Related Experiment Video
Updated: Sep 12, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Evaluation of deep learning models for anterior segment OCT image segmentation during scleral lens wear
Yoel Garcia Marin1, David Alonso-Caneiro2, Michael J Collins1
1Contact Lens and Visual Optics Laboratory, Optometry and Vision Science, Centre for Vision and Eye Research, Queensland University of Technology, Brisbane, Queensland, Australia.
Purpose:
The accurate segmentation of corneal and contact lens boundaries in anterior segment optical coherence tomography (AS-OCT) images provides essential clinical information. The purpose of this study was to evaluate the performance of sixteen different deep learning (DL) models developed to segment AS-OCT images obtained during scleral lens wear.
Methods:
AS-OCT images were obtained from 15 participants with normal corneas after 0 and 480 min of scleral lens wear. An experienced observer manually annotated the boundaries of interest in each image (considered the ground truth) including the anterior and posterior scleral lens surfaces, the anterior corneal epithelial surface, the anterior stromal interface, and the endothelium. Four different architectures were adapted for semantic segmentation (U-Net, U-Net++, FPN, and MA-Net) each of which was tested with five different encoders (EfficientNet-B4, DenseNet201, VGG19, ResNet34, and Xception). Following training, the segmentation performance of each model was evaluated using the Dice coefficient (measurement of the area overlap) and the mean absolute boundary error.
Results:
All DL models displayed a high level of performance for classification and segmentation of the scleral contact lens and fluid reservoir (with Dice coefficients typically > 99 % and mean absolute error values of < 1 pixel). Misclassification issues arose for some models, likely linked to the lower reflectivity and homogeneity of the interface between Bowman's layer and the corneal stroma. Overall, only minor differences were observed between models, with the U-Net++/VGG19 combination displaying the best performance with an overall Dice score of 99.26 % and per class Dice scores ranging from 99.11 to 99.77 %.
Conclusion:
The U-Net++/VGG19 DL model displayed the best performance for AS-OCT image segmentation during scleral lens wear based on the overall Dice coefficient. Further assessment of DL models involving the segmentation of eyes with corneal disease and altered tissue morphology during scleral lens wear is warranted.
