Deep-learning models can predict common eyelid position abnormalities
Deep-learning models can detect common eyelid position abnormalities from photographs, potentially helping aid diagnosis in the telemedicine setting, according to a study.
Researchers performed manual annotation of ocular landmarks on photos to train semantic segmentation of deep-learning networks. A total of 597 photos were annotated and diagnosed by 4 graders, and 557 and 40 of these images were used to train the deep-learning networks and validate the network, respectively.
The DeepLab v3+ was the best performing deep-learning network trained in this study. It had an F1 score of 0.93 and a receiver operating characteristic curve of 0.9999.
The network correctly diagnosed 71% of ptosis cases and 80% of eyelid retraction cases.
Grob SR, et al. Automatic identification of eyelid position abnormalities using computer vision. Presented at: 2020 ASOPRS Virtual Meeting.