Differential Diagnosis of Benign and Malignant Thyroid Nodules Using Deep Learning Radiomics of Thyroid Ultrasound Images

Purpose: We aimed to propose a highly automatic and objective model named deep learning Radiomics of thyroid (DLRT) for the differential diagnosis of benign and malignant thyroid nodules from ultrasound (US) images. Methods: We retrospectively enrolled and finally include US images and fine-needle aspiration biopsies from 1734 patients with 1750 thyroid nodules. A basic convolutional neural network (CNN) model, a transfer learning (TL) model, and a newly designed model named deep learning Radiomics of thyroid (DLRT) were used for the investigation. Their diagnostic accuracy was further compared with human observers (one senior and one junior US radiologist). Moreover, the robustness of DLRT over different US instruments was also validated. Analysis of receiver operating characteristic (ROC) curves were performed to calculate optimal area under it (AUC) for benign and malignant nodules. One observer helped to delineate the nodules. Results: AUCs of DLRT were 0.96 (95% confidence interval [CI]: 0.94-0.98), 0.95 (95% confidence interval [CI]: 0.93-0.97) and 0.97 (95% confidence interval [CI]: 0.95-0.99) in the training, internal and external validation cohort, respectively, which were significantly better than other deep learning models (P < 0.01) and human observers (P < 0.001). No significant difference was found when applying DLRT on thyroid US images acquired from different US instruments. Conclusions: DLRT shows the best overall performance comparing with other deep learning models and human observers. It holds great promise for improving the differential diagnosis of benign and malignant thyroid nodules.