MixUNet: A Hybrid Retinal Vessels Segmentation Model Combining The Latest CNN and MLPs.

The success of Vision Transformer has led to an increased emphasis on combining global and local context, and its high training cost spawned numerous alternative works. The latest MLP architecture has achieved excellent results as the alternative to Transformer, but the large number of parameters make it difficult to perform on segmentation independently. A U-shaped network MixUNet combining CNN and MLP is proposed to reduce the limitation among data samples by External AttentionMLP in the encoder and decoder, while adding a global context path based on Mix-Scale MLP to capture global information directly, and designing a Multi-scale Vision Attention Module containing multi-scale features and dynamic weights. The experimental validation on DRIVE and STARE shows that MixUNet can achieve better comprehensive performance compared with other medical image segmentation networks.