Abstract
Purpose :
Acquiring high quality ophthalmic images typically requires a trained operator in close proximity to a patient. Motorized alignment can enable remote operation, e.g. for physical distancing during a pandemic, and automated alignment can enable quality imaging when a trained operator is not readily available, e.g. in low-resource environments. We evaluated and compared manual and automated operation of an ophthalmic device on a motorized stage.
Methods :
We acquired data on 5 normal subjects (10 eyes) by both manual (via a trained operator) and automated operation using a prototype with three off-axis iris cameras (one below and on each side of the device) and a 3-axis motorized stage (Fig. 1). The motorized stage allowed all data to be acquired with a plexiglass shield between the subject and operator for remote operation during the COVID-19 pandemic. Auto-alignment was achieved using custom software to align the device to the patient pupil, with real-time pupil detection via a deep-learning algorithm. Auto-capture was triggered once the pupil was detected at the target location. We evaluated alignment success rate (alignments that triggered a capture), mean ± SD time to align, and accuracy of alignment (mean ± SD distance between detected pupil center and target).
Results :
Of 24 auto-alignments (1-3 per eye), 23 successfully triggered a capture; 1 failed due to a software bug. All 14 manual alignments (1-3 per eye) proceeded to capture. Time to align for auto-alignment (29 ± 15 s) was significantly faster than for manual alignment (72 ± 37 s). The accuracy of alignment was 0.94 ± 0.59 mm for auto and 0.97 ± 0.41 mm for manual operation. A two-sample t-test assuming independent alignments for auto and manual results did not demonstrate a statistically significant difference in accuracy (p > 0.05).
Conclusions :
This comparison of manual and auto-alignment and capture for a motorized device indicated that this auto-alignment method is 2.5x faster than motorized manual alignment, saving an average of 40 s per alignment. Auto-operation was comparable in accuracy, and reliable on normal eyes. While further investigations are needed for a clinical population, this method shows promise for utility in a clinical setting.
This is a 2021 ARVO Annual Meeting abstract.