Thyroid nodules are common and their accurate classification as benign or malignant is essential for treatment. Ultrasound is the main diagnostic tool, but its accuracy depends on the operator. The proposed two-branch deep learning framework combines node segmentation and classification. Key is the node mask-driven feature enhancement module, which uses probability masks from the segmentation branch to focus the classification branch on relevant regions and suppress irrelevant information. The framework was tested on ultrasound data from three medical centers. It achieved excellent classification accuracy compared to basic methods. This approach has potential as a reliable computer-aided diagnostic tool.