August 2019
Volume 60, Issue 11
Open Access
ARVO Imaging in the Eye Conference Abstract  |   August 2019
OCT image noise reduction using deep learning without additional priors
Author Affiliations & Notes
  • Arindam Bhattacharya
    Carl Zeiss Meditec, Dublin, California, United States
  • Mary K Durbin
    Carl Zeiss Meditec, Dublin, California, United States
  • Footnotes
    Commercial Relationships   Arindam Bhattacharya, Carl Zeiss Meditec (E); Mary Durbin, Carl Zeiss Meditec (E)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science August 2019, Vol.60, PB092. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Arindam Bhattacharya, Mary K Durbin; OCT image noise reduction using deep learning without additional priors. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB092.

      Download citation file:

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

  • Supplements

Purpose : Optical coherence tomography (OCT) imaging quality is often limited by various noise sources, which may hinder the ability to visualize fine tissue features. Recent advances in deep learning techniques have made it possible to perform effective image noise reduction. Such methods based on image translation require prior knowledge such as higher-resolution images or pre-defined noise characteristics. We propose a formulation of the noise reduction problem based on the U-net architecture which does not require additional prior knowledge by directly learning from the input images.

Methods : A CIRRUS™5000 HD-OCT with AngioPlex® OCT Angiography (ZEISS, Dublin, CA) was used to acquire 3x3mm scans (245 B-scans, 245 A-scans/B-scan and 1024 pixels/A-scan). Each B-scan location was repeated 4 times. We trained a custom convolutional neural network on pairs of images from B-scans from the same scan. A pre-processing step registered the four B-scans within a cluster, and each cluster of registered B-scans was used to generate three pairs of training sets. Existing approaches using averaged images as priors suffer from ‘blurring’ caused by imperfect registration. Those instead using additive noise are biased towards the noise statistic added. The proposed network instead only learns to denoise the content noise characteristics.
Our network is a modified five-layer U-net architecture with deep supervision. A custom loss function enforces a symmetric loss between training images. We also use a linear combination of L1 and L2 loss and deep supervision for faster convergence and accurate results.

Results : Fig 1 shows the results of one of our test cases. First column shows single input slice and the corresponding output slice. The volume rendering shows the rendering of 50 slices (1024x245x50). The cropped images show closeups of regions of interest. The network learns anisotropic feature preserving smoothing.
Fig 2 shows the histogram along the yellow line (one A-scan) for a single original slice in blue and network output in red. Following the curves, we see that the network decreases the noise without deteriorating the peaks representing the signal ranges.

Conclusions : We present a robust feature preserving denoising method which can automatically learm the characteristic noise in OCT data without additional information.

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.