Automated corneal endothelial cell detection and segmentation in specular microscopy images using Attention and Faster U-Net

Monitoring the density of corneal endothelial cells (CECs) is important for assessing corneal health. Manual calculation is time consuming and prone to error, which drives the need for reliable automated methods. U-Net, a neural network for biomedical image analysis, has shown promising results in the automated detection of cells in healthy and good-quality corneas. The purpose of this study was to assess the performance of two deep learning models namely, Attention and Faster U-Net, on real-world CEC images of variable quality and differing ocular disease from the publicly available SREP-18-33533B dataset. The outcome measures were precision, recall of CEC detection, correctness of CEC density estimation and detection of ungradable images. A classical marker-controlled watershed approach was used for comparison. There was good agreement between the automated image analysis and manual annotation from Attention and Faster U-Net. R² derived from the regression was 0.95 for the Attention U-Net and 0.94 for the Faster U-Net. Recall of CEC was averaged to 0.65 for both models and precision averaged to 0.66 and 0.64 respectively. The marker-controlled watershed underperformed with an R² derived from regression was only 0.12. This framework provides a reproducible foundation for future clinical research in ophthalmology.