Abstract
BACKGROUND: The rising global incidence of thyroid nodules necessitates improved non-invasive methods for differentiating benign from malignant lesions. However, research on artificial intelligence (AI) models using multiphase CT imaging to differentiate benign from malignant thyroid nodules is limited. METHODS: This retrospective study analyzed multiphase CT data (noncontrast, arterial, and venous phases) from 604 patients with thyroid nodules confirmed by postoperative pathology. We developed and compared multiple machine learning and deep learning models using extracted radiomics features, raw 3D DICOM data, and key clinical factors (sex, age, thyroglobulin and thyrotropin levels). Model performance was evaluated using receiver operating characteristic (ROC) analysis, and Gradient-weighted Class Activation Mapping (Grad-CAM) was used for visualization. RESULTS: Models incorporating imaging data significantly outperformed a clinical-only model (AUC = 0.811). Nomograms combining either a radiomics score (Rad-Score) or a deep learning score (AI-Score) with clinical data demonstrated the highest diagnostic accuracy. The nomogram based on Rad-Score and clinical data achieved a peak AUC of 0.885. Similarly, the AI-Score-based nomogram reached an AUC of 0.881. Both integrated approaches proved superior to models relying on a single data type. CONCLUSIONS: AI models integrating multiphase CT radiomics or deep learning features with clinical data provide a robust and highly accurate approach for differentiating benign from malignant thyroid nodules. These integrated models show significant potential for improving clinical decision-making.