SAA-Net: A Medical Image Segmentation Framework Based on Stream-Across Attention

Deep convolutional networks are widely applied in numerous medical image analysis tasks. Although most convolutional neural network (CNN)-based methods have strong feature representation ability, they are hindered by lousy encoding of long-range interaction information. To alleviate this problem, we propose a novel U-shaped image segmentation network with the Stream-Across Attention module, termed as SAA-Net. Specifically, SAA module contains two types of attention branches, the multi-scale spatial and channel attention branches. Our spatial attention branch attempts to model long-range dependencies and capture local information at the same time by using kernels with different sizes. Channel attention branch can enhance channel information by selectively aggregating the features among all channel maps. In addition, to avoid negative impacts of network deepening, we propose DenseMLP to learn multi-level semantic features. Experimental results show that SAA-Net is competitive in the comparisons with current state-of-the-art (SOTA) methods on two public datasets.