New deep AI‐based algorithm to assess quality of primary corneal endothelial cell cultures

Aims/Purpose: The development of cell therapies to treat corneal endothelial pathologies requires characterizing the cells during the cell culture process. The endothelial cell density (ECD) and morphometry are essential parameters for identifying cultures with sufficient phenotype and yield. However, they are difficult to measure on cells in culture using conventional image analysis tools. Aims: (1) To develop an AI‐based automatic cell segmentation method for reliable ECD and morphology measurement. (2) To select new morphometry parameters. (3) To classify endothelial cell cultures by quality.Methods: Experiments were conducted on primary cultures of human corneal endothelial cells (hCECs) after immunofluorescent labeling against NCAM revealing lateral membranes. Cultures from donors < 30 years old (high ECD and regular mosaic) were compared with cultures form older donors comprising senescent cells and cells in endothelial‐mesenchymal transition. Images were acquired with an epifluorescence microscope. We compared two AI segmentation models, a custom model and the published Cellpose model1,2. Models were trained with manually labeled images and tested on separate images. Standard image analysis methods in Python measured cell number, morphology, comparing reference and AI‐generated images3. Additional metrics were developed to qualify endothelial cell culture and correlated with known metrics. K‐Means and other classifications were tested to classify cell culture quality.Results: Cellpose outperformed the custom model in segmentation accuracy (mean relative error ± SD respectively: 0.054 ± 0.020 vs 0.0499 ± 0.0136). New reliable morphological measurements were defined, enabling classification of endothelial cell culture clusters based on quality. Coronal CV (coefficient of variation) was developed to complement classical CV used to characterize clinical endothelial cell.Conclusions: AI segmentation of hCEC in culture significantly improves quality measurement during bioengineering, facilitating rapid analysis and classification of cell batches, thereby enhancing quality assessment and success rates for cell‐based endothelial therapy.References Stringer et al (2020). Pachitariu et al (2022). Gavet et al (2012).