Comparison of Autonomous AS-OCT Deep Learning Algorithm and Clinical Dry Eye Tests in Diagnosis of Dry Eye Disease.

OBJECTIVE

To evaluate a deep learning-based method to autonomously detect dry eye disease (DED) in anterior segment optical coherence tomography (AS-OCT) images compared to common clinical dry eye tests.

METHODS

In this study, 27,180 AS-OCT images were prospectively collected from 151 eyes of 91 patients. Images were used to train and test the deep learning model. Masked cornea specialist ophthalmologist diagnoses were used as the gold standard. Clinical dry eye tests were performed on patients in the DED group to compare the results of the model. The dry eye tests performed were tear break-up time (TBUT), Schirmer's test, corneal staining, conjunctival staining, and Ocular Surface Disease Index (OSDI).

RESULTS

Our deep learning model achieved an accuracy of 84.62%, sensitivity of 86.36%, and specificity of 82.35% in the diagnosis of DED. The positive likelihood ratio was 4.89, and the negative likelihood ratio was 0.17. The mean DED probability score was 0.81 ± 0.23 in the DED group and 0.20 ± 0.27 in the healthy group (P < 0.01). The deep learning model accuracy in the diagnosis of DED was significantly better than that of corneal staining, conjunctival staining, and Schirmer's test (P < 0.05). There was no significant difference between the deep learning diagnostic accuracy and that of the OSDI and TBUT.

CONCLUSION

Based on preliminary results, reliable autonomous diagnosis of DED with our deep learning model was achieved, when compared with standard dry eye clinical tests that correlated significantly more or similarly to diagnoses made by cornea specialist ophthalmologists.