Abstract ID: 26-633
Artificial intelligence, oculomics and oculometrics in thyroid eye disease: a scoping review of orbital magnetic resonance imaging applications
Author: Marcus Eden-Haigh Base Hospital / Institution: UCL Institute of Ophthalmology, United Kingdom
Presentation Type: ePoster Presentation
Purpose
Thyroid eye disease (TED), a common orbital inflammatory condition, causes significant patient morbidity and potential sight loss. Clinical activity score (CAS)-based prediction of intravenous glucocorticoid (IVGC) response is used to guide management. However, newer, more diverse TED severity and phenotype, diagnostics and management studies require more rigorous, objective methods. We aimed to examine the clinical readiness of artificial intelligence (AI) for orbital magnetic resonance imaging (O-MRI) segmentation and interpretation to support tailored TED management.
Methods
Using expanded search terms: “TED” AND “AI” AND “MRI” via Medline and Embase databases, we found 422 papers. After excluding 411, 11 English-language retrospective studies from China remained. We respectively defined AI diagnostic and segmentation accuracy using the area under the curve (AUC) and the Dice coefficient (AUC and Dice: 0.8-0.9 = good and >0.9 = excellent).
Results
Relevant studies explored different AI tasks, including a 2022 study that found clinicians with AI (AUC=0.921) outperformed clinicians (AUC=0.801) at detecting dysthyroid optic neuropathy. A 2024 study found that a neural network model (NNM) with gradient boost (0.899) outperformed an NNM (0.872) in activity staging. A 2024 study found a logistic regression model (0.936) outperformed radiologists (0.818) at diagnosis. A 2026 study found radiomics with CAS (0.894) outperformed CAS-alone at predicting IVGC success (p ≤ 0.05). A 2026 study found that an NNM with volumetric and functional O-MRIs (0.982) outperformed NNM with volumetric O-MRIs (0.908) in diagnosis and segmentation (Dice > 0.8).
Conclusion
Despite limited evidence in terms of quality and volume, AI O-MRI segmentation and interpretation show potential to support TED patient management decisions. Large, prospective, multicentre external validation and clinical impact studies across diverse populations and countries, TED classifications, MRI sequences and modalities would be better placed to produce more reliable, generalisable results, tailored to heterogeneous TED subtypes.
Additional Authors
| First name | Last name | Base Hospital / Institution |
|---|---|---|
| Jimmy | Uddin | Moorfields Eye Hospital, United Kingdom |
| Mohsan | Malik | Moorfields Eye Hospital / Institute of Ophthalmology, United Kingdom |