Related Experiment Video
Updated: Jul 30, 2025
10:46
Corneal Donor Tissue Preparation for Descemet's Membrane Endothelial Keratoplasty
Published on: September 17, 2014
15.3K
Deep Learning Using Preoperative AS-OCT Predicts Graft Detachment in DMEK
Alastair Patefield1, Yanda Meng1, Matteo Airaldi2
1Department of Eye and Vision Sciences, Institute of Life Course and Medical Sciences, University of Liverpool, Liverpool, UK.
Translational Vision Science & Technology
|May 15, 2023
Summary
A novel deep learning algorithm shows high sensitivity in detecting potential graft detachment after Descemet membrane endothelial keratoplasty (DMEK). This artificial intelligence tool could aid in routine screening for DMEK suitability.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Descemet membrane endothelial keratoplasty (DMEK) is a surgical procedure to treat corneal endothelial dysfunction.
- Graft detachment is a potential complication following DMEK, necessitating timely detection and intervention.
- Anterior segment optical coherence tomography (AS-OCT) is a key imaging modality for assessing DMEK outcomes.
Purpose of the Study:
- To evaluate a novel deep learning algorithm for distinguishing eyes with or without graft detachment.
- To assess the algorithm's performance using pre-operative AS-OCT images in the context of DMEK.
Main Methods:
- A retrospective cohort study utilizing a multiple-instance learning artificial intelligence (MIL-AI) model (ResNet-101 backbone).
- AS-OCT images from 74 eyes were split into training (50 eyes) and testing (24 eyes) sets.
- Model performance was evaluated using F1 score, precision, specificity, sensitivity, and AUC, and compared to an ophthalmologist's classification.
Main Results:
- The MIL-AI model achieved an F1 score of 0.77, precision of 0.67, specificity of 0.45, sensitivity of 0.92, and AUC of 0.63 on the test set.
- The algorithm demonstrated higher sensitivity (92%) compared to manual classification (31%) for detecting potential graft detachment.
- Heatmaps were generated to visualize areas contributing to the AI's classification.
Conclusions:
- The MIL-AI model exhibits high sensitivity in predicting eyes that may experience post-DMEK graft detachment requiring rebubbling.
- Further large-scale clinical trials are recommended to validate the model's efficacy.
- MIL-AI models show promise for integration into routine screening processes for DMEK suitability.
