Multiscale attention aggregation network for 2D vessel segmentation

Vessel segmentation is essential for clinical diagnosis and surgical planning. However, it is quite challenging for automatic blood vessel segmentation due to low contrast, complex structure, and variable scale, especially when the annotated data is scarce. In this paper, we propose a novel multiscale attention aggregation network (MAA-Net) for vessel segmentation. In MAA-Net, based on a U-shaped encoder-decoder architecture, the dual attention module with scale factors is employed behind the decoder at each stage to generate multi-resolution feature maps adaptively weighted by channel and location attention. In this way, multiscale contextual information with long-range dependencies can be captured to tackle scale variations of vessels. Meanwhile, these attention feature maps are gradually integrated into multi-level aggregate supervision to assemble multiscale context information for refining segmentation results. The proposed method was evaluated on the retinal vessel and coronary angiography dataset (DRIVE and DCA1). Results demonstrate that MAA-Net achieves state-of-the-art performance for vessel segmentation. The code will be available at: \url{https://github.com/lseventeen/MAA-Net-Vessel-Segmentation}.