DI-Unet: Dimensional interaction self-attention for medical image segmentation

In recent years, Unet network based on convolution has become a general structure for medical image segmentation tasks. However, it cannot effectively model the long-distance dependence between features due to the limitation of the receptive field. The successful application of Transformer in computer vision solves the problem of the limited receptive field of neural networks. However, the computational complexity limits its further application in medical image segmentation. In addition, the self attention mechanism in Transformer only explores the spatial dimension relationship of the feature maps, and lacks the interaction with the channel dimension, which limits the performance improvement of the network. Here, we proposes DI-Unet, which develops Dimensional Interactive (DI) self-attention for effective feature extraction processing. When inputting high - resolution images, it can effectively reduce the amount of model calculations and capture crossdimensional information before calculating attention weights. The overwhelming superiority of DI-Unet is demonstrated by extensive experiments in multiple databases. In large datasets, the proposed method outperforms other methods in segmentation tasks. The study provides a research foundation and important reference value for the research and application of Transformer structure in medical image segmentation tasks.