U-Attention Nested U-Transformer for Medical Image Segmentation

Automatic medical image segmentation facilitates accurate computer-aided diagnosis and computer-assisted surgery. Current segmentation approaches usually use convolutional neural networks that generally get trapped in representing long-range or global features. Based on a combination of convolutional neural networks and vision transformer, this work designs a new deep learning architecture called U- attention nested U-transformer (UANUT) networks for multiorgan segmentation. Specifically, we build an U-attention module by an encoder-decoder U-structure and a channel- spatial attention and create a new convolutional transformer module that replaces multihead self-attention (token mixer) by one-dimensional convolution for feature fusion. The convolutional transformer module in parallel with skip connections at multiple stages effectively detects global and local features, while it also reduces the computation of multihead self-attention. We evaluate our proposed method on two publicly available clinical data, with the experimental results showing that it outperforms other network methods, improving the multiorgan segmentation average dice and average symmetric surface distance from (91.36%, 1.68) to (95.11%, 0.40).