Distinguishing benign from malignant thyroid nodules via virtual biopsy: a study on using quantitative parameters and classical radiomics features from dual-energy CT imaging.

BACKGROUND: To evaluate the predictive value of dual-energy computed tomography (DECT)-based quantitative parameters and radiomics models for the preoperative differentiation between benign and malignant thyroid nodules, and to compare their performance with radiologists' interpretations.

METHODS: A retrospective analysis was conducted on 215 patients who underwent contrast-enhanced DECT of the thyroid, with pathological outcomes for thyroid nodules obtained. Patients were randomly assigned to training and testing groups in a 7:3 ratio. The images were evaluated by radiologists. Quantitative parameters derived from DECT were identified through univariate and multivariate logistic regression analyses to construct a DECT model. Radiomics features were extracted from the 40, 70, and 100 keV virtual monochromatic images, as well as iodine-based material-decomposition (IMD) images in the arterial phase (AP) and venous phase (VP) to develop radiomics models from both individual and combined images using a support vector machine (SVM), with the optimal performing model selected as the final radiomics model. Subsequently, a fusion model combining DECT parameters and the radiomics model was established. The diagnostic performances of the radiologist, DECT, radiomics and fusion models were evaluated using receiver operating characteristic (ROC) curves, and clinical usefulness was assessed through decision curve analysis (DCA).

RESULTS: The normalized iodine concentration (NIC) in DECT emerged as an independent factor for the preoperative differentiation between benign and malignant thyroid nodules. The multi-image radiomics model demonstrated excellent predictive performance in the test cohort, achieving an area under the curve (AUC) of 0.966 and was selected as the final model among the radiomics models. The ROC curves indicate that the radiomics model outperformed the radiologist model in predicting the thyroid nature in both the training (0.990 vs. 0.746) and test cohorts (0.966 vs. 0.697) (all p < 0.05). The diagnostic efficacy of the fusion model showed slight improvement over the radiomics model, although this difference was not statistically significant. Furthermore, the fusion model performed exceptionally well in DCA.

CONCLUSIONS: The fusion model, which combined the NIC and the optimal multi-image radiomics model, demonstrates strong diagnostic capability in predicting benign and malignant thyroid nodules.