U-MLP: MLP-based ultralight refinement network for medical image segmentation

The convolutional neural network (CNN) and Transformer play an important role in computer-aided diagnosis and intelligent medicine. However, CNN cannot obtain long-range dependence, and Transformer has shortcomings in computational complexity and a large number of parameters. Recently, compared with CNN and Transformer, the Multi-Layer Perceptron (MLP)-based medical image processing network can achieve higher accuracy with smaller computational and parametric quantities. Hence, in this work, we propose an encoder-decoder network, U-MLP, based on the ReMLP block. The ReMLP block contains an overlapping sliding window mechanism and a Multi-head Gate Self-Attention (MGSA) module, where the overlapping sliding window can extract local features of the image like convolution, then combines MGSA to fuse the information extracted from multiple dimensions to obtain more contextual semantic information. Meanwhile, to increase the generalization ability of the model, we design the Vague Region Refinement (VRRE) module, which uses the primary features generated by network inference to create local reference features, thus determining the pixel class by inferring the proximity between local features and labeled features. Extensive experimental evaluation shows U-MLP boosts the performance of segmentation. In the skin lesions, spleen, and left atrium segmentation on three benchmark datasets, our U-MLP method achieved a dice similarity coefficient of 88.27%, 97.61%, and 95.91% on the test set, respectively, outperforming 7 state-of-the-art methods.