Machine Learning Based Malignancy Prediction in Thyroid Nodules Malignancy: Radiomics Analysis of Ultrasound Images

The aim of this work was to use sonographic image features as biomarker to assess the malignancy of thyroid nodules in patients recommended to FNA according to ACR TI-RADS guideline. Two hundred and ten patients with FNA test report were included in this study. Eighty Different quantitative radiomic features were extracted from sonographic images. Minimum Redundancy Maximum Relevance (MRMR) and logistic regression (LR) algorithms were used as feature selector and classifier, respectively. The evaluation of the models was performed using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). AUC of MRMR feature selection algorithm and LR classifier was 0.87 (with accuracy of 0.74, sensitivity of 0.85 and specificity of 0.60). In the validation dataset, the AUC was 0.92 (with accuracy of 0.70, sensitivity of 0.81 and specificity of 0.58). The proposed model could be potentially used as alternative to FNA as noninvasive tools in clinical setting.