June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Unsupervised OCT Denoising using speckle split
Author Affiliations & Notes
  • Julia Schottenhamml
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
    Department of Ophthalmology and Eye Hospital, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
  • Tobias Würfl
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
  • Stefan B Ploner
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
  • Lennart Husvogt
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
    Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
  • Bettina Hohberger
    Department of Ophthalmology and Eye Hospital, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
  • James G Fujimoto
    Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
  • Andreas Maier
    Pattern Recognition Lab, Friedrich-Alexander-Universitat Erlangen-Nurnberg, Erlangen, Bayern, Germany
  • Footnotes
    Commercial Relationships   Julia Schottenhamml None; Tobias Würfl None; Stefan Ploner IP related to VISTA-OCTA, Code P (Patent); Lennart Husvogt None; Bettina Hohberger None; James Fujimoto Optovue, Code C (Consultant/Contractor), Topcon, Code F (Financial Support), Optovue, Code I (Personal Financial Interest), Optovue, Code P (Patent), Carl Zeiss Meditec, Code P (Patent), IP related to VISTAOCTA, Code P (Patent); Andreas Maier None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 182 – F0029. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Julia Schottenhamml, Tobias Würfl, Stefan B Ploner, Lennart Husvogt, Bettina Hohberger, James G Fujimoto, Andreas Maier; Unsupervised OCT Denoising using speckle split. Invest. Ophthalmol. Vis. Sci. 2022;63(7):182 – F0029.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Optical coherence tomography (OCT) scans are often degraded by noise, which complicates further analysis. We demonstrate an unsupervised denoising method that uses OCT angiography (OCTA) information to keep the informative speckle used for OCTA computation and removes flow-unrelated speckle noise. Qualitative and user-study analyses demonstrate noise removal and improved sharpness compared to other unsupervised denoising methods.

Methods : In a first step, a U-Net is trained via unsupervised Noise2Void (N2V) to denoise OCTA B-scans. In a second step, a second U-Net is trained via N2V on OCT B-scans. A constraint in the loss function ensures that the OCTA image computed from two denoised OCT B-scans is similar to the denoised OCTA B-scans from the first step. The workflow of this algorithm is depicted in Figure 1. The training/test set consisted of 90/18 measurements (with 500 B-scans per volume and 2 volumes per patient) from 30/9 patients. Field sizes included 3x3 mm and 6x6 mm areas around the fovea and included various pathologies (choroidal neovascularization, diabetes mellitus without diabetic retinopathy (DR), non-proliferative DR, proliferative DR, early age-related macular degeneration, geographic atrophy, non-arteritic anterior ischemic optic neuropathy and healthy controls). We compared our method (speckle split N2V (SSN2V)) to other unsupervised denoising algorithms (N2V (without the constraint), BM3D, TV, WNNM). A user study was performed on the test set. One expert ophthalmologist graded 35 randomly selected B-scans and the results of the denoising algorithms from 0 (does not like the image) to 5 (likes the image).

Results : The user study shows that the clinical expert preferred the results produced by our proposed algorithm (input: 1.11±0.622; BM3D: 2.2±0.92; TV: 2.74±0.73; WNMM: 3.37±0.93; N2V: 1.51±0.77; SSN2V: 4.09±0.87). Qualitative results of the results of the different denoising algorithms are shown in Figure 2.

Conclusions : Using the additional constraint for the unsupervised OCT denoising improves the denoising capability and performs superior to the compared methods in an expert user study.

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

 

Workflow of the proposed algorithm. ‘l’ denotes the N2V loss function for the respective input image.

Workflow of the proposed algorithm. ‘l’ denotes the N2V loss function for the respective input image.

 

Qualitative examples from the denoising results of the different algorithms.

Qualitative examples from the denoising results of the different algorithms.

×
×

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.

×