TAOD-Net: Automated detection and analysis of thyroid-associated orbitopathy in facial imagery

Abstract

Thyroid-Associated Orbitopathy (TAO), a common autoimmune thyroid disease, significantly impacts patients’ quality of life. The conventional method for assessing TAO disease activity relies on the Clinical Activity Score (CAS), which is evaluated by skilled experts. However, the high cost of securing expert evaluators and inconsistencies in their assessments highlight the need for an expert-level, data-driven CAS assessment system. In response, we introduce TAOD-Net (Thyroid-Associated Orbitopathy Detection Network), an advanced data-driven system designed to identify five key CAS components related to inflammatory signs. Leveraging patient facial images as input, our system incorporates a novel learning strategy for multi-label classification and utilizes domain knowledge for optimized image cropping. The performance of TAOD-Net was rigorously validated using 2040 digital facial images collected from 1020 TAO patients at the Department of Ophthalmology, Seoul National University Bundang Hospital. Our results demonstrate that TAOD-Net surpasses existing models in diagnosing TAO disease activity, underscoring its potential to exceed current standards.