Automated quantification of lens epithelial cell density using deep learning: validation and large-scale clinical...

Lens epithelial cells (LECs) have a critical role in nutrient transport, ion balance, and the synthesis of essential molecules required to preserve the lens's transparency. We hypothesize that lens epithelial cell density (LECD) may be correlated with the formation and severity of cataracts. Studying this relationship is limited by current quantification methods. This study aimed to develop an AI-driven model capable of automatically performing epithelial cell (LEC) counts in excised capsules, including density, distribution, and determining factors that affect LECD. We developed an AI-based software that leverages deep learning algorithms to automate the enumeration of LECs from light micrographs of harvested anterior lens capsules. To evaluate the performance and reliability of our AI model, we compared its results against traditional manual cell counting methods. Validation analyses included repeated-measures ANOVA, Bland-Altman analysis, mean absolute percentage error (MAPE), the intraclass correlation coefficient (ICC), and a comparison against inter-observer agreement between two independent expert observers to quantitatively assess agreement between AI-generated cell counts and manual enumeration. Participants were patients with age-related cataracts scheduled for phacoemulsification. Over 43,000 individual cellular targets were analyzed across 20 validation images. The AI-driven software showed excellent agreement with consensus manual counts (98.1% accuracy; MAPE 1.87%, 95% CI 1.27-2.47%; Pearson r = 0.99; ICC 0.994, 95% CI 0.982-0.997), tighter than the agreement between the two expert observers themselves (MAPE 3.74%; ICC 0.972). The AI tool provides a rapid, objective, and repeatable method for LECD analysis.