LM-UNet: Lateral MLP Augmented U-Net for Medical Image Segmentation

With the rapid development and popularization of medical imaging equipment, imaging technology has been widely used in clinical practice, and it has become an indispensable aid for disease diagnosis, surgical planning, and prognosis assessment. Medical images often play a crucial role in the diagnosis and treatment process. In this work, we propose a segmentation model applied to clinically informative medical images, which utilizes a multi-scale hybrid multilayer perceptron (MLP) to establish the connection between local and global information. A hierarchical encoder-decoder architecture based on U-Net and MLP network is proposed. The designed tokenized feature maps, in the Lateral-MLP architecture (LM-UNet), are utilized as lateral connections in multiple levels. The output tokenized sequences inherently have non-local relationships. Hence, it is capable of capturing multi-scale global features, and meanwhile enhancing the boundary information. We have implemented the segmentation experiments on three publicly available datasets. Our LM-UNet outperforms the UNeXt model, achieving an average Dice improvement of 2.07\%, 2.48\% and 0.11\% on the ISIC 2018, OAI-ZIB 2D and BUSI datasets, respectively. This study improves the accuracy and generalization of medical image segmentation. This study explores the merits of MLP sequences and convolution features. The integration of U-Net and MLP-Mixer blocks achieves the improvement on segmentation accuracy in practical medical applications.