aditya Nair, Carl Zeiss Meditec, Inc. (E); Lars Omlor, Carl Zeiss Meditec, Inc. (E); Rainer Leitgeb, Carl Zeiss Meditec, Inc. (F), Carl Zeiss Meditec, Inc. (C); Wolfgang Drexler, Carl Zeiss Meditec, Inc. (F), Carl Zeiss Meditec, Inc. (C); Tilman Schmoll, Carl Zeiss Meditec, Inc. (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science August 2021, Vol.62, 65. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Michael Niederleithner, Anja Britten, Philipp Matten, Niranchana Manivannan, Aditya Nair, Lars Omlor, Rainer Leitgeb, Wolfgang Drexler, Tilman Schmoll; 3D deep learning algorithm for denoising OCTA volumes acquired at 1.68 MHz A-scan-rate. Invest. Ophthalmol. Vis. Sci. 2021;62(11):65.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To produce the best possible capillary contrast, high quality optical coherence tomography angiography (OCTA) scans still rely on more than two repeat measurements. This causes long acquisition times, especially when imaging large fields of view (FOV). We present a three-dimensional (3D) deep learning algorithm to denoise 2-repetition OCTA volumes to realize an image quality so far only accessible with a larger number of measurements.

Methods : We developed a 3D deep learning algorithm for denoising OCTA volumes. The algorithm is based on a 3D U-net network architecture using 32 x 32 x 32 voxel patches. For the OCTA acquisition we used a custom OCT device with a laser source having a sweep rate of 1.68 MHz and a maximum FOV on the retina of up to 18 mm x 18 mm. For creating the training data, we acquired 8 volumes from 3 different healthy subjects at different positions on the retina with a FOV of 9 mm x 9 mm (1024 x 1024 A-scans) with 8 repetitions per slow axis position. For the input data we processed only the first 2 repetitions of a cluster to create low quality OCTA volumes, while we used all 8 repetitions for the high quality target data. After training, the algorithm was used to denoise wide field OCTA volumes with a FOV of 18 mm x 18 mm (2048 x 2048 A-scans). To demonstrate the improvement, we compared the image quality of low quality data sets by using only the first 2 repetitions of a cluster and denoising to the same data set processed with all 4 repetitions, but without denoising.

Results : The algorithm was trained to denoise widefield OCTA volumes with 2 B-scan repetitions per slow scanning position. The resulting volumes in separate test data show significant noise reduction. En face projections of the volume before and after denoising are shown in Figure 1. A comparison of projections of the same dataset processed with only 2 repetitions per cluster and denoised to one processed with all 4 repetitions is shown in Figure 2.

Conclusions : Deep learning algorithms can be used to effectively denoise low quality OCTA volumes and therefore reduce acquisition time significantly. This increases patient comfort drastically and enables larger FOVs in a single shot acquisition.

This is a 2021 Imaging in the Eye Conference abstract.

 

Comparison between before and after applying denoising algorithm to 2-repetition volume (en face projection)

Comparison between before and after applying denoising algorithm to 2-repetition volume (en face projection)

 

Comparison between volume processed with 2 repetitions plus denoising and all 4 repetitions (en face projection)

Comparison between volume processed with 2 repetitions plus denoising and all 4 repetitions (en face projection)

×
×

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.

×