Explainable Deep Learning System for Automatic Detection of Thyroid Eye Disease Using Facial Images

Purpose

To report an explainable deep learning (XDL) system to automatically detect thyroid eye disease (TED) using facial images.

Design

Prospective study to develop and evaluate a deep-learning diagnostic algorithm.

Methods

A dataset consisting of 302 and 289 facial images of newly diagnosed, treatment-naïve, TED patients and healthy subjects were compiled, annotated, and applied to train the XDL model. It consisted of a periocular landmarks localization network that identified the periocular landmarks on facial images, and the TED detection network (TDN) that uses a binary classification to detect TED using facial images. The generalizability of the XDL system was evaluated using a threefold cross-validation strategy and further validated using 100 facial images of TED patients from an independent thyroid eye clinic.

Results

The area under the receiver operating characteristic curve was 99.7%, sensitivity 99.7%, and specificity 94.5% (95% confidence interval: 99.6%-99.9%). Heatmaps demonstrated upper and lower eyelids as key regions of interest. The validation cohort achieved area under the receiver operating characteristic curve of 98.9%, sensitivity 92%, and specificity 93%.

Conclusions

This XDL system detected TED using facial images with excellent accuracy and explainability. It should be further evaluated in prospective Graves’ disease cohorts at nonspecialist setting for early detection and referral of progressive TED.