June 2020
Volume 61, Issue 7
Free
ARVO Annual Meeting Abstract  |   June 2020
Denoising Optical Coherence Tomography Images Using Conditional Generative Adversarial Networks
Author Affiliations & Notes
  • Dewei Hu
    Electrical Engineering and Computer Science, Vanderbilt University, Nashville, Tennessee, United States
  • Yigit Atay
    Electrical Engineering and Computer Science, Vanderbilt University, Nashville, Tennessee, United States
  • Joseph Malone
    Biomedical Engineering, Vanderbilt University, Nashville, Tennessee, United States
  • Yuankai Tao
    Biomedical Engineering, Vanderbilt University, Nashville, Tennessee, United States
  • Ipek Oguz
    Electrical Engineering and Computer Science, Vanderbilt University, Nashville, Tennessee, United States
  • Footnotes
    Commercial Relationships   Dewei Hu, None; Yigit Atay, None; Joseph Malone, None; Yuankai Tao, None; Ipek Oguz, None
  • Footnotes
    Support  R01 EY030490, The SyBBURE Searle Undergraduate Research Program, R01-NS094456, Vanderbilt Discovery Program
Investigative Ophthalmology & Visual Science June 2020, Vol.61, 2028. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Dewei Hu, Yigit Atay, Joseph Malone, Yuankai Tao, Ipek Oguz; Denoising Optical Coherence Tomography Images Using Conditional Generative Adversarial Networks. Invest. Ophthalmol. Vis. Sci. 2020;61(7):2028.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Optical coherence tomography (OCT) is widely used in ophthalmology, dermatology and cardiology. One of the shortcomings of OCT is that it is prone to speckle noise. Conditional generative adversarial networks (cGAN) show promise for denoising OCTs of healthy retinas. The goal of our study is to develop a cGAN that can be generalized to OCT images of both healthy and pathological anatomy without having to re-train specifically for pathology.

Methods : To achieve our goal, both healthy and pathological retinas are included in training dataset. Focal laser-induced injury was performed to 9 mice who were then imaged under anesthesia using a custom-built spectral domain OCT system with 107dB SNR at 125kHz line-rate and 2.08um axial resolution. The sampling was 4096*500*500pix. with 5 repeated frames at each position for a total of 2500 frames. ‘Ground truth’ was obtained by averaging the five repeated frames. To generalize the relativistic cGAN, we introduce horizontal and vertical edge loss functions for our generator to improve edge preservation in any orientation. This proved to be essential as pathological retinas contain edges in various orientations around lesions or deformities, unlike healthy retinas that typically consist of mostly horizontal tissue layers.

Results :
Our experiments show promising results that successfully preserve edges and fine-level details of the anatomy while reducing speckle for both healthy (Figure 1) and pathological (Figure 2) anatomy. Furthermore, our synthesized images out-perform the ‘truth’ images in normalized cross-correlation (0.95 vs. 0.88), structural similarity index (0.47 vs. 0.39), and contrast-to-noise ratio (3.03 vs. 2.52), although we recommend supplementing these metrics with the qualitative assessment of the actual images.

Conclusions : Our results illustrate the strong performance of our method, which produces visually plausible images. Generalizability of the algorithm is improved by better preservation of arbitrarily oriented edges.

This is a 2020 ARVO Annual Meeting abstract.

 


Figure 1: Healthy mouse B-scan. Upper left: Noisy input image. Upper right: Ground truth denoised image. Bottom: Denoised image with our method


Figure 1: Healthy mouse B-scan. Upper left: Noisy input image. Upper right: Ground truth denoised image. Bottom: Denoised image with our method

 


Figure 2: Mouse B-scan with a retinal lesion. Upper left: Noisy input image. Upper right: Ground truth denoised image. Bottom: Denoised image with our method


Figure 2: Mouse B-scan with a retinal lesion. Upper left: Noisy input image. Upper right: Ground truth denoised image. Bottom: Denoised image with our method

×
×

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.

×