Machine learning model incorporating spectral CT quantitative parameters and clinical characteristics for differentiation of benign and malignant thyroid nodules

This study aims to explore the clinical value of multiple machine learning (ML) methods in predicting the benign or malignant nature of thyroid nodules based on preoperative dual-layer spectral computed tomography (DLCT) quantitative parameters and clinical characteristics. The DLCT images and clinical data of 432 thyroid nodule patients were retrospectively analyzed. The patients were randomly divided into a training (n = 302) and validation sets (n = 130). DLCT quantitative parameters included conventional CT value, 40-keV CT value, 70-keV CT value, 100-keV CT value, spectral curve slope, iodine concentration (IC), normalized IC, effective atomic number (Zeff), and normalized Zeff in the arterial phase (AP) and venous phase, and arterial enhancement fraction (AEF). Pearson correlation analysis and feature importance ranking were employed for dimensionality reduction. Nine ML algorithms were evaluated for predicting the benign or malignant nature of thyroid nodules. The area under the curve (AUC) was obtained by plotting the receiver operating characteristic curve. Finally, SHapley Additive exPlanations (SHAP) values were used to interpret variable contributions. Six variables remained after dimension reduction of all DLCT quantitative parameters and clinical characteristics, namely AEF, age, AP-40 keV CT value, FT4, TgAb, and TPOAb. In the validation set, the AUCs of the Extreme Gradient Boosting, Logistic Regression, Random Forest, Adaptive Boosting, Light Gradient Boosting Machine, Multilayer Perceptron Classifier, HistGradientBoostingClassifier, BaggingClassifier, and StackingClassifier models constructed based on the six reduced-dimension variables were 0.79, 0.78, 0.78, 0.82, 0.81, 0.76, 0.78, 0.76, and 0.79, respectively. The SHAP method showed the top three variables ranked by contribution degree were: AEF, age, and AP-40 keV CT value. The multiple ML models constructed based on dual-energy CT quantitative parameters and clinical characteristics in this study can effectively predict the benign or malignant nature of thyroid nodules, among which AEF contributes the most.