10 Feasibility assessment of automated corneal endothelium quantification using CELDA in donor corneas

Résumé fourni par la source

Purpose Endothelial cell density (ECD) is essential for determining the quality of donor corneal tissue. However, current assessment methods rely on manual counting and are subject to inter-operator variability. CEL-DA is a GUI (Graphical User Interface) that deploys a deep-learning model trained on specular micros-copy images from patients’ corneal endothelium. This feasibility study evaluates CELDA’s performance for automated ECD determination in ex vivo donor corneas at ETB-BISLIFE Eye Bank in Haarlem, The Netherlands.Materials and methods CELDA was tested on 34 heterogeneously focused images, acquired from six donor corneas, to deter-mine which image characteristics it handles best. Scale and blur preprocessing approaches were tested to enhance robustness. Twelve corneas were prospectively imaged at one central and four peri-central locations targeting the identified image characteristics. CELDA’s outputs were compared with manual ECD counts using means, Lin’s concordance correlation and Bland-Altman analyses.Results CELDA performed well on images with low-intensity backgrounds, high-intensity cell bodies, and redu-ced detail. Preprocessing with a 0.3 scaling factor improved robustness, with cells recognized in 93% of prospectively acquired images; cells were not recognized in four images from a single cornea due to insufficient quality. CELDA based its ECD calculation on more cells per cornea than manual counting (2475 [744 - 4390] vs 132 [118 - 140]) with an increasing difference between the methods at higher ECDs (mean difference 182 cells/mm 2; 95% CI -167.3 to 533.1), with decreased agreement toward the periphery (rc = 0.77 for central vs 0.64 for full cornea). Limited segmented area (<50% vs >50%) was associated with increased difference between both methods (mean 255 vs 87 cells/mm2; p = 0.008). CELDA showed promising intra-corneal agreement (rc = 0.64; 0.94 after outlier removal).Conclusions CELDA shows great potential for automated ECD quantification in donor corneas. However, standardisa-tion of image acquisition, model retraining, and extensive validation are required before implementation of this fully automatic approach at Eye Banks using light microscopy. CELDA’s intra-corneal agreement was high, and the method shows potential to improve reproducibility and efficiency in ECD determina-tion based on a single image of the central donor cornea.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.