What question did this study set out to answer?

The aim is to create and validate a deep learning model for accurate corneal segmentation using AS-OCT images.

April 12, 2026Open Access

An Open-Source Deep Learning Framework for Automated Corneal Segmentation in Anterior Segment Optical Coherence Tomography With Cross-Device External Validation

Key Points

The aim is to create and validate a deep learning model for accurate corneal segmentation using AS-OCT images.
Developed CUNEX model using a dataset of AS-OCT scans from 37,499 patients.
Selected 100 images each from normal, keratoconus, and Fuchs dystrophy cases for training.
Performed external validation using AS-OCT images from four different devices.
Benchmarking CUNEX against other established models, CorneaNet and ScLNet.
CUNEX achieved Dice similarity coefficients of 94%-95% on internal validation set.
External tests showed CUNEX maintained high performance across devices with DSC and IoU >70%.
Segmentation impact on classification accuracy was minimal, though sex prediction accuracy dropped from 81% to 68%.

Abstract

Purpose: To develop a deep learning model—Cornea nnU-Net Extractor (CUNEX)—for full-thickness corneal segmentation of anterior segment optical coherence tomography (AS-OCT) images and evaluate its utility in artificial intelligence research. Methods: CUNEX was developed and evaluated using our Moorfields AS-OCT dataset of 194,599 scans from 37,499 patients, from which 100 normal, 100 keratoconus, and 100 Fuchs endothelial corneal dystrophy eyes were selected for model training. External validation was performed without fine-tuning using AS-OCT images acquired on 4 different OCT devices: Casia SS-1000, Anterion, Visante, and a ultrahigh-resolution custom build. CUNEX was benchmarked against 2 established models, CorneaNet and ScLNet. We then evaluated whether CUNEX-based segmentation influenced downstream deep learning classification of age, sex, and disease staging. Results: On the internal MS-39 validation set, CUNEX achieved Dice similarity coefficient (DSC) and intersection over union (IoU) scores of 94% to 95% and 88% to 91% across healthy, keratoconus, and Fuchs endothelial corneal dystrophy eyes, comparable to ScLNet (within 3%) and superior to CorneaNet (8%–35% lower). On external datasets, CUNEX was the only model to produce anatomically meaningful segmentations across devices (DSC and IoU >70%), whereas ScLNet and CorneaNet showed substantial performance degradation (DSC and IoU <20%). Segmentation minimally affected classification accuracy except for sex prediction, where accuracy dropped from 81% to 68%, suggesting sex-related features may lie outside the cornea. Conclusions: CUNEX delivers the first open-source AS-OCT corneal segmentation model evaluated across multiple independent OCT platforms. It offers a reproducible foundation for integrating segmentation into clinical and artificial intelligence research workflows. The code and pretrained weights are available at https://github.com/lkandakji/CUNEX.

AI에게 질문

Bookmark

View Full Paper