关键词: 关键词：甲状腺结节, U-Net, 空洞卷积, 多尺度特征, 坐标注意力

Abstract: Abstract： The ultrasound images of thyroid nodules have serious noise and low contrast between different tissues. The existing ultrasound image segmentation algorithm of thyroid nodules have some problems of blurred edge information and inaccurate segmentation of small nodules. Therefore this paper proposes an ultrasound image segmentation algorithm of thyroid nodules fused with multi-scale spatial features. Based on the U-Net model， the coordinate attention mechanism is introduced to embed the position information into the channel attention to achieve the model’s localization of the thyroid nodule region in the coding part. At the same time， the fused multiscale feature module extracts the spatial aspect features. To retain more detailed features， we uses convolution operation in the process of down sampling and the binary cross-entropy loss and Dice coefficient loss as the comprehensive loss. The experimental results show that compared with the benchmark model U-Net， the proposed algorithm model improves the F1 evaluation index by 9.9 percentage points， and the accuracy rate is increased to 92.8%. Thus the feasibility and effectiveness is verified.

Key words: Key words： thyroid nodule, U-Net, atrous convolution, multi-scale features, coordinate attention