In this study, I introduce a novel workflow for extracting useful features in thyroid ultrasound images using deep learning and machine learning methods. The methodology combines Convolutional Auto-Encoder, Local Binary Patterns, Histogram of Oriented Gradients and professional image characterization together to extract useful information from medical images. Multiple machine learning classifiers are used to build an effective thyroid tumor diagnosis model from extracted features. The experimental results show that Support Vector Machine with a specifically designed preprocessing scheme and a customized objective function outperforms human on the test set. The final model can effectively reduce the number of unnecessary biopsies and the number of missing malignancies.