Deep Learning System for the Auxiliary Diagnosis of Thyroid Eye Disease: Evaluation of Ocular Inflammation, Eyelid Retraction, and Eye Movement Disorder

Yu Han¹Jun Xie^1*Xiaoyu Li¹Xinying Xu¹Bin Sun²Liu Han^2*Chunfang Yan^2*

• ¹Taiyuan University of Technology, Taiyuan, China
• ²Shanxi Eye Hospital, Taiyuan, Shanxi Province, China

Objective This study aims to construct a semantic segmentation-based auxiliary diagnostic model for thyroid eye disease (TED) focusing on eyelid retraction, eye movement disorders, ocular inflammation related to Clinical Activity Score (CAS), facilitating rapid and non-invasive diagnosis for suspected TED patients and enhancing the efficiency of treatment and diagnosis. Methods Data were collected from 153 subjects exhibiting symptoms of eyelid retraction, eye movement disorders, and ocular inflammation related to CAS. After quality screening, datasets for the primary position (303 eyes), gaze positions (1199 eyes), and a multi-label inflammatory classification dataset (272 eyes) were constructed. The constructed TBRM-Net adopts a dual-branch feature extraction and fusion strategy to extract inflammation features for multi-label classification and recognition; the constructed DSR-Net performs segmentation of ocular structures and has designed a quantitative diagnostic algorithm. Results The semantic segmentation-based auxiliary diagnostic model for TED demonstrated a mean pixel accuracy (MPA) of 94.1% in the primary position dataset and 95.0% in the gaze positions dataset. The accuracy for diagnosing eye movement disorders, upper eyelid retraction, and lower eyelid retraction reached 85.4%, 95.1%, and 87.0%, respectively. The accuracy for Redness of Eyelids, Swelling of Eyelids, Redness of Conjunctiva, Swelling of Conjunctiva, and Swelling of Caruncle or Plica reaches 81.8%, 78.8%, 90.6%, 73.5%, and 83.9%, respectively, with an average accuracy of 81.7%. Segmenting and classifying images of structures affected by ocular inflammation can effectively exclude interfering features. The designed quantitative algorithm provides greater interpretability than existing studies, thereby validating the effectiveness of the diagnostic system. Conclusion The deep learning-based auxiliary diagnostic model for TED established in this study exhibits high accuracy and interpretability in the diagnosis of ocular inflammation related to CAS, eyelid retraction, and eye movement disorders. It holds significant medical value in assisting doctors in formulating treatment plans and evaluating therapeutic effects.