Validation of a Deep Learning-Assisted Evaluation of Total Corneal Endothelial Cells Viability

Abstract

Purpose: To describe the validation of a novel automated analysis of preoperative pan-corneal endothelial cell viability. Methods: Preclinical experimental study. Dead endothelial cells and denuded areas of Descemet membrane of corneoscleral rims were stained with trypan blue (TB) 0.05%. Endothelial mortality was estimated by an experienced eye bank technician ("gold standard") and by deep learning-aided automated segmentation of TB-positive areas (TBPAs) on images of the stained corneas ("V-CHECK method"). V-CHECK mortality was calculated for the whole cornea and for concentric 2-mm rings. The agreement in the estimation of endothelial mortality between the two methods was assessed with Bland-Altman analysis and correlation tests. Results: Nineteen corneas deemed unsuitable for transplantation were used for the experiment. The automated V-CHECK method was able to accurately segment the corneal endothelium and the TBPAs. The gold standard and the V-CHECK method showed a strong positive correlation for all rings (Pearson's ρ, range 0.76-0.81, all P < 0.001). The V-CHECK method resulted in a higher average estimated endothelial mortality (mean difference range +6.5% to +9.5%). Conclusions: The V-CHECK method enables reproducible estimation of endothelial cell viability in donor corneas. Incorporating this technique into the preoperative assessment of donor corneal tissues (in the eye bank and in the operating theater) can provide a reliable evaluation of endothelial health, thereby improving the consistency of tissue quality and further supporting optimal surgical results. Translational Relevance: The V-CHECK deep learning-assisted computer vision protocol will allow surgeons and eye bank technicians to perform an independent, preoperative assessment of global corneal endothelial viability.