ORIGINAL RESEARCH article
Front. Med.
Sec. Ophthalmology
This article is part of the Research TopicEfficient Artificial Intelligence in Ophthalmic Imaging – Volume IIIView all articles
Deep Learning-Assisted Widefield Endothelial Imaging in Descemet Membrane Endothelial Keratoplasty
Provisionally accepted
Ezekiel Ze Ken Cheong1,2*
Khin Yadanar Win3,4
Qiu Ying Wong3,5
Ralene Sim3,5Clarissa Ng Yin Ling3,5Hla Myint Htoon2,3
Hon Shing Ong2,3,5
Jodhbir Mehta2,3,5
Bingyao Tan2,3,5
Leopold Schmetterer2,3,4,5Marcus Ang2,3,5*- 1Singapore National Eye Center, Singapore, Singapore
- 2Ophthalmology and Visual Sciences Academic Clinical Programme, Duke-NUS Medical School, Singapore, Singapore
- 3Singapore Eye Research Institute, Singapore, Singapore
- 4SERI-NTU Advanced Ocular Engineering (STANCE) Program, Singapore, Singapore
- 5Singapore National Eye Centre, Singapore, Singapore
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
Purpose: Significant endothelial cell density (ECD) losses in Descemet membrane endothelial keratoplasty (DMEK) can precipitate graft failure. Currently clinical practice only evaluates the central corneal endothelium and not the paracentral or peripheral cornea. Here, we evaluate a deep learning (DL) algorithm for automated image quality assessment and detection of low corneal ECD in widefield specular microscopy (WFSM) images following DMEK, at central, paracentral and peripheral regions. Methods: Prospective observational study in eyes with clear, stable DMEKs. WFSM imaging (CEM-530; Nidek, Japan) performed at a baseline visit (mean of 23 months post-DMEK) and 6 months after. Images at 15 locations were captured; 1 central, 8 paracentral, and 6 peripheral. Image quality and corneal endothelial cell parameters including ECD were assessed. SqueezeNet architecture was employed for binary and multi-class automated image quality and low ECD detection (threshold set < 1000 cells/mm2). Main outcome measures were regional ECD and area under curve (AUC) for DL tasks. Results: DMEK was performed in 53 eyes (43 subjects, mean age 67.4 ± 10.1 years) with Fuchs endothelial corneal dystrophy (FECD) (71.7%; 38 eyes) and pseudophakic bullous keratopathy (PBK) (28.3%; 15 eyes). The DL classification models were trained and validated on 1362 images, achieving AUCs of 0.979 (binary image quality), 0.907 (multi-class image quality) and 0.980 (low ECD detection). WFSM imaging was able to detect both central ECD loss (1709 vs 1555 cells/mm2; P < 0.001) and peripheral ECD loss (1498 vs 1347 cells/mm2; P = 0.016). Overall, we estimated a mean annual central ECD loss of 5.81% (95% CIs: 3.54% - 8.08%) in the stable phase of our DMEK cohort. Conclusions: Serial WFSM was able to image a larger area of the endothelium in eyes with DMEK. Our DL model was a useful adjunct that analysed the large number of specular microscopy images to identify scans of adequate quality and ECD based on pre-defined thresholds. This approach may support the use of DL-assisted WFSM for early detection of ECD loss which may aid in monitoring endothelial health and graft survival following DMEK.
Keywords: Descemet membrane endothelial keratoplasty, Corneal endothelium, widefieldspecular microscopy, deep learning, image quality, Endothelial cell density, Endothelial cell loss
Received: 29 Sep 2025; Accepted: 29 Oct 2025.
Copyright: © 2025 Cheong, Win, Wong, Sim, Ng Yin Ling, Htoon, Ong, Mehta, Tan, Schmetterer and Ang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
* Correspondence:
Ezekiel Ze Ken Cheong, ezekiel@u.duke.nus.edu
Marcus Ang, marcus.ang@snec.com.sg
Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.