Topic: ESOPRS 2021 ePoster sessions
Time: Sep 17, 2021 16:00 Amsterdam, Berlin, Rome, Stockholm, Vienna, 15:00 London
(plain text version here)
Automated eyelid measurements from patient videos using artificial intelligence
Author: Christopher Schulz
ePoster Number: 103
Purpose
Advances in artificial intelligence could pave the way for automated video assessment of the ocular adnexa. The current study aimed to develop and test a custom program to automate the assessment of eyelid position (MRD1 & MRD2), blink lagophthalmos (bLag) and average ocular surface area exposure (OSAE) from consumer-grade videos.
Methods
A custom program was designed to read and analyse each frame from a video of a person’s face. A neural network was trained to label periocular anatomy using 7101 pre-labelled images with 98.2% accuracy. Using this program, clinical markers (MRD1, MRD2, bLag & OSAE) were determined and compared for 77 videos of individuals with acute onset facial nerve palsy (FNP), 33 with ptosis, 33 with thyroid eye disease (TED) and 65 controls. For 31 individuals with FNP, videos were assessed prior to onset of FNP, at onset and at 6 months. A further 28 eyes underwent both clinical assessment and automated video assessment, evaluating agreement by Bland-Altmann analysis.
Results
Mean automated MRD1 was 3.6mm in controls, 2.2mm in patients with ptosis (p<0.001) and 3.9mm in FNP (p=0.049). Mean MRD2 was 5.9mm in controls and 6.4mm in TED (p<0.001). Mean bLag was 0.0mm in controls, 3.7mm in FNP (p<0.001) and 1.6mm in TED (p=0.009). Mean OSAE was calculated as 13.1mm2 in controls, 12.2mm2 in ptosis (p<0.001), 13.3mm2 in FNP (p=0.50) and 13.7mm2 in TED (p=0.002). At an individual level, OSAE and bLag increased after the onset of FNP (p=0.035 and p<0.001, respectively); no significant change was observed in MRD1 or MRD2. At 6-month follow-up there was a significant decrease in MRD2 (p<0.001), OSAE (p=0.002), and bLag (p=0.004). Automated and clinician-derived measures of MRD1 and MRD2 demonstrated 95% limits of agreement of 1.1mm and 0.75mm respectively.
Conclusion
Automated video assessment of the adnexa is feasible, demonstrating promising validity and reliability. Such technology might facilitate remote clinical assessment and self-monitoring. It might also facilitate the measurement of novel, sensitive and clinically meaningful disease markers e.g. blink lagophthalmos and ocular surface exposure.
Additional Authors
There are no additional authors to display.
Abstract ID: 21-167