Multi-sequence MRI-based radiomics model for predicting radiotherapy response in thyroid eye disease

Radiotherapy is a pivotal treatment for active, moderate-to-severe thyroid eye disease (TED), but responses vary. Conventional indicators such as the clinical activity score (CAS) are subjective and have limited response predictive value. We constructed a multi-sequence predictive model integrating clinical data, conventional imaging, and radiomic features of magnetic resonance imaging (MRI) to predict radiotherapy response in TED patients. 85 TED subjects treated with standardized radiotherapy (20 Gy) were retrospectively enrolled. Pre-treatment orbital MRI, including fat-suppressed T1-weighted contrast-enhanced (T1CE) and T2-weighted imaging (T2WI), was analyzed. Radiomic features from extraocular muscles (EOMs) were extracted and selected to develop single-sequence and multi-sequence models as well as integrated models incorporating clinical data, imaging metrics, and radiomic features using nine machine learning algorithms. Models were evaluated using DeLong’s test, calibration curves, Brier scores, decision curve analysis, and SHAP analysis. Among 85 patients, 46 were responders and 39 were non-responders. Baseline CAS was the only clinical variable significantly different between groups. Clinical models showed limited predictive performance (AUC = 0.700), whereas radiomic models achieved higher discrimination, with AUCs of 0.850, 0.875, and 0.933 for T1CE, T2WI, and multi-sequence radiomics, respectively. The integrated model combining radiomic features and imaging metrics achieved the highest AUC (0.942) and significantly outperformed the clinical model (P = 0.040), with favorable reliability and clinical utility in calibration and decision curve. Multi-sequence MRI-based radiomics integrated with conventional imaging metrics provides an objective approach for predicting radiotherapy response in TED. The combined model demonstrated superior discrimination, favorable calibration, and clinical utility compared with clinical prediction approaches, supporting individualized risk stratification after prospective validatio/n. Thyroid eye disease is an autoimmune orbital disorder that causes orbital inflammation and visual impairment. Orbital radiotherapy is widely used to treat active disease, but treatment responses vary substantially across patients, leaving many individuals exposed to unnecessary radiation without clinical benefits. Currently, no reliable preoperative tools can accurately predict radiotherapy outcomes. Here, we developed a simple and non-invasive predictive model based on multi-sequence MRI radiomics. By integrating multi-sequence radiomic features, orbital conventional imaging metrics and clinical indicators, our optimized model accurately distinguishes radiotherapy responders from non-responders before treatment. This quantitative tool helps guide individualized decision-making, reduce ineffective radiotherapy, and advance personalized management for thyroid eye disease.