Abstract
Purpose :
The Tear Break-up Time (TBUT) is a commonly used test to measure the stability of the tear film in people with Dry Eye Disease (DED). A short break-up time indicates poor tear film quality and can indicate DED. However, the TBUT test is subjunctive and depends on the skill and experience of the ophthalmologist performing the test. Our project aims to use AI to automate this test, reduce subjectivity and achieve more consistent and reliable results.
Methods :
A digital slit-lamp was used to record the TBUT measurement for each of the 54 study participants within a specific protocol. We split each video into frames, which an expert annotated into three classes: blinking, unbroken tear film, and broken tear film. This database was divided into 40 videos for training, 7 for validation, and 7 for testing. We propose 2 approaches: The baseline model, a standard classification network between unbroken and broken tear film classes using 124 500 frames, EfficientNet-B0-NS, and a BinaryCrossEntropy loss. For the second approach, we developed an original method called"Dual-task siamese network". For this approach, 124 500 image pairs were generated. We used the same backbone as the baseline, and a weighted sum of three losses, two BinaryCrossEntropy losses used for the classification task, and a contrastive loss for the similarity learning task.
Results :
For the classification task between unbroken and broken tear films, using the baseline model, we obtained an AUC of 0.92 while with the"Dual-task siamese network" model, the AUC reaches 0.98. To go further, we compute the TBUT using the results obtained with the dual-task siamese network. After applying Gaussian smoothing to the model predictions and taking 0.5 as a decision threshold, as shown in table 1, we got 5.36 ± 3.69 seconds as a predicted TBUT. These results do not significantly differ from the ground-truth (5.35 ± 3.79 seconds) as the Wilcoxon p-value is greater than 0.05.
Conclusions :
The results show the good performance of the proposed AI approach for the TBUT test quantification. However, we see the potential for even further improvement by incorporating more data and another annotation by an ophthalmologist into our method. This will allow us to continue to refine and enhance the TBUT quantification proposed approach.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.