Multi-Attentional U-Net for Medical Image Segmentation

Abstract

Deep learning (DL) approaches for image segmentation have been gaining state-of-the-art performance in recent years. Particularly, in deep learning, U-Net model has been successfully used in the field of image segmentation. However, traditional U-Net methods extract features, aggregate remote information, and reconstruct images by stacking convolution, pooling, and up sampling blocks. The traditional approach is very inefficient due of the stacked local operators. In this paper, we propose the multi-attentional U-Net that is equipped with non-local blocks based self-attention, channel-attention, and spatial-attention for image segmentation. These blocks can be inserted into U-Net to flexibly aggregate information on the plane and spatial scales. We perform and evaluate the multi-attentional U-Net model on three benchmark data sets, which are COVID-19 segmentation, skin cancer segmentation, thyroid nodules segmentation. Results show that our proposed models achieve better performances with faster computation and fewer parameters. The multi-attention U-Net can improve the medical image segmentation results.