VAC-UNet: Visual Attention Convolution U-Net for 3D Medical Image Segmentation

Deep learning-based methods have achieved significant success in medical image segmentation. However, the intrinsic locality of convolutional operations limits the learning of global semantic information. In this paper, we propose a novel method, namely VAC-UNet that is a convolutional transformer U-Net architecture for enhanced global representation. Specially, we introduce a novel Visual Attention Convolution (VAC) block integrating transformers and CNNs for joint local-global feature learning. VAC blocks replace standard convolutions in the encoder-decoder backbone. Asymmetric application on encoders improves localization. Additionally, we design Visual Attention Gates using learned convolution filters to suppress irrelevant regions in skip connections. Finally, we evaluated VAC-UNet on two medical image segmentation tasks against state-of-the-art methods. The results demonstrate VAC blocks effectively capture global context while maintaining local precision. The visual attention improves feature propagation for better segmentation.