Eyelid position abnormalities can be detected from photographs using deep-learning models, which may aid in the early detection of ocular diseases, according to a poster presented at ASOPRS 52nd Annual Fall Scientific Symposium.
Using photos from a database, manual and automatic scores for ptosis and eyelid
The deep-learning network (DLN) DeepLab v3+ performed best during manual annotation of ocular landmarks on photographs from the Chicago Face Database (F1 score = 0.93).
DLN was able to identify 71% (n = 77 images) of cases of ptosis and 80% (n = 57 images) of cases of retraction.
There was agreement between trained ophthalmologists and the Eyemeter program in margin reflex distance 1 values (Intra-Class Coefficient [ICC] of 0.610. Ophthalmologists and Eyemeter evaluated 10 clinical photos of patients with ptosis and 24 photos of patients with thyroid eye disease (TED). Good correlation was seen for detection of retraction in TED (ICC = 0.79), but there was a higher rate of failure in segmenting the corneoscleral region of TED patients compared to healthy patients for the DLN.
The authors concluded that “Additional optimization of our program is necessary, along with analyzing a large photo database to increase the accuracy of disease detection.”
Reference
Grob, et al. Use of Deep-Learning to Automatically Detect Eyelid and Orbital Disease from Photographs. Presented at: ASOPRS 52nd Annual Fall Scientific Symposium.
Related Content
