To evaluate the clinical application and diagnostic efficiency of the radiomics features extracted from thyroid B-mode ultrasound (B-US) images, contrast-enhanced ultrasound (CEUS) images, and high-frame-rate contrast-enhanced ultrasound (H-CEUS) images in classifying and predicting benign and malignant thyroid nodules.

We retrospectively collected and ultimately included B-US images, CEUS videos, and H-CEUS videos of 186 nodules from 186 patients. Next, we selected the following eight images to manually select the regions of interest: nodule in transverse section, nodule in longitudinal section, and the "rise time" frame, "time to peak" frame, and "mean transit time" frame of both CEUS and H-CEUS videos, and then extracted the corresponding radiomics features from these images. We established three radiomics models (B-US radiomics model, B-US+CEUS radiomics model, and B-US+H-CEUS radiomics model) based on these radiomics features by Logistic regression and random forest, and assessed the diagnostic performance of these radiomics features. Finally, the area under curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), accuracy (ACC), and F1-score in the test set (n=38) were calculated and compared.

A total of 9000 significant radiomics features of 186 nodules from 186 patients were extracted. The AUC in the differential diagnosis of the nature of thyroid nodules was 0.81 for the B-US radiomics model, 0.87 for the B-US+CEUS radiomics model, and 0.91 for the B-US+H-CEUS radiomics model. The accuracy, sensitivity, specificity, PPV, NPV, and F1 scores of the B-US radiomics model in the test set were 81.6%, 68.8%, 90.9%, 84.6%, 80.0%, and 0.76, respectively; the corresponding values were 84.2%, 68.8%, 95.5%, 91.7%, 80.8%, and 0.79 for the B-US+CEUS radiomics model, and 86.8%, 87.5%, 86.4%, 82.4%, 90.5%, and 0.85 for the B-US+H-CEUS radiomics model.

Radiomics features based on B-US images, CEUS, and H-CEUS key frame images have appreciated clinical value in differentiating benign and malignant thyroid nodules. The diagnostic efficacy of the B-US+H-CEUS radiomics model is better than that of other radiomics models.