Oral Abstract Presentation

OP-3 Deep-learning identification of stromal hyperreflectivity on AS-OCT and its influence on visual outcomes after DMEK surgery

Abstract

Purpose To assess the influence of preoperative stromal hyperreflectivity on visual outcomes of Descemet Membrane Endothelial Keratoplasty (DMEK) surgery.

Design Monocentric, cohort study.

Methods Anterior Segment Optical Coherence Tomography (AS-OCT) imaging of eyes which underwent uncomplicated DMEK surgery at the Royal Liverpool University Hospital were collected before and after surgery. Patient electronic records were reviewed to collect visual acuity outcomes. A deep-learning algorithm was developed to segment the corneal boundaries and identify clusters of hyperreflectivity. The loss function utilised in this study was a combination of dice loss and cross-entropy loss and an Adam-based optimizer was employed for optimisation.

Results A total of 19 eyes from 18 patients were analysed. Visual acuity improved in all eyes after DMEK (mean [SD], 0.59 [0.31] vs. 0.26 [0.22] LogMAR, p<0.001). Stromal hyperreflectivity correlated with preoperative central corneal thickness (p=0.88, p<0.001) but not with preoperative visual acuity (p=0.11, p=0.65). At 6 months after DMEK surgery, patients with preoperative stromal hyperreflectivity higher than the median values had lower final visual acuity than those with lower values of stromal hyperreflectivity (mean [SD], 0.29 [0.3] vs. 0.23 [0.12] LogMAR, p<0.04).

Conclusions Clusters of stromal hyperreflectivity can be identified and monitored with deep-learning based segmentation algorithms. Preoperative stromal hyperreflectivity was associated with lower visual acuity recovery after DMEK surgery. Tools to identify stromal hyperreflectivity corresponding to clinical stromal scarring can help clinicians in stratifying candidate patients for DMEK and gauging the expected visual acuity recovery rate.

Article metrics
Altmetric data not available for this article.
Dimensionsopen-url