Abstract
Corneal endothelial cell morphology is clinically important to understand corneal health, and is routinely examined using specular microscopy. Traditional metrics such as endothelial cell density (ECD), coefficient of variation (CV), and hexagonality percentage provide limited morphological information and are unreliable when used on large diverse datasets. This paper presents an automated phenotyping framework using Elliptic Fourier Descriptors (EFDs), Zernike moments, and geometric features combined with unsupervised K-Means clustering for endothelial cell classification. The framework was validated on two independent datasets. The primary dataset comprised 179 cells (1,253 post augmentation) from four specular microscopy images. The external validation set (SREP-18-33533B) comprised 35,890 cells across 385 images. EFD based features achieved a mean silhouette score of 0.6213 (SD = 0.056) on the primary dataset, representing a large effect size over traditional measures (Cohen’s d = 5.290). Across datasets, traditional measures declined by nearly 40% (0.5986 to 0.3628), whereas EFD based scores remained stable with under 1% change (0.6213 to 0.6119), as confirmed by the Wilcoxon signed-rank test (p < 0.001). Three reproducible phenotypes were identified. Cluster 0, contained regular compact cells (64.4%). Cluster 1, contained moderately regular cells (29.2%) and Cluster 2, contained irregular or elongated cells (6.4%). These results indicate that EFD based shape analysis remains discriminative on a 35,890 cells where traditional clinical metrics fail. The framework can support real world analysis of corneal endothelial cell phenotypes when combined with clinical expertise of researchers and ophthalmologists.