DAGU-Net: Cascaded multi-scale aware network based on dual attention grouping module for medical image segmentation

Abstract

Automatic segmentation of medical images based on convolutional neural networks has achieved outstanding success in the computer-aided diagnosis owing to the powerful feature representation. Besides, numerous image feature extraction methods based on attention mechanisms have been proposed to improve the accuracy of medical image segmentation, such as methods based on spatial attention, channel attention or Transformer. However, attention based methods utilizing the specialized modules to extract valuable information from basic features increase the complexity of models only to obtain better features for specific targets. An encoder–decoder architecture based on the dual attention grouping module and cascaded multi-scale structure (DAGU-Net) is proposed for medical image segmentation, which can adaptively extract features for input images and utilize multi-scale features to generate more precise probability maps. Concretely, the dual attention grouping module designed by the spatial and channel attention is taken as the basic convolutional block of the U-shape network. In addition, the cascaded multi-scale structure is conducted on encoder features to pass multi-scale contexts to the decoder part, significantly improving the quality of semantic segmentation. Extensive comparative experiments show that our method DAGU-Net surpasses eight state-of-the-art segmentation methods on three publicly available medical image datasets.