Adequate ultrasound image classification of the thyroid is important for timely detection and efficient treatment of thyroid diseases. A machine learning pipeline is proposed in this study that makes use of histogram-based feature extraction (HOG) and variance thresholding for extracting features to increase classification accuracy. Out of a collection of 5,893 ultrasound images, more than 26,244 features were first extracted, and dimensionality reduction was achieved through a 0.01 variance threshold. The processed feature set was trained in three classifiers namely Support Vector Machine (SVM), Random Forest (RF) and Extreme Gradient Boosting (XGBoost). Out of these, a validation accuracy of 0.83 and balanced recall of both malignant and benign cases were obtained by XGBoost. These findings suggest that the proposed approach is an interpretable and efficient solution for computer-aided disease diagnosis in the context of the thyroid.