ANR-NET: Attention-Guided Non-Local Residual Network for Thyroid Segmentation in Ultrasound Image.

Thyroid gland segmentation in ultrasound images is essential in thyroid disease diagnoses. However, the relatively low quality and the speckle noise in ultrasound images make the organ tissues ambiguous and inhomogeneous. Besides, the thyroid anatomical parts may be separated by a significant distance. In this paper, we proposed a novel Attention-guided Non-local Residual Network (ANR-Net) for thyroid segmentation in ultrasound image, which extracts the short-long relationship to boost thyroid gland segmentation performance. In the proposed network, we deploy two non-local modules to learn the global information of the image at shallow layers. Moreover, the residual features extracted in the first step are used to guide the training of the network in the second step. Experimental results on datasets demonstrate that our proposed network outperforms the state-of-the-art segmentation methods. Our code is available at https://github.com/zyairelu/ANR-Net.