Purchase this article with an account.
Sripad Krishna Devalla, Jean-Martial Mari, Tin A Tun, khai sing chin, NIcholas Strouthidis, Tin Aung, Alexandre H. Thiéry, Michael J A Girard; A Device-Independent Deep Learning Approach to Digitally Stain Optical Coherence Tomography Images of the Optic Nerve Head. Invest. Ophthalmol. Vis. Sci. 2018;59(9):3500. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a device-independent artificial-intelligence (deep learning) approach to digitally stain neural and connective tissues in optical coherence tomography (OCT) images of the optic nerve head (ONH).
A horizontal B-scan was acquired through the center of the ONH for 1 eye of each of 100 subjects (40 healthy; 60 glaucoma) using 3 commercially-available OCT devices (Spectralis [Heidelberg], Cirrus [Zeiss], DRI [Topcon]). All images were preprocessed using compensation (to improve deep-tissue visibility), histogram equalization (to harmonize intensities across devices) and resampling (to bring all images to the same isotropic resolution: 4 x 4 µm). A deep learning algorithm (custom U-NET) that was able to capture both the local tissue information (texture) and contextual information (spatial arrangement of tissues) was trained with each device to digitally stain 6 tissue layers of the ONH from any device. Due to small data size, we generated additional images through data augmentation (e.g. image deformations, application of noise, intensity shifts). In each case, digital staining accuracy was assessed against manual segmentation. Performance of digital staining across devices, the effect of compensation and performance comparison between glaucoma and healthy subjects were studied.
When trained with images from 1 device, our algorithm was able to isolate neural and connective tissues in unseen OCT images from any of the 3 devices. On average, accuracies where always higher than 90% for any given ONH tissue layer, for any device, irrespective of the device that was used for training (Figure 1). Irrespective of the device used for training, there was no significant difference in digital staining accuracy (p>0.05). The performance of digital staining was significantly better when trained with compensated images (p<0.05). Finally, there was no significant difference in digital staining performance between healthy and glaucoma images (p>0.05).
We offer a device-independent solution to digitally stain neural and connective tissues of the ONH. Our work suggests that data from 1 OCT device can be used to interpret images from any other device. Our work offers a framework to automatically measure multiple key structural parameters of the ONH (irrespective of the OCT device) that may be critical to improve glaucoma management.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only