Abstract
Purpose :
To design a convolutional neural network (CNN) for keratoconus and Fuchs endothelial corneal dystrophy (FECD) detection using optical coherence tomography (OCT) corneal maps.
Methods :
Normal volunteers, keratoconus, and FECD patients were recruited. The central 6mm of the cornea was imaged using a radial OCT scan pattern (Optovue Avanti, visionix). We constructed and optimized a convolutional neural network (CNN) model to classify corneas into 3 categories (keratoconus, FECD, and normal). The model inputs are five OCT maps: pachymetry, epithelial thickness, posterior surface mean curvature, enhanced posterior surface float elevation, and Descemet’s layer and endothelial reflectance (i.e., guttae) maps of the cornea. All maps were down-sampled to 16 × 16 pixels to reduce the number of features to be trained on. Each map type was treated as a different color channel in the neural network. A grid search hyperparameter optimization resulted in a streamlined CNN architecture. The multi-label classification was implemented using two output neurons with sigmoid activation. We chose multi-label output instead of classic softmax output, because it allows the model to make predictions for more than one cause of corneal abnormality at a time. Repeated 5-fold cross-validation (train 70%, validation 10%, and test 20%) was used to evaluate model performance.
Results :
We included 128 eyes of 64 normal, 105 eyes of 78 keratoconus, and 74 eyes of 38 FECD patients in our preliminary study. The CNN model provided excellent accuracy in classifying keratoconus, FECD, and normal eyes. The average accuracy was 98.1 ± 2.6% for manifest keratoconus, 98.7 ± 5.3% with subclinical keratoconus, 95.9 ± 6.0% with FECD, and 96.0 ± 6.0% in normal eyes.
Conclusions :
OCT mapping of the cornea and CNNs can be used to detect keratoconus and FECD with high accuracy. This approach could be expanded to automate the classification of corneal diseases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.