Joint Skeleton and Boundary Features Networks for Curvilinear Structure Segmentation.

Curvilinear structure segmentation has wide-ranging practical applications across many fields. However, existing methods have low topological accuracy when segmenting curved structures, and face difficulties in maintaining complete topological connectivity in their segmentation results. To address these problems, we propose a joint skeleton and boundary feature Encoder-Decoder segmentation network for curved structures. Our method incorporates three decoding branches that extract semantic, skeleton, and boundary features, respectively. Additionally, each decoder output undergoes feature fusion via a joint unit after every layer. Furthermore, adaptive connection units are added between the encoder and decoder to selectively capture information from the encoder. Finally, we perform evaluations on three public datasets for curvilinear structure segmentation tasks, including retinal images for clinical diagnosis, coronary angiography images, and road crack images. Experimental results show that the method outperforms other existing state-of-the-art methods in terms of pixel-level accuracy and topological connectivity.