August 2019
Volume 60, Issue 11
Open Access
ARVO Imaging in the Eye Conference Abstract  |   August 2019
Keeping it Clean: Artificial Intelligence Based Denoising Improves Segmentation of Optical Coherence Tomography Images
Author Affiliations & Notes
  • Sripad Krishna Devalla
    Biomedical Engineering, National University of Singapore, Singapore, Singapore
  • Tan Hung Pham
    Biomedical Engineering, National University of Singapore, Singapore, Singapore
    Singapore Eye Research Institute, Singapore, Singapore, Singapore
  • Liang Zhang
    Biomedical Engineering, National University of Singapore, Singapore, Singapore
  • Tun Tin
    Singapore Eye Research Institute, Singapore, Singapore, Singapore
    Biomedical Engineering, National University of Singapore, Singapore, Singapore
  • Rajan Mohan
    Rajan Eye Care Hospital, Chennai, Tamil Nadu, India
  • Sujatha Mohan
    Rajan Eye Care Hospital, Chennai, Tamil Nadu, India
  • Tin Aung
    Singapore Eye Research Institute, Singapore, Singapore, Singapore
  • Leopold Schmetterer
    Singapore Eye Research Institute, Singapore, Singapore, Singapore
    Lee Kong Chian School of Medicine,NTU, Singapore, Singapore
  • Alexandre H. Thiéry
    Statistics and Applied Probability, National University of Singapore, Singapore, Singapore
  • Michael Girard
    Biomedical Engineering, National University of Singapore, Singapore, Singapore
    Singapore Eye Research Institute, Singapore, Singapore, Singapore
  • Footnotes
    Commercial Relationships   Sripad Krishna Devalla, None; Tan Hung Pham, None; Liang Zhang, None; Tun Tin, None; Rajan Mohan, None; Sujatha Mohan, None; Tin Aung, None; Leopold Schmetterer, None; Alexandre Thiéry, None; Michael Girard, None
  • Footnotes
    Support  National University of Singapore Young Investigator Award Grant (NUSYIA_FY13_P03; R-397-000-174-133 [MJAG])
Investigative Ophthalmology & Visual Science August 2019, Vol.60, 027. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Sripad Krishna Devalla, Tan Hung Pham, Liang Zhang, Tun Tin, Rajan Mohan, Sujatha Mohan, Tin Aung, Leopold Schmetterer, Alexandre H. Thiéry, Michael Girard; Keeping it Clean: Artificial Intelligence Based Denoising Improves Segmentation of Optical Coherence Tomography Images. Invest. Ophthalmol. Vis. Sci. 2019;60(11):027.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To develop an artificial intelligence (deep learning) denoising framework to improve segmentation of optical coherence tomography (OCT) images of the optic nerve head (ONH), independently of the device being used.

Methods : Volume scans were acquired through the centre of the ONH using two commercial OCT devices (Spectralis [Heidelberg]:97 B-Scans; Cirrus [Zeiss]: 200 B-Scans) for both the eyes of 1000 subjects. From each volume, noisy (raw) and clean (average of 5 consecutive slices) images were obtained (38,800 Spectralis and 80,000 Cirrus images of each kind in total). A multi-device deep learning network to denoise the OCT images and visually harmonize the device specific characteristics (speckle noise, intensity, contrast, etc.) was developed. The network was trained with 30,000 images (noisy/clean) from each device (one network for each device; shared weights between each network) and tested on the remaining noisy images. The denoising performance was assessed qualitatively and quantitatively using signal-to-noise ratio (SNR) on the unseen images from both the devices. Further, a custom 3D segmentation network was trained on 6 OCT volumes (Spectralis only) to segment 6 ONH tissues. Segmentation quality was then assessed on unseen volumes (for each device) qualitatively (200 volumes) and quantitatively (200 images) using the Dice coefficient (DC; between 0-1; mean of all tissues). To assess the benefits of denoising on segmentation performance, the entire process (training and testing) was repeated on the denoised OCT volumes.

Results : With the proposed multi-device denoising network, we were able to successfully denoise (an increase in the SNR greater than 123% for all cases) the noisy OCT images from both devices (Figure 1 (a)). When trained on the noisy Spectralis volumes, the 3D segmentation network was able to isolate 6 ONH tissues with a DC of 0.87/0.66 (Spectralis/Cirrus) on unseen volumes. Upon denoising, a significant improvement (p<0.05) in the DC (0.93/0.89) was observed (Figure 1 (b)).

Conclusions : We have developed a custom deep learning approach to denoise OCT images from multiple devices simultaneously. Through this process, the network was able to visually harmonize the images from both the devices. We believe this could facilitate the development of clinically robust device-independent deep learning tools for segmentation & diagnosis applications.

This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.

 

×
×

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.

×