Abstract
Purpose :
Previous studies have shown corneal biomechanical properties including central corneal thickness (CCT) and corneal hysteresis (CH) to be associated with risk of glaucoma development and progression. It has been speculated that such association may reflect a relationship between biomechanical properties of the anterior segment (cornea) and those of the posterior segment of the eye, specifically the peripapillary area, optic nerve head, and lamina cribrosa. The purpose of this study was to investigate whether CCT and CH could be predicted from deep learning analysis of optic nerve head scans obtained with enhanced depth imaging spectral domain-optical coherence tomography (EDI SD-OCT).
Methods :
The full dataset included 42,624 optic nerve head (ONH) B-scans of 888 eyes from 476 subjects extracted from the Glaucoma Repository of the Vision, Imaging and Performance (VIP) Laboratory at Duke University. EDI SD-OCT (Heidelberg Engineering; Heidelberg, Germany) images were analyzed. A total of 48 radial scans (each scan 7.5 degrees apart) of the ONH were evaluated for each eye. CH was measured using the ORA device (Reichert; Depew, NY) Deep learning convolutional neural network (ResNet) was applied to each B-scan to predict CH and CCT. The dataset was divided, with 80% of the images utilized for training and validation of the network, while 20% was used to test the network.
Results :
The test set was derived from 178 eyes. Mean measured CCT in the test set was 549.9±40.7 µm, whereas mean predicted CCT from deep learning analysis of ONH scans was 550.9±18.2 µm. There was a statistically significant correlation between observed and predicted CCT (r = 0.402; P<0.001) with a mean absolute deviation error of 30.3µm. Mean measured CH in the test set was 9.76±1.62 mmHg, whereas mean predicted CH from deep learning analysis of ONH scans was 9.74±0.51 mmHg. There was a statistically significant correlation between observed and predicted CH (r = 0.400; P<0.001), with a mean absolute deviation error of 1.18mmHg.
Conclusions :
Deep learning analysis of ONH images was able to predict measurements of CCT and CH. These findings suggest a relationship between CCT and CH with ONH structures, perhaps reflecting the shared role of extracellular matrix components.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.