Smartphone-Based Deep Learning Exophthalmometry in Thyroid Eye Disease: A Multicenter, Multiethnic Validation Study.

Background

Proptosis is the principal quantitative measure of disease severity in thyroid eye disease (TED), but Hertel exophthalmometry has substantial interobserver variability and requires in-person specialist visits. Existing deep learning (DL) proptosis-from-photograph models require studio-acquired images. We aimed to validate, in a multicenter multiethnic study, a DL model fine-tuned to estimate proptosis (mm) from nonstudio facial photographs-smartphone selfies or clinician-taken photographs.

Methods

This validation study included 361 patients with confirmed or suspected TED at three institutions. Data were collected prospectively at two Korean sites-Severance Hospital (SH; 203 patients/3654 facial images; internal validation) and Seoul National University Bundang Hospital (SNUBH; 79/933; same-ethnicity external validation)-and retrospectively, from routine-care images, at one Spanish site, Virgen Macarena University Hospital (HUVM; 79/336; different-ethnicity external validation). Korean patients self-acquired selfies on six smartphone models per site using a capture-guidance application; Spanish images were clinician-acquired during routine care without such guidance. The model-a dual-stream ResNet-18 jointly processing a facial image and a per-pixel depth map (ZoeDepth)-outputs a proptosis estimate (mm) per eye from one photograph and was fine-tuned, from a previously published model trained on studio digital single-lens reflex photographs, for these nonstudio acquisitions. Hertel measurements by oculoplastic specialists served as the reference standard.

Results

On single-image analysis, the mean absolute error (MAE) versus Hertel measurement was 1.14 mm (Pearson correlation coefficient [PCC] 0.83) in SH, 1.44 mm (PCC 0.85) in SNUBH, and 1.58 mm (PCC 0.77) in HUVM, with predictions within 2 mm of Hertel in 83.6%, 71.3%, and 69.7% of images, respectively. Performance was consistent across 12 smartphone models (per-device MAE 1.11-1.54 mm). For detecting severe proptosis (Hertel ≥21 mm), the area under the receiver operating characteristic curve was 0.94 (confidence interval 0.93-0.96), 0.91 (0.89-0.92), and 0.86 (0.82-0.91) in SH, SNUBH, and HUVM, respectively. Bland-Altman analysis revealed underestimation at the upper extreme.

Conclusions

An artificial intelligence model fine-tuned for nonstudio facial photographs achieved Hertel-level agreement and reliable severity discrimination across 2 ethnicities and 12 smartphones, both for patient self-acquired smartphone selfies and for clinician-acquired routine clinic photographs, providing a scalable approach for patient-facing and clinic-based monitoring of TED severity, pending further longitudinal validation.