Fig. 2

The overall study process

The most representative image of each tumour on the thyroid multimodal ultrasound image was selected. Radiomics features, including shape, histogram, absolute gradient, grey-level co-occurrence matrix, run length matrix, autoregressive model, and wavelet transform, were extracted. Radiomics features were generated using the Fisher coefficient, mutual information, probability of classification error and average correlation coefficient methods (F + MI + PA) and LASSO. These selected features were used to train the linear SVM in fivefold cross-validation and test in an test set. Univariate analysis was performed to determine the association between the clinical variables and ETE. Another SVM classifier was built using radiomics features plus clinical variables and observed ultrasound characteristics. SVM: support vector machine