Automated MRI Analysis Accurately Grades Thyroid Eye Disease Severity

Standardizing the Assessment of Orbital Pathology in Thyroid Eye Disease

Thyroid eye disease is a complex autoimmune disorder that primarily affects the extraocular muscles and periorbital fat, often leading to significant visual impairment and disfigurement [1, 2]. While the Clinical Activity Score remains the standard for bedside evaluation, its inherent subjectivity has driven a search for more reliable imaging biomarkers to guide the use of modern immunomodulators [3]. Magnetic resonance imaging (MRI) has demonstrated fair diagnostic performance in staging disease activity, particularly when utilizing T2-weighted sequences to detect edema within the extraocular muscles [4, 3]. However, the lack of standardized, rapid tools for extracting quantitative data from these scans remains a significant barrier to routine clinical implementation [5, 6]. To address this limitation, researchers have developed an automated deep learning framework designed to provide precise, multiparametric quantification of orbital soft tissues, offering clinicians an objective tool to grade disease severity.

Multicenter Development and Validation of TED-Net

The researchers conducted a retrospective multicenter study to develop and validate the TED-Net framework, utilizing a robust dataset to ensure clinical reliability. The primary development cohort consisted of 330 patients with thyroid eye disease recruited from a single institution. To refine the analysis of disease progression, the researchers categorized a subset of this primary group into 182 patients with mild thyroid eye disease and 138 patients with moderate-to-severe thyroid eye disease. To test the generalizability of the tool across different imaging environments and patient populations, the study also enrolled 113 patients from two external centers for independent validation. The technical foundation of the framework is TED-Net, a deep learning model that integrates ConvNeXt and Transformer architectures. In clinical terms, these are computational frameworks that combine local feature recognition, such as identifying the specific borders of a muscle, with global relationship modeling, which allows the system to understand how different orbital structures relate to one another in three-dimensional space. This integrated architecture enables the model to perform automated, high-precision segmentation of several critical orbital structures, including the extraocular muscles, the lacrimal gland, orbital fat, and the eyeball. By isolating these specific tissues, the system provides a standardized method for quantifying the volumetric and functional changes that characterize thyroid eye disease severity, potentially allowing physicians to track subtle anatomical shifts before they become clinically obvious.

Automated Extraction of Morphological and Functional Biomarkers

To capture the full spectrum of orbital pathology, the researchers utilized a comprehensive imaging protocol consisting of 3 T MRI with water-fat separation and fat-suppressed T2 mapping sequences. This high-field imaging approach allowed the TED-Net framework to automatically extract two distinct categories of quantitative data. The first category involves morphological parameters, specifically the volume and volume ratio of orbital structures, which provide a macroscopic view of tissue hypertrophy or atrophy. The second category consists of functional parameters, which offer insight into the underlying tissue composition and inflammatory status. These functional metrics include the water fraction, the fat fraction, and the fat-suppressed T2 relaxation time (a measure of tissue water content and edema that often correlates with active inflammation in the extraocular muscles). The technical reliability of TED-Net was rigorously evaluated to ensure that its automated measurements could be trusted in a clinical setting. The model achieved Dice similarity coefficients greater than 0.80 across all orbital structures, a statistical metric indicating a high degree of overlap between the automated segmentation and the manual reference standards established by expert radiologists. Furthermore, the volumetric measurements demonstrated high consistency across different imaging sequences, yielding an intraclass correlation coefficient of 0.843 (a measure of reliability showing that the results remain stable across repeated assessments). This level of reproducibility suggests that the tool can provide objective biomarkers for thyroid eye disease regardless of the specific MRI sequence used for the volumetric analysis, facilitating a more standardized assessment of disease severity in routine practice.

Diagnostic Accuracy in Distinguishing Disease Severity

The primary objective of the research was to identify quantitative imaging biomarkers that could facilitate the precise grading of thyroid eye disease severity. In the primary cohort analysis of 182 patients with mild disease and 138 patients with moderate-to-severe involvement, the researchers identified significant differences in both morphological and functional parameters between the two groups (all p < 0.05). These differences spanned the full range of automated metrics, including structural volumes and volume ratios, as well as functional indicators such as water fraction, fat fraction, and fat-suppressed T2 relaxation times. By capturing these distinct physiological changes, the system provides a multidimensional view of orbital pathology that moves beyond simple visual inspection of imaging scans. To evaluate the clinical utility of these metrics, the researchers assessed diagnostic performance using receiver operating characteristic analysis (a statistical method used to determine the accuracy of a diagnostic test by balancing sensitivity and specificity) and decision curve analysis (a technique used to estimate the clinical net benefit of a model across different threshold probabilities). The results indicated that while structural changes alone are highly informative, they do not capture the full clinical picture. The volumetric-only model demonstrated an area under the curve (AUC) of 0.908, but the combined volumetric-functional model achieved a higher AUC of 0.982 for diagnostic accuracy. This high level of precision suggests that integrating tissue composition data with structural measurements significantly improves the ability to distinguish between disease stages. For the practicing clinician, this automated approach offers a standardized method to grade orbital involvement, potentially reducing the inter-observer variability associated with manual measurements or subjective clinical scores. By providing a precise, quantitative baseline and follow-up metrics, these biomarkers may assist in more accurate disease staging and the monitoring of treatment response in patients with complex orbital presentations.

References

1. Poon SHL, Cheung JJ, Shih KC, Chan YK. A systematic review of multimodal clinical biomarkers in the management of thyroid eye disease.. Reviews in endocrine & metabolic disorders. 2022. doi:10.1007/s11154-021-09702-9

2. Zhang H, Liu Y, Zhang Z, et al. Neuroimaging in thyroid eye disease: A systematic review.. Autoimmunity reviews. 2024. doi:10.1016/j.autrev.2024.103667

3. Vaishnav YJ, Mawn LA. Magnetic Resonance Imaging in the Management of Thyroid Eye Disease: A Systematic Review.. Ophthalmic plastic and reconstructive surgery. 2023. doi:10.1097/IOP.0000000000002511

4. Liu S, Sun C, Liu S, Zhu T, Kikkawa DO, Lu W. Orbital MRI for thyroid eye disease activity staging: a systematic review and meta-analysis.. BMC ophthalmology. 2026. doi:10.1186/s12886-026-04714-y

5. Zhang H, Li Z, Chan HC, Song X, Zhou H, Fan X. Artificial intelligence in thyroid eye disease imaging: A systematic review.. Survey of ophthalmology. 2026. doi:10.1016/j.survophthal.2025.07.008

6. Wong NTY, Yuen KFK, Aljufairi FMAA, et al. Magnetic resonance imaging parameters on lacrimal gland in thyroid eye disease: a systematic review and meta-analysis.. BMC ophthalmology. 2023. doi:10.1186/s12886-023-03008-x