Purchase this article with an account.
Elias Pavlatos, David Huang, Yan Li; Combining OCT Corneal Topography and Thickness Maps to Diagnose Keratoconus Using a Convolutional Neural Network. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2109 – F0098.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To design a convolutional neural network (CNN) for keratoconus detection using optical coherence tomography (OCT) corneal topography and thickness maps.
Normal subjects (n = 52) and patients with either keratoconus (n = 131) or contact lens-related warpage (n = 20) were recruited. Keratoconus eyes were divided into 3 groups: 1) Manifest (n = 89): slit-lamp or topographic signs of keratoconus and corrected distance visual acuity (CDVA) < 20/20, 2) Subclinical (n = 16): topographic signs of keratoconus but CDVA ≥ 20/20, and 3) Forme fruste (n = 26): normal-appearing eye with keratoconus in the contralateral eye. The central 6mm of the cornea was imaged using a radial OCT scan pattern (Avanti, Optovue Inc.), and maps of pachymetry, epithelial thickness, anterior surface mean curvature, and posterior surface mean curvature were generated (Li et al, Ophthalmology, 2012; Pavlatos et al, BOE, 2020; Figure 1). All maps were down-sampled to a size of 16×16 pixels. The 4 map types were each treated as different color channels, and a grid search was used to optimize the network architecture and hyperparameters. Binary classification was performed to separate the keratoconus cases and non-keratoconus cases (normal or contact lens warpage). Repeated 5-fold cross-validation was used to evaluate model performance, and class activation maps were generated.
The average balanced accuracy of the CNN during cross-validation was 94 ± 2%. The precision and recall were 98 ± 3% and 91 ± 4%, respectively. The area under the receiver operating characteristic curve was 0.94 ± 0.02. The network was able to detect 100% of the manifest and subclinical keratoconus cases. The accuracy for the forme fruste keratoconus cases was 56 ± 19%. The network demonstrated good specificity, with 97 ± 6% of normal eyes and 96 ± 9% of warpage eyes being classified as non-keratoconus cases. The class activation maps indicated that the regions of the topography and thickness maps which contained the keratoconic cone were most important to the CNN for disease detection (Figure 2).
OCT mapping of the cornea and CNNs can be used to detect keratoconus with high accuracy. This approach could be expanded to automate the classification of corneal diseases.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
Fig 1. OCT scan pattern, image segmentation, and maps for a manifest keratoconus eye.
Fig 2. Average class activation maps for each group.
This PDF is available to Subscribers Only