Generalisation of a deep learning-based exophthalmometry system using facial photographs for non-thyroidal orbital disease.

Abstract

To evaluate agreement between photo-based automated exophthalmometry and manual Hertel exophthalmometry in healthy individuals and patients with non-thyroidal orbital disease, with a deep learning system trained exclusively in thyroid eye disease (TED) cohorts applied without retraining or fine-tuning. This was a retrospective observational study. A total of 269 individuals were included in the study, including 95 healthy controls and 174 patients with unilateral non-TED orbital pathology. Automated measurements obtained using Glandy EXO from same-day frontal facial photographs were compared with corresponding manual Hertel exophthalmometry measurements. Agreement was assessed using Pearson correlation coefficient (PCC), mean absolute error (MAE), intraclass correlation coefficient (ICC) and Bland-Altman analysis. Diagnostic performance for clinically significant intereye asymmetry (>2.0 mm) was also evaluated. The system showed good agreement with Hertel exophthalmometry in the eyes of healthy controls (PCC 0.80; MAE 0.86 mm; ICC 0.80) as well as the affected (PCC 0.81; MAE 1.41 mm; ICC 0.75) and fellow (PCC 0.76; MAE 1.03 mm; ICC 0.75) eyes of patients. Overall, 84.8% of measurements were within 2.0 mm of the manual Hertel values. The sensitivity, specificity and accuracy of the system for detecting clinically significant intereye asymmetry were 59.6%, 95.0% and 83.3%, respectively. The TED-trained photoexophthalmometry system showed good agreement with Hertel measurements in healthy individuals and patients with non-TED orbital disease. These findings support its use as an adjunct to clinical examination for objective quantification and documentation of globe position.