The Pulse
Retrospective study links MRI radiomics model to radiotherapy response in thyroid eye disease
A study reports that MRI radiomics combined with conventional imaging metrics predicted radiotherapy response in thyroid eye disease. In 85 patients, the combined model achieved an AUC of 0.942, compared with 0.700 for a clinical model.
AI.info Team · Written with AI · fact-checked ·
A model built from orbital MRI
Researchers report that an MRI-based model combining radiomic features with conventional imaging metrics predicted response to radiotherapy in people with thyroid eye disease. In a retrospective study, the combined model achieved an area under the curve (AUC) of 0.942, compared with 0.700 for a model based on clinical data.
The study addresses a treatment whose outcomes vary: the authors describe radiotherapy as a treatment for active, moderate-to-severe thyroid eye disease, while noting that patients respond differently. They say conventional indicators such as the clinical activity score (CAS) are subjective and have limited value for predicting response. The article says the model is intended to predict outcomes before treatment using MRI information alongside clinical and imaging measures.
How the researchers tested it
The researchers retrospectively enrolled 85 patients with thyroid eye disease who received standardized radiotherapy at 20 Gy. Before treatment, patients underwent orbital MRI that included fat-suppressed T1-weighted contrast-enhanced imaging, or T1CE, and T2-weighted imaging, or T2WI.
The team extracted radiomic features from the extraocular muscles and selected features for analysis. It developed models using individual MRI sequences, both sequences together, and combinations that added clinical data and conventional imaging metrics. Nine machine learning algorithms were used to develop these models.
To evaluate them, the researchers used DeLong’s test, calibration curves, Brier scores, decision curve analysis and SHAP analysis.
Performance and limits
Of the 85 patients, 46 were classified as responders and 39 as non-responders. Baseline CAS was the only clinical variable that differed significantly between the groups. The clinical model had an AUC of 0.700, while models based on T1CE radiomics, T2WI radiomics and radiomics from both sequences had AUCs of 0.850, 0.875 and 0.933, respectively.
The model combining radiomic features with conventional imaging metrics reached an AUC of 0.942. The authors report that it significantly outperformed the clinical model, with P = 0.040, and describe its calibration and decision-curve results as showing favorable reliability and clinical utility.
The authors say the approach could support individualized risk stratification, but frame that prospect as following prospective validation. The article was published in Communications Medicine on 10 November 2026.