Abstract
Purpose :
To report a multi-disease deep convolutional neural network of common corneal disease (keratoconus and Fuch’s endothelial dystrophy) classification using OCT corneal thickness, curvature, and guttae maps.
Methods :
Keratoconic eyes had topographic signs of keratoconus (e.g., asymmetric bowtie with skewed radial axis, or central or inferior steep zone). Fuch’s patients were diagnosed using the simplified slit-lamp edema scale (grades 1-3) and the modified Krachmer guttae scale (grades 0-6). Normal controls had a normal slit-lamp examination, normal topography map appearance, and corrected distance visual acuity of 20/20 or better. A spectral-domain OCT (Avanti, Visionix/Optovue) was used to acquire 6 mm wide images of the central cornea. Custom algorithms were used to generate pachymetry, epithelial thickness, posterior surface mean curvature, enhanced posterior surface float elevation, and posterior surface signal intensity (i.e., guttae) maps of the cornea. All maps were downsampled to a size of 16×16 pixels, and each map type was treated as a different color channel in the neural network. A multi-label classification approach was implemented using two output neurons with sigmoid activation. Classification accuracy was computed for different model architectures using repeated 5-fold cross-validation. Within each fold, the data was split into the train (70%), validation (10%), and test (20%) sets.
Results :
58 normal, 111 keratoconic, and 39 Fuch’s dystrophy eyes were involved in the cross-validation analysis. The convolutional neural network with the best performance consisted of three convolutional and max pooling layers followed by four dense layers, including the output layer. Cross-validation was repeated five times. The overall average accuracy was 98.0% (95% confidence interval 96.9% to 98.8%) with balanced accuracy of 98.9% for keratoconus, 99.2% for Fuch’s, and 96.9% for normal.
Conclusions :
Our deep convolutional neural network can accurately detect irregular corneas and distinguish between keratoconus and Fuch’s endothelial dystrophy. It can potentially be extended to create a comprehensive classification system for corneal diseases.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.