GAUNet: Gated Attention U-Net for Medical Image Segmentation

Abstract

Despite of their robust power on modeling global dependency, Transformer-based methods for medical segmentation usually depend on complex computation and large-scale pre-training. We aim to design a simple approach without complex computation and pre-training. Gated attention units (GAU) with a single head self-attention performs well for modeling global context features, thus we apply GAU to extract semantic information but it still lacks of localization due to insufficient local details. In this paper, we propose a novel GAU-based hybrid cascaded U-shape GAUNet for medical segmentation method. To compensate the shortcoming of GAU, we design a GAU-Conv module for encoder-decoder to extract global context dependency and reinforce localization. To further improve the performance, we redesign U-Net skip connection with a ReLU to strengthen the detailed localization when decodering. GAUNet achieves the state of the art performance on Synapse multi-organ and cardiac datasets without any pre-trained model.