Objective The purpose of this study was to improve the differentiation between malignant and benign thyroid nodules using deep learning (DL) in category 4 and 5 based on the Thyroid Imaging Reporting and Data System (TI-RADS, TR) from the American College of Radiology (ACR). Design and Methods From June 2, 2017 to April 23, 2019, 2082 thyroid ultrasound images from 1396 consecutive patients with confirmed pathology were retrospectively collected, of which 1289 nodules were category 4 (TR4) and 793 nodules were category 5 (TR5). Ninety percent of the B-mode ultrasound images were applied for training and validation, and the residual 10% and an independent external dataset for testing purpose by three different deep learning algorithms. Results In the independent test set, the DL algorithm of best performance got an AUC of 0.904, 0.845, 0.829 in TR4, TR5, and TR4&5, respectively. The sensitivity and specificity of the optimal model was 0.829, 0.831 on TR4, 0.846, 0.778 on TR5, 0.790, 0.779 on TR4&5, versus the radiologists of 0.686 (P=0.108), 0.766 (P=0.101), 0.677 (P=0.211), 0.750 (P=0.128), and 0.680 (P=0.023), 0.761 (P=0.530), respectively. Conclusions The study demonstrated that DL could improve the differentiation of malignant from benign thyroid nodules and had significant potential for clinical application on TR4 and TR5.