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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Thyroid eye disease (TED) is an autoimmune condition affecting the eye area.
• Early detection and accurate diagnosis are crucial for managing TED and preventing vision loss.

Purpose of the Study:

• To develop and validate an explainable deep learning (XDL) system for automated TED detection from facial images.
• To assess the accuracy and generalizability of the XDL system in identifying TED.

Main Methods:

• A prospective study trained an XDL model on 302 TED and 289 healthy subject facial images.
• The XDL system included a periocular landmark localization network and a TED detection network (TDN).
• Evaluation involved threefold cross-validation and validation on an independent cohort of 100 TED patient images.

Main Results:

• The XDL system achieved 99.7% area under the ROC curve, 99.7% sensitivity, and 94.5% specificity in the training cohort.
• Heatmaps highlighted eyelids as key regions for TED detection.
• The independent validation cohort showed 98.9% area under the ROC curve, 92% sensitivity, and 93% specificity.

Conclusions:

• The XDL system demonstrates excellent accuracy and explainability in detecting TED from facial images.
• Further evaluation in diverse clinical settings, including non-specialist settings, is recommended for early detection of progressive TED.
• This AI-driven approach holds potential for improving patient outcomes through timely diagnosis and referral.