Screening for referrable dry eye disease using deep learning: a multicenter Asian study

Abstract

Purpose : Managing dry eye disease in the community requires a reliable way to detect cases that need specialist care. We train and validate an algorithm for screening of referrable dry eye disease based on deep learning from corneal staining images.

Methods : Images acquired with slit lamp microscopy on dry eye patients with significant corneal staining and those without such staining were evaluated by an ophthalmologist for training the deep learning algorithm. We used EfficientNet, a deep neural network architecture known for its high accuracy and low computational cost, based on scaling up a neural network systematically. The model was trained with data augmentation techniques, including random resize, random rotation, random flip and colour jitter. The model was trained for 50 epochs and the best model was taken based on the best validation AUC over the training.
To externally validate the algorithm, we used images from 775 eyes of 775 patients (classified into referrable or non-referrable) from Singapore and China.
We generated saliency maps using Gradient-weighted Class Activation Mapping to identify the regions of an input image used by the neural network to make a prediction.

Results : We used images from 261 eyes from dry eye patients in Singapore to train the initial algorithm. The validation was performed with four datasets (62 images from Singapore, 338 images from Wenzhou, 217 and 220 images from Beijing). Validation on the Singapore dataset showed area under the curve [AUC] of 0.951 (95% CI 0.891 - 0.989), Accuracy of 0.855 (0.758 - 0.935), Sensitivity of 0.763 (0.619 - 0.886) and specficity of 1.000 (1.000 - 1.000).
In the Wenzhou dataset, AUC was 0.895 (0.859 - 0.926), Accuracy was 0.793 (0.749 - 0.831), Sensitivity was 0.728 (0.671 - 0.784) and specificity was 0.949 (0.899 - 0.989). In one of the Beijing dataset, AUC was 0.929 (0.891 - 0.964), accuracy was 0.783 (0.728 - 0.839), sensitivity was 0.750 (0.685 - 0.815) and specificity 0.946 (0.861 - 1.000).
The saliency heat maps showed that the hot areas highlighted the corresponding areas stained by fluorescein dye (green).

Conclusions : A single well-focused cornea image taken after dye staining has the potential to be used to differentiate cases that required specialist care from those that don't in Asian patients with dry eye.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.