Corneal guttae, abnormal growths of extracellular matrix in the corneal endothelium, appear in specular microscopy images as black droplets that obscure the endothelial cells. These guttae are a hallmark of Fuchs’ endothelial dystrophy and play a crucial role in disease diagnosis and progression assessment. However, guttae are currently assessed using slit-lamp examination, a subjective method that does not allow for accurate quantification or consistent grading. This limitation underscores the need for standardized, objective methods to assess guttae.
The primary objective of this thesis is to develop a U-Net capable of detecting guttae in corneal images. Two datasets from earlier studies regarding Baerveldt implants and corneal transplants were used in this study, comprising a total of 82 images included in the analysis. The study first investigates inter- and intra-grader variability in manual guttata segmentation, revealing significant discrepancies in human annotations.
To address these challenges, several model configurations were explored, including incorporating postprocessing techniques, regions of interest masks to focus on key areas, data augmentation and patch-based training. All models were evaluated against manual annotations and the baseline model with postprocessing (BP) achieved a Dice score of 0.82 ± 0.11 on the test set, demonstrating performance comparable to that of human graders.
While the BP model highlights the potential of AI in achieving consistency and accuracy in
guttata segmentation, challenges related to the uncertainty in manual labels remain. Future work should focus on improving label quality through iterative validation and uncertainty modeling. Despite these challenges, this study establishes a promising foundation for AI-driven tools in the assessment of Fuchs’ endothelial dystrophy, paving the way for more objective clinical and research applications.