Abstract
Purpose :
To report a deep segmentation network for the segmentation of the air-epithelium (Epi), epithelium Bowman’s (Bw), and the endothelium-aqueous (En) interfaces of pathological corneas in optical coherence tomography (OCT) images.
Methods :
One hundred and twenty OCT images were obtained randomly from 67 eyes of 53 patients. Eye scans were obtained using OCT machine (Envisu R2210, Bioptigen, Buffalo Grove, IL). The images include 40 images for eyes with keratoconus, 40 images for dry eyes, 40 images with Fuchs’ endothelial dystrophy, and 40 images for normal controls. The images were manually segmented by a trained operator. The images were randomly divided into 96 images for training, 12 images for validation, and 12 images for testing. A deep U-shape network was developed with multi-channel output (i.e., one channel for each interface). The network was trained on the training data and evaluated using the validation and testing data. The outliers in the network segmentation were removed using a random sample consensus method. The network prediction accuracy between the output segmentation and the manual segmentation was computed. The mean and standard deviation of the absolute errors in pixels were computed between the network segmentation and the manual segmentation. Examples of segmented OCT images are shown in Fig. 1.
Results :
The prediction accuracies of the segmentation network were 99.88%, 99.79%, 99.80%, for the training, validation, and testing data. The mean and standard deviation of the absolute errors between the network and the operator were 0.33±0.36, 0.32±0.33, and 0.34±0.37 pixels for the Ep, Bw, and En interfaces, respectively, on the training data. The mean and standard deviation of the absolute errors between the network and the operator were 0.54±0.50, 0.62±0.75, and 0.43±0.46 pixels for the Ep, Bw, and En interfaces, respectively, on the validation data. The mean and standard deviation of the absolute errors between the network and the operator were 0.52±0.57, 0.53±0.71, and 0.41±0.41 pixels for the Ep, Bw, and En interfaces, respectively, on the testing data.
Conclusions :
The developed segmentation network can reliably segment three interfaces in OCT images of pathological as well as healthy corneas with errors less than one pixel as compared to the manual segmentation.
This is a 2020 ARVO Annual Meeting abstract.