IB-TransUNet: Combining Information Bottleneck and Transformer for Medical Image Segmentation

Medical image segmentation plays an important role in disease diagnosis and surgical guidance. There are two problems in the current field of medical image segmentation. First, due to the inherent locality of convolution operations, it is difficult for convolutional neural network models to capture global context information. Second, the data set is usually small and the model is at risk of overfitting. To solve the above problems, we innovatively introduced Transformer and information bottlenecks based on the UNet model (IB-TransUNet). Transformer can capture global context information. Information bottleneck can compress redundant features and reduce the risk of overfitting in medical image segmentation tasks. Furthermore, we add a multi-resolution fusion mechanism to skip connections, which helps high-resolution feature maps to have both spatial texture information and semantic information. Finally, a channel attention block with residuals is added to the decoder to help the model learn relevant features. To verify the performance and efficiency of the proposed model, we conduct ablation experiments on two public datasets and compare them with those state-of-the-art models. Experimental results demonstrate the advantages of the proposed model.