Abstract
Purpose :
To develop an automated classification system using artificial intelligence to distinguish the clinically unaffected eyes in patients with keratoconus from a normal control population based on the combination of Scheimpflug camera and Spectral-Domain OCT imaging data.
Methods :
In total, 78 eyes from 78 participants had previously been classified by 2 cornea experts as either normal eyes (30 eyes), keratoconus (28 eyes) or subclinical keratoconus (20 eyes). All eyes were imaged with a Scheimpflug camera and a Spectral-Domain OCT. Convolutional neural networks were used to extract corneal features from the imaging data of these two machines. Random forest classifier was used to train a model based on these features to distinguish the subclinical keratoconus from normal eyes. Fisher score was used to rank the differentiable power of each parameter.
Results :
Among all individual features, the maximum localized thinning in the vertical meridian in corneal epithelium extracted from OCT ranked the first one to differentiate the subclinical keratoconus from normal eyes. The combination of Spectral-Domain OCT features reached a higher differentiable power (AUC = 0.94, sensitivity = 93%, specificity = 94%) than Schimpflug camera (AUC = 0.86). The developed classification model to combine all features from the two machines dramatically improved the differentiable power to discriminate between normal and subclinical keratoconus eyes (AUC = 1.00 with 100% sensitivity and 100% specificity).
Conclusions :
The automated classification system using artificial intelligence based on the combination of Scheimpflug camera and Spectral-Domain OCT imaging data showed very good performance to discriminate the subclinical keratoconus eyes from normal eyes. The epithelial features extracted from OCT images played an important role in the discrimination. This classification system has the potential to improve the differentiable power of subclinical keratoconus and the efficiency of keratoconus screening.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.