Manual and Automatic Segmentations of the Fibrillar Layer in Fuchs Endothelial Corneal Dystrophy Eyes

Purpose: Subendothelial collagen deposits, referred to as fibrillar layer (FL), are present in about 80% of eyes with advanced Fuchs endothelial corneal dystrophy (FECD). We investigated reliability and reproducibility of FL segmentations on Scheimpflug corneal densitometry images and trained a deep neural network for automatic segmentation.
Methods: In a retrospective monocenter study, high-quality preoperative Scheimpflug images of patients undergoing DMEK or triple DMEK for advanced FECD were included. FL-negative cases were excluded. Images were manually segmented by two graders with the open-source software MITK. A MATLAB-based code was generated for quantitative and spatial analysis calculated the region of interest (ROI), maximum caliper diameter, horizontal and vertical diameter, and Dice similarity coefficient for FL-areas. The FL localization was visualized by spatial heatmaps. The intraclass correlation coefficient (ICC) was calculated by IBM SPSS. The segmentations of Grader 1 and 2 were combined and used as ground truth for training a neural network (U-Net) for FL segmentation. 25 images were reserved as test set. Of the remaining data, 176 (80%) images were used for training, 44 (20 %) for validation. Accuracy metrics (Dice and Jaccard Coefficient) were calculated on the test set to quantify the degree of similarity between the manual segmentations (ground truth) and the model.
Results: Among 308 eyes, 248 were FL-positive (80,5%) and 60 FL-negative (19,5%). Intra-observer Grader 1 vs intra-observer Grader 2 vs inter-observer results include: ROI 7.91mm2 ±4.8 vs 9.29mm2±5.0, ICC (ROI) at 0.957 vs 0.930 vs 0.954, and Dice similarity coefficient 0.86 vs 0.89 vs 0.85. Spatial heatmaps demonstrated inferotemporal localization of increased densitometry areas in FL-positive eyes. The model was trained over 300 epochs. High accuracy in Dice (0.92) and Jaccard Coefficent (0.85) was achieved after 150 epochs on the test set.
Conclusions: The FL can be reliably and reproducibly visualized on corneal densitometry images in the clinical routine and projects to the inferotemporal corneal quadrant. A neural network can be trained for automatic segmentation that achieves comparable or even better results than manual segmentation. In the future, the FL could play a role in the individualization of descemetorhexis in DMEK and in the allocation of rare donor tissue.