FAM: Fully Attention Module for Medical Image Segmentation

Semantic segmentation plays crucial role in image analysis, which needs rich spatial information and contextual representations. Recently, deep convolution neural networks have achieved much progress in semantic segmentation. However, it still faces many challenges due to the complexity of images, like the blur, noise, and similarity, which imped the progress of the image segmentation. In this paper, we rethink the feature fusion attention methods, such as spatial attention and channel attention, and propose a novel Fully Attention Module (named FAM) based on every pixel of feature map without reducing the size of the input feature maps. We integrate the proposed FAM into U-Net and Fast-SCNN to assess its effectiveness on Synapse dataset. The extensive experiments show that FAM could improve the performance of both two architectures with acceptable cost in terms of speed and total number of parameters. Specifically, we improve the average Dice Similarity Coefficient (DSC) 4.31% / 1.07% and average Hausdorff Distance (HD) 17.62mm / 8.21mm comparing to U-Net and FastSCNN, respectively. In summary, our proposed FAM could boost the segmentation performance by extracting the fully attention of each pixel in the feature maps. The code will be made publicly available at https://github.com/apple1986/FAM.