Toward automated assessment of conjunctival hyperemia: A semisupervised artificial intelligence approach

This paper develops an automated approach for conjunctival hyperemia grading from slit-lamp images using semisupervised learning. We conducted a retrospective study including slit-lamp images from two study sites. Two independent graders assessed the severity of hyperemia according to the Efron Grading Scales. Segmentation of the conjunctiva and its vessels was performed using semisupervised segmentation with limited labeled data. Conjunctival vessel densities were estimated from the model outputs and compared against the manual clinical Efron gradings. Three hundred and seventeen slit-lamp images from the primary site and 164 from an external site were included. The semisupervised models with unlabeled data demonstrated significantly improved segmentation compared to a baseline fully supervised model using only the labeled data (p < 0.001). Calculated conjunctival vessel densities showed correlations of 0.86 [0.76, 0.93] with ground truth vessel densities. Comparisons of vessel densities against mean manual clinical Efron gradings showed correlations of 0.83 and 0.80 for the test and external datasets, which were comparable to the inter-rater agreements of 0.82 [0.68, 0.90] and 0.75 [0.67, 0.81] in the datasets, respectively. Conjunctival vessel densities obtained with semisupervised learning showed good agreement with clinical grading of conjunctival hyperemia. This approach may be applied toward an automatic, objective assessment of the conjunctiva.