Figure 5.
The U-Net++ and attention U-Net architectures are advanced variants of the original U-Net, designed to improve performance in image segmentation tasks, particularly in medical imaging. U-Net++ aims to improve the standard U-Net by addressing the semantic gap between the encoder and decoder feature maps connected by skip connections. It achieves this through a nested, densely connected structure. Instead of a single U-shaped path, U-Net++ incorporates multiple U-Net-like structures nested within each other. This creates pathways of varying depths (Nested U-Net Structure). Attention U-Net enhances the U-Net architecture by integrating attention mechanisms, specifically Attention Gates (AGs), into the skip connections. AGs are placed in the skip connections between the encoder and decoder. They learn to suppress irrelevant regions in the input image and highlight salient features relevant to the target structure.
