Multiple imaging techniques are beneficial in contemporary scleral lens fitting. Advanced scleral technologies now enable practitioners to prescribe fully customized lens designs.¹ Anterior segment optical coherence tomography (AS-OCT) provides high-resolution cross-sectional imaging that allows practitioners to evaluate central and limbal clearance, quantify scleral toricity, and design fully customized scleral lenses.² A study examined 16 unique deep learning (DL) models developed to segment AS-OCT images acquired during scleral lens wear.

AS-OCT images were obtained from 15 healthy adults (mean age 22 ± 1 years) who had normal corneas and best-corrected visual acuity of 0.00 logMAR or better in both eyes. None had a prior history of ocular injury or surgery. An irregular corneal design diagnostic scleral lens was fit with hexafocon A material with a central thickness of 300 µm and an overall diameter of 16.5 mm. Lenses were fitted to achieve an initial post-lens fluid reservoir thickness of 200 µm to 400 µm and were filled with preservative-free saline.

AS-OCT imaging was performed immediately after lens application (T0) and after 8 hours of wear, before lens removal (T8). The instrument provides a digital axial resolution of 4 µm. A high-definition 6 mm radial line scanning protocol was utilized, capturing 12 B-scans, each composed of 10 averaged images centered on the pupil.

An experienced observer manually marked the boundaries of interest in each image (considered the ground truth). These included the anterior and posterior scleral lens surfaces, the anterior corneal epithelial surface, the anterior stromal interface, and the corneal endothelium. Four different architectures were modified for semantic segmentation: U-Net, UNet++, Feature Pyramid Network (FPN), and MA-Net. Each was assessed with 5 different encoders (EfficientNet-B4, DenseNet201, VGG19, ResNet34, and Xception). UNet++ was selected as the benchmark architecture based on its superior performance, demonstrating the highest Dice coefficient and the lowest mean boundary error. Segmentation performance was quantified using the Dice coefficient (to assess area overlap) and the mean absolute boundary error.

Several deep learning-based semantic segmentation models were evaluated as alternatives to the established U-Net model and its variants for AS-OCT image analysis during scleral lens wear. All models demonstrated high levels of accuracy in segmenting the scleral lens, fluid reservoir, corneal epithelium, Bowman’s layer, and corneal stroma. Misclassifications most frequently occurred within the fairly homogeneous layers of the anterior stroma and Bowman’s layer.

The UNet++ model paired with a VGG19 encoder achieved the best result based on the overall Dice coefficient, indicating superior segmentation performance. This model may be used in future studies assessing segmentation accuracy in eyes that had corneal pathology and altered tissue morphology during scleral lens wear.

References

1. Silverman JIM, Huffman JM, Zimmerman MB, Ling JL, Greiner MA. Indications for wear, visual outcomes, and complications of custom imprint 3D scanned scleral contact lens use. Cornea. 2021 May 1;40:596-602. doi: 10.1097/ICO.0000000000002588

2. Marin YG, Alonso-Caneiro D, Collins MJ, Vincent SJ. Evaluation of deep learning models for anterior segment OCT image segmentation during scleral lens wear. Cont Lens Anterior Eye. 2026 Feb;49(1):102484. doi: 10.1016/j.clae.2025.102484

Melissa Barnett, OD, FAAO, FSLS, practices at Baker Optometry. She also is an International Coaching Federation (ICF) associate certified coach. She has received remuneration from ABB, AbbVie, Alcon, Azura, Bausch + Lomb, BCLA, Bruder, Cloudbreak Pharma, Dompé, Epion, OCuSOFT, ScienceBased Health, STAPLE Program, Tarsus, and Théa Pharma.

 This content was supported by Contamac.