Enhancing Corneal Endothelium Segmentation: A Comparison of Classification and Distance-Map Regression UNets

This paper compares two UNet-based architectures for corneal endothelium segmentation: a classification approach (UNet-mask) and a distance-map regression approach (UNet-dm). Our results show that the UNet-dm outperforms the UNet-mask with an average Dice coefficient of 0.8180 compared to 0.6583. Moreover, the UNet-dm model generates well-defined cell boundaries and produces mor-phometric parameters closer to the reference values. This study highlights the potential of distance-map regression-based UNet models for accurate corneal endothelium segmentation.