Objective: To investigate the diagnostic value of radiomics in thyroid nodules of American College of Radiology-Thyroid Imaging Reporting and Data System (ACRTI-RADS) grades 4 and 5. Methods: A retrospective study of 221 patients with ACRTI-RADS grades 4 and 5 who underwent thyroidectomy was performed. The data on clinicopathological and ultrasound examinations of the patients were analyzed. The images were randomized into a training set and a validation set. MaZda, a software for calculating texture parameters in digitized images, was employed to manually outline the images and extract radiomics features. Features were filtered through the univariate Logistic analysis and LASSO method. The Logistic regression model, Bayes model and KNN model were trained using the selected texture features. Results: 314 radiomics features were extracted from each patient’s ultrasound image ROI using the MaZda software package. The use of LASSO further filtered five most significant features. The KNN model, based on these features, performed the best, with ROC values of 0.849 for the training group and 0.885 for the testing group, and accuracy rates of 0.779 and 0.761, respectively. Compared to readings by experienced radiologists, the KNN model demonstrated superior performance. Conclusion: The ultrasound omics model based on thyroid ultrasound, which is superior to experienced physician-performed ultrasound diagnosis, has shown outstanding performance in diagnosing thyroid ACR grade 4, 5 nodules and provided an effective reference for the identification of clinical routine ACR4, 5 nodules.