Deep Learning−Driven Exophthalmometry through Facial Photographs in Thyroid Eye Disease

Objective

To develop and evaluate a deep learning (DL)-assisted system for proptosis measurement using facial photographs in thyroid eye disease (TED).

Design

A retrospective cohort study.

Participants

This study included 1108 patients with TED from Severance Hospital (SH) and 171 from Seoul National University Bundang Hospital (SNUBH).

Methods

The DL-assisted system was trained using 1610 facial images paired with Hertel exophthalmometry measurements from SH and externally validated using 511 SNUBH images. The system employs a dual-stream ResNet-18 neural network, utilizing both red-green-blue images and depth maps generated by the ZoeDepth algorithm.

Main Outcome Measures

Accuracy was assessed using mean absolute error (MAE), Pearson correlation coefficient, intraclass correlation coefficient (ICC), and area under the curve of the receiver operating characteristic curve.

Results

The DL-assisted system achieved an MAE of 1.27 mm for the SH dataset and 1.24 mm for the SNUBH dataset. Pearson correlation coefficients were 0.82 and 0.77, respectively, with ICCs indicating strong reliability (0.80 for SH and 0.73 for SNUBH). The receiver operating characteristic curve analysis showed area under the curves of 0.91 for SH and 0.88 for SNUBH in detecting proptosis. The system detected significant proptosis changes (≥ 2 mm) with 74.6% accuracy.

Conclusions

The DL-assisted system offers an accurate, accessible method for exophthalmometry in patients with TED using facial photographs. This tool presents a promising alternative to traditional exophthalmometry, potentially improving access to reliable proptosis measurement in both clinical and nonspecialist settings.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.