Purchase this article with an account.
Kerry Jayne Halupka, Hiroshi Ishikawa, Matthew Lee, Gadi Wollstein, Joel Schuman, Simon Wail, Bhavna Josephine Antony; “One-size fits all” OCT image enhancement via deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1513. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Convolutional neural networks (CNN) are promising image enhancing tools, but often do not generalise well. The purpose of this study was to test the performance of a denoising CNN, trained on optic nerve head (ONH) scans then applied to macular scans.
A CNN was trained in our previous work to denoise b-scans from OCT ONH volumes (Cirrus HD-OCT, Zeiss, Dublin, CA) of healthy eyes using either (1) mean-squared error (CNN-MSE), or (2) a generative adversarial network with Wasserstein distance and perceptual similarity (CNN-WGAN). The two approaches respectively favour pixel-based similarity and perceptual similarity. To test their generalizability, OCT scans were obtained on the macular region of 10 healthy and 10 glaucoma subjects at a range of signal strengths (SS, ranged 6-10). Specifically, OCT scans were attained from 2 subjects for each group at each of the SS levels, and 10 evenly-distributed b-scans from each volume were denoised using the two pre-trained CNNs, resulting in 200 b-scans. A trained observer was shown the three versions of each B-scan (raw image, CNN-MSE and CNN-WGAN) and asked to choose the image which most clearly showed inner retinal structures.
All images showed notable improvement in quality at all signal strengths, though CNN-MSE resulted in visibly smoother images than CNN-WGAN, while the latter showed more detail in layer textures (Fig. 1). Qualitative results suggest that CNN-MSE more clearly showed inner retinal structures compared to both raw b-scans and CNN-WGAN, and this was consistent for all signal strengths (Fig. 2). All structural features were retained, particularly the foveal dip and the photoreceptor bands. Furthermore, no features or artefacts were introduced by the networks.
The deep learning networks are region-independent and perform well on macular scans despite being trained on healthy optic nerve head scans regardless of the SS. This method shows promise as a clinically useful OCT post-processing tool.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Central slice of healthy (a-c) and Glaucomatous (d-f) macular OCT volumes with signal strength of 7 (a,d), and the result of processing with CNN-MSE (b,e) and CNN-WGAN (c,f).
Qualitative test results showing the number of images of each type perceived to most clearly show inner retinal structures, for 40 b-scans of varying signal strengths.
This PDF is available to Subscribers Only