SSTrans-Net: Smart Swin Transformer Network for medical image segmentation

Medical image segmentation has achieved impressive results through some recent transformer-based works. Especially Swin Transformer has shown the superiority of the method in some segmentation tasks. However, the identical and fixed masks in the Swin Transformer prevent all interactions among ultra-long-range pixels in all channels. It is beneficial to capture long-range dependencies in some channels for multi-organ segmentation. In this paper, we propose a u-shaped Smart Swin Transformer Network (SSTrans-Net) for multi-organ segmentation. In SSTrans-Net, the Smart Shifted Window Multi-Head Self-Attention (SSW-MSA) module is used to replace the module based on the masks in Swin Transformer for learning the different channel-wise features, which focuses on the relevant dependencies among organs. Especially, keep an effective long-range dependency in the channels that exclusively focus on the target distribution and remove that dependency from the channels that concentrate on the local context. In addition, we introduce the Dice and Focal loss functions to supervise the optimization of the Smart Swin Transformer to improve its ability to balance global and local features. Experiments on Synapse and ACDC datasets demonstrate that our strategy requires less computational resources than most segmenters and can significantly improve the segmentation performance of the model. Our code is available at https://github.com/suofer/Smart-Swin-Transformer.