Corneal Endothelial Images Segmentation Using Deep Learning Model

The objective assessment of corneal endothelial pathology is frequently carried out using manual and semi-automated software tools. These software applications enable the quantification and analysis of corneal endothelial cells. However, this approach is subjective and not widely adopted by ophthalmologists. Developing methods for the automatic assessment of the size and structural organization of endothelial cells represents a significant challenge in the field of ophthalmology. Therefore, this article introduced a cell segmentation method using a convolutional neural network based on the U-Net architecture. More specifically, the network is trained to discriminate pixels at the borders between cells. The main contribution of this study lies in modifying the original U-Net by reducing the number of parameters in width and using average pooling, aiming to enhance the efficiency of corneal endothelial cell segmentation and analysis. The proposed method was evaluated on a dataset consisting of corneal endothelial cell images. The accuracy values are around 93.99% and the sensitivity has a high value (99.33
^%
) compared to other architectures.