An Open-Source Deep Learning Framework for Automated Corneal Segmentation in Anterior Segment Optical Coherence...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate corneal segmentation is crucial for diagnosing and monitoring eye diseases using anterior segment optical coherence tomography (AS-OCT).
• Existing segmentation models often struggle with performance variability across different OCT devices.

Purpose of the Study:

• To develop and evaluate the Cornea nnU-Net Extractor (CUNEX), a deep learning model for full-thickness corneal segmentation in AS-OCT images.
• To assess CUNEX's utility and generalizability in artificial intelligence (AI) research across diverse OCT platforms.

Main Methods:

• CUNEX was trained on a large AS-OCT dataset and validated internally and externally on images from four different OCT devices.
• The model's performance was benchmarked against established models (CorneaNet, ScLNet).
• The impact of CUNEX-based segmentation on downstream AI classification tasks (age, sex, disease staging) was evaluated.

Main Results:

• CUNEX achieved high segmentation accuracy (Dice Similarity Coefficient [DSC] 94-95%, Intersection over Union [IoU] 88-91%) on internal validation.
• The model demonstrated superior cross-device generalizability, maintaining meaningful segmentations (DSC/IoU >70%) where other models failed.
• Segmentation minimally impacted classification accuracy, except for sex prediction, suggesting sex-related features are extracameral.

Conclusions:

• CUNEX is the first open-source AS-OCT corneal segmentation model validated across multiple independent OCT platforms.
• It provides a reproducible and robust foundation for integrating corneal segmentation into clinical and AI research.
• The availability of code and pretrained weights facilitates its adoption in the research community.