Related Experiment Video
Updated: May 28, 2026
Intense Pulsed Light for the Treatment of Dry Eye Owing to Meibomian Gland Dysfunction
Published on: April 1, 2019
Marker-Free, Automated Eyelid Assessment in Thyroid Eye Disease Using Artificial Intelligence: A Multicenter
Jae Hoon Moon1,2,3, Jongchan Kim1, Joonhyeon Park1
1THYROSCOPE INC., Ulsan, Republic of Korea.
Objective:
To validate "Glandy LID," a novel marker-free, deep learning-based software designed for the automated assessment of eyelid morphology using intrinsic corneal diameter for calibration.
Design:
Multicenter retrospective study.
Participants:
The internal validation cohort consisted of 119 patients with thyroid eye disease (TED) from Seoul National University Bundang Hospital (South Korea). The external validation cohort included 140 patients from Hospital Universitario Virgen Macarena (Spain).
Methods:
An artificial intelligence (AI) algorithm segmented eyelid and corneal regions from facial photographs. Eyelid metrics were calculated using population-specific corneal diameters as a reference scale for pixel-to-millimeter conversion. The internal cohort utilized digital single-lens reflex images compared against ground truth derived from physical markers. The external cohort utilized smartphone-captured images compared against clinical medical records to assess clinical generalizability.
Main Outcome Measures:
Geometric precision was assessed using Intersection over Union. Accuracy of quantitative measurements (margin reflex distance 1 [MRD1] and 2 [MRD2]) was evaluated using Pearson correlation coefficient (PCC), mean absolute error, and mean absolute percentage error (MAPE).
Results:
In the internal validation, the AI system demonstrated high geometric precision (Intersection over Union 0.94). Quantitative measurements showed excellent agreement with the ground truth, achieving a PCC of 0.98 for MRD1 and 0.94 for MRD2, with low MAPEs of 5.69% and 4.57%, respectively. In the external validation using smartphone images without markers, the system maintained high reliability. For MRD1, it achieved a PCC of 0.94 and a MAPE of 9.06%. Although MRD2 showed a slightly higher MAPE (15.07%), likely due to manual measurement variability in clinical records, the correlation remained strong (PCC 0.93).
Conclusions:
This AI system provides an accurate and robust automated solution for eyelid measurement without the need for physical reference markers. By overcoming the limitations of manual assessment and marker-based systems, this tool offers a practical and accessible method for objectively monitoring TED in both routine clinical practice and remote health care settings.
Financial Disclosures:
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
