RE-TransUNet: An Abdominal Multi-Organ Segmentation Network Based on RoPE and TransUNet

Abdominal multi-organ segmentation can accurately extract regions of interest, provide a key basis for disease diagnosis, treatment, and evaluation, and is of significant research importance. However, due to the blurred boundaries of abdominal images, complex anatomical structures, and variable shapes and sizes, achieving effective image segmentation remains a highly valuable research problem. Firstly, in order to enhance the model's ability to model both short-range and long-range features, a network architecture named RE-TransUNet has been proposed by combing Vision Transformer and ResNet50. Secondly, rotational position encoding (RoPE) is introduced to better capture the relative positional information between pixels. Finally, a multidimensional feature fusion (MDFF) module has been designed to enhance the representation capability of multi-scale features. The performance of RE-TransUNet has been evaluated on the Synapse dataset, achieving an average Dice similarity coefficient of $\mathbf{8 2. 8 8 \%}$ and an average Hausdorff distance of 23.41 mm. The experimental results show that the proposed RE-TransUNet network structure can improve the effect of medical image segmentation with almost no increase in computational burden.