Boundary Attention and Multi-scale Feature Aware Network for Thyroid Nodule Segmentation

Segmentation of thyroid nodules is the key to the diagnosis of thyroid diseases and is important for medical image analysis and clinical diagnosis. Although some progress has been made in this field, we still face many challenges, such as segmentation difficulties due to fuzzy target boundaries, insufficient attention to boundary segmentation accuracy, and the diversity of target scales in medical images. To address these challenges, we propose a boundary attention and multi-scale feature aware network. Specifically, we design a edge prediction module to extract the attention maps of the nodule boundaries to improve the segmentation performance, especially the accuracy of the nodule boundaries. In addition, we propose a pyramid fusion module as well as a multi-scale context fusion module for integrating multi-level global context information and multi-scale feature extraction and aggregation, respectively. In the decoding stage, gradually fusion and supervision are performed based on boundary attention maps and segmentation masks. Numerous quantitative and qualitative experimental results show that the proposed network performs well in the tests of several comparison models.