A thyroid ultrasound image region-of-interest (ROI) localization method based on attention mechanisms is proposed to address the problems of background interference and limited dataset size in thyroid ultrasound images. A cross-scale attention interaction strategy is adopted to improve the feature network of the localization model, enabling effective fusion of hierarchical features at different scales. Knowledge distillation is used to transfer the feature extraction ability of the feature network, addressing the problem of network overfitting caused by insufficient data size. Based on statistical data on thyroid morphology in anatomy, a t-distribution mask is designed and combined with attention masks to calculate feature loss, guiding the network to learn key channels and pixel information in thyroid ultrasound images and complete ROI localization. Experimental results show that when the IoU threshold is 0.5, the AP of thyroid ultrasound image ROI localization reaches 92.7%, which is clinically significant and valuable for assisting physicians in diagnosing thyroid diseases.