A Multi-Scale Grid Attention Based Network for Blood Vessel Extraction from Retinal Fundus Images

Accurate blood vessel extraction in medical images is a key step in the diagnose of vascular-related diseases and can also provide important guiding information for surgery. However, blood vessels are characterized by elongated and tortuous, irregular shapes, making the task of vessel segmentation challenging. Multi-scale contextual features can help to better understand the structure and morphology of an image, and attention mechanisms plays a quite essential role in the abstraction of contextual features. In this paper, we propose a multi-scale dual-branching blood vessel image segmentation method based on the grid attention mechanism. In the encoder, a dual combination of CNN and grid attention is used for multi-scale contextual information extraction, and then channel attention is utilized to fusion the feature representations extracted in a weighted manner, which can be better adapted to complicated vascular structures. In the decoder, each feature map of the encoder layer is up-sampled by sub-pixel convolution, and subsequently, the high-resolution feature maps obtained are concatenated to generate the final segmentation mask by further convolution operations. This design can fully utilize the multi-scale feature information extracted by the encoder while preserving the spatial structure and details of the image. We conducted experiments on two public retinal fundus image datasets and the results demonstrate that our approach outperforms CNN-based, attention-based and state-of-the-art CNN-Attention combined models.