Multi-dimensional CT feature screening, construction, and validation of a clinical diagnostic model for thyroid eye disease.

OBJECTIVE

Thyroid eye disease (TED) is an autoimmune condition associated with thyroid dysfunction, often presenting with complex and variable orbital manifestations that pose challenges for early and objective diagnosis. Current diagnostic reliance on clinical assessment and conventional imaging may lack sensitivity or standardization, highlighting an unmet need for quantitative and reproducible tools. To address this gap, this study developed and validated a clinical diagnostic model for TED screening through multidimensional analysis of computed tomography (CT)-derived three-dimensional reconstruction and multiplanar reconstruction data. By systematically screening sensitive imaging indicators across various parameters, the research quantitatively established, expanded, and verified the clinical utility of CT-based measurements in TED detection.

METHODS

Based on orbital CT imaging data, we utilized Mimics and RadiAnt Viewer software to perform semi-automated quantitative measurements of four-dimensional parameters (point: CT values of extraocular muscles, orbital fat, and optic nerves; line: exophthalmos degree and interocular difference; plane: cross-sectional areas of extraocular muscles and optic nerves; volume: volumetric measurements of extraocular muscles, optic nerves, orbital fat, and orbital cavity) in both thyroid eye disease (TED) patients and control subjects. Statistical analyses were conducted using SPSS software and R language. The dataset was randomly split into training and validation sets in a 7:3 ratio. In the training set, least absolute shrinkage and selection operator (LASSO) regression with cross-validation was employed for feature screening to reduce overfitting risk, followed by logistic regression analysis to further determine modeling indicators. Subsequently, a clinical diagnostic model was constructed and visualized as a nomogram. Model performance was evaluated by calculating the area under the receiver operating characteristic (ROC) curve in both training and validation sets. The classification accuracy at different risk cutoff points in the internal validation set was computed to assess model performance across varying risk thresholds. Finally, calibration curves and decision curve analysis (DCA) were performed to validate the probability accuracy and net benefit thresholds ofthe model.

RESULTS

The distribution of data in the control group was more concentrated and positively skewed than that in the TED group; Among the four-dimensional CT measurements, the most relevant indicator for TED screening was the volume of the inferior rectus muscle (AUC = 0.92). At the two-dimensional level, the maximum cross-sectional area along the long axis of the inferior rectus muscle showed the highest correlation (AUC = 0.84); We successfully developed a highly accurate CT-based TED diagnostic model, achieving AUC values of 0.959 (training set) and 0.918 (validation set).

CONCLUSION

The clinical diagnostic model constructed based on CT-derived measurements demonstrates considerable accuracy in TED screening, showing improved performance compared to single screening indicators. This advancement holds significant clinical value. Integrating machine learning, deep learning, and radiomics approaches could further enhance its accuracy, practicality, and clinical workflow efficiency.