Fig 1.
Corneal segmentation and scoring method.
(A) Corneal segmentation grid and proportion (right eye). The horizontal and vertical ratios of each zone of the grid are 1:1.6:1. (B) Two examples of NEI scale evaluation. PEE of the five zones is assessed and scored using the NEI scale. (C) Corneal segmentation grid and proportion (left eye). NEI, National Eye Institute; PEE, punctate epithelial erosion.
Fig 2.
Diagram of the proposed deep learning system.
In step 1, corneal region in fluorescein-stained slit lamp image is segmented using U-Net architecture with 1100 images and their corneal region labeled masks. In step 2, CNN-based classification model was trained with 200 images and their PEE and non-PEE labeled data to find the PEE candidate regions within the corneal region. In step 3, PEE quantification is performed using PEE density map and presented as MDV. PEE, punctate epithelial erosion; MDV, maximum density value.
Fig 3.
Schematic diagram of dataset splitting for deep learning analysis.
For the PEE candidate regional model, 200 cases were divided into the training, development, and validation sets in a 7:2:1 ratio. For the corneal region segmentation model, PEE detection, and quantification model, 1100 cases were divided 5-folds. Also, data from 94 cases were used for external validation of the entire system, and data from another 100 cases were used for serial data analysis. PEE, punctate epithelial erosion.
Fig 4.
Examples of PEE and non-PEE labeling generated using Microsoft paint software.
(A–C). Labeling of definite PEE (red color) and definite non-PEE (yellow color). (D). Extraction of patches sized 192 × 192 pixels for learning. PEE, punctate epithelial erosion.
Fig 5.
Ground truth NEI score of the development dataset (hospital 1 data) at each zone (A) and total NEI score (B), and ground truth NEI score of the external validation dataset (hospital 2 data) at each zone (C) and total NEI score (D). NEI, National Eye Institute.
Table 1.
Diagnosis of the patients in the development (hospital 1) and external validation (hospital 2) datasets.
Fig 6.
(A) Correlation between median NEI score and MDV, and (B) correlation between mean NEI score and MDV by zone. NEI, National Eye Institute; MDV, maximum density value.
Table 2.
Agreement in corneal fluorescein staining (CFS) scores based on the National Eye Institute (NEI) scale.
Fig 7.
Examples of the training dataset, input, and final output used in the corneal segmentation step.
The input is a fluorescein-stained corneal image (first panel). The ground truth (second panel) is used to train the cornea segmentation model. The red lines in the second panel indicate corneal regions. The third panel displays the predicted image of corneal segmentation. The final output (last panel) is a grid mask on a predictive image obtained using computer vision algorithms.
Fig 8.
Illustration of classification model and density map results.
(A) CAM of PEE candidate region classification. The red and yellow boxes represent true positive and true negative of the PEE classification model, respectively. (B) Density map results. Blue indicates low PEE density, and red indicates high PEE density. CAM, class activation map; PEE, punctate epithelial erosion.
Fig 9.
Correlation between total NEI score and MDV.
(A) Spearman correlation result between total NEI score and MDV of the development dataset (hospital 1). The Spearman correlation is 0.868 (p<0.001). (B) Spearman correlation results between the total NEI score and MDV of external validation dataset (hospital 2). The Spearman correlation coefficient is 0.863, and the p-value is <0.001. NEI, National Eye Institute; MDV, maximum density value.
Table 3.
Agreement between the entire model and ground truth data for the assessment of improvement or deterioration in 50 eyes (n = 100 images).