The ultrasound-based radiomics-clinical machine learning model to predict papillary thyroid microcarcinoma in TI-RADS 3 nodules

Background: Conventional ultrasound (CUS) technology has proven to be successful in the identification of thyroid nodules. Moreover, the American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) was developed for the purpose of evaluating the risk of thyroid nodules based on ultrasound imaging. Nevertheless, identifying papillary thyroid microcarcinoma (PTMC) from TI-RADS 3 nodules using this system can be difficult due to overlapping morphological features. The main objective of this study was to investigate the efficacy of a machine learning model that utilizes ultrasound-based radiomics features and clinical information in accurately predicting the presence of PTMC in TI-RADS 3 nodules. Methods: A total of 221 patients with TI-RADS 3 nodules were included, consisting of 91 cases of PTMC and 130 benign thyroid nodules. They were randomly divided into training and test cohort in an 8:2 ratio. Radiomics features were extracted from CUS images by manually outlining the targets, while clinical parameters were obtained from electronic medical records. The radiomics model, clinical model, and combined model were constructed and validated to distinguish between PTMC and benign thyroid nodules. Radiomics variables were extracted via the Pyradiomics package (V1.3.0). Moreover, least absolute shrinkage and selection operator (LASSO) regression was used for feature selection. Light Gradient Boosting Machine (LightGBM) was employed to build both radiomics and clinical models. Ultimately, a radiomics-clinical model, which fused radiomics features with clinical information, was developed. Results: Among a total of 1,477 radiomics features, fifteen features that were found to be associated with PTMC through univariate analysis and LASSO regression were selected for the development of the radiomics model. The combined "radiomics-clinical" model demonstrated superior diagnostic accuracy compared to the clinical model for distinguishing PTMC in both the training dataset [area under receiver operating curve (AUC): 0.975 vs. 0.845] and the validation dataset (AUC: 0.898 vs. 0.811). We constructed a radiomics-clinical nomogram, and the clinical applicability was confirmed through decision curve analysis. Conclusions: Utilizing an ultrasound-based radiomics approach has proven to be effective in predicting PTMC in patients with TI-RADS 3 nodules.