A deep learning–based magnetic resonance imaging framework accurately segmented orbital soft tissues and differentiated mild from moderate-to-severe thyroid eye disease in a retrospective multicenter study published in European Radiology. 

Researchers, led by Linhan Zhai of Tongji Hospital in Wuhan, China, developed TED-Net, an automated segmentation model merging ConvNeXt and Transformer architectures, to quantify orbital tissue involvement in patients with thyroid eye disease using 3-T magnetic resonance imaging with water-fat separation and fat-suppressed T2 mapping sequences.  

The study included 330 patients with thyroid eye disease from a primary center for model development and 113 patients from two external centers for validation. From the primary cohort, 182 patients with mild disease and 138 with moderate-to-severe disease were selected for further analysis.  

The model segmented extraocular muscles, lacrimal glands, orbital fat, and the eyeball, then automatically extracted morphologic and functional imaging parameters, including tissue volume, volume ratio, water fraction, fat fraction, and fat-suppressed T2 relaxation time.  

The researchers highlighted three principal findings. First, TED-Net achieved Dice similarity coefficients greater than 0.80 across all orbital structures, indicating high agreement between automated and reference segmentations.  

Second, combining morphologic measurements with functional magnetic resonance imaging biomarkers—including water fraction, fat fraction, and fat-suppressed T2 values—improved discrimination between mild and moderate-to-severe disease compared with volumetric analysis alone, increasing the area under the curve from 0.908 to 0.982.  

Third, the model maintained strong performance during external multicenter validation, supporting the potential utility of quantitative magnetic resonance imaging biomarkers for objective assessment of thyroid eye disease severity across institutions.  

The researchers added that decision curve analysis also supported the clinical utility of the combined model for evaluating orbital involvement.  

Limitations included the study’s retrospective design and reliance on imaging protocols from selected centers, which may affect generalizability. The researchers added that additional prospective multicenter validation is needed before broader clinical implementation.  

No conflicts of interest were reported.