Thyroid nodules classification in ultrasound images is actively researched in the field of medical image processing. However, due to the low quality of ultrasound images, severe speckle noise, the complexity and diversity of nodules, etc., the classification and diagnosis of thyroid nodules are extremely challenging. At present, deep learning has been widely used in the field of medical image processing, and has achieved good results. However, there are still many problems to be solved. To address these issues, we propose a mask-guided hierarchical deep learning (MHDL) framework for the thyroid nodules classification. Specifically, we first develop a Mask RCNN network to locate thyroid nodules as the region of interest (ROI) for each image, to remove confounding information from input ultrasound images and extract texture, shape and radiology features as the low dimensional features. We then design a residual attention network to extract depth feature map of ROI, and combine the above low dimensional features to form a mixed feature space via dimension alignment technology. Finally, we present an AttentionDrop-based convolutional neural network to implement the classification of benign and malignant thyroid nodules in the mixed feature space. The experimental results show that our proposed method can obtain accurate nodule classification results, and hierarchical deep learning network can further improve the classification performance, which has immense clinical application value.