The automatic classification of thyroid nodules in ultrasound images is a critical research focus in medical imaging. However

publicly available thyroid ultrasound datasets remain scarce. In this study

we developed the Ultrasound Dataset for Thyroid Nodules (UD-TN)

a comprehensive dataset containing 10

495 labeled images classified as benign or malignant based on pathology-confirmed results. To establish a benchmark

we proposed the Thyroid Ultrasound Image Neural Network (ThyUNet)

a deep learning model designed for accurate nodule classification. By incorporating high-resolution feature enhancement

instance normalization

and dilated convolutions into residual blocks

ThyUNet excels in extracting fine-grained features

particularly for small nodules. Experimental results demonstrate that ThyUNet achieves state-of-the-art performance

with an accuracy of 89.7%

a sensitivity of 0.879

and a specificity of 0.910 on the testing set. These results surpass those of other advanced architectures

highlighting the model’s effectiveness. UD-TN and ThyUNet contribute significantly to advancing intelligent medical diagnostics. Dataset details and access instructions are available at https://github.com/18811755633/Sample-of-UD-TN.