Investigating the Value of B-Mode and Contrast-Enhanced Ultrasound Based Radiomics Features in Differentiating Chinese TI-RADS Category 4a and 4b Micro-nodules

Abstract Background: This study aimed to evaluate the clinical utility of radiomics features in differentiating Chinese Thyroid Imaging and Data System (C-TIRADS) category 4a and 4b thyroid micro-nodules using B-mode ultrasound (BMUS) and contrast-enhanced ultrasound (CEUS) images. Methods: Radiomics features were extracted from BMUS and CEUS images using Intelligence Foundry software. Three radiomics models (BMUS, CEUS, and BMUS+CEUS) were developed using machine learning algorithms. Diagnostic performance of these models and experienced radiologist's diagnosis were evaluated using receiver operating characteristic curves (ROC) area under the curve (AUC). Delong test was used to compare diagnostic performance differences among these models. Results: The BMUS+CEUS radiomics model exhibited the highest diagnostic performance in both the training (AUC=0.996, 95% CI, 0.966-1.000) and validation (AUC=0.897, 95% CI, 0.816-0.951) cohorts compared to the other two radiomics models. All three radiomics models demonstrated better diagnostic performance than the experienced radiologist's diagnosis, which achieved an AUC of 0.717 (95% CI, 0.634-0.791) in the training cohort and 0.692 (95% CI, 0.587-0.784) in the validation cohort. Conclusion: A noninvasive model combining BMUS and CEUS radiomics features has the potential to accurately distinguish the nature of C-TIRADS category 4a and 4b thyroid micro-nodules preoperatively. The BMUS radiomics model could also be a good clinical choice when CEUS is absent.