MLU-Net: A Multi-Level Lightweight U-Net for Medical Image Segmentation Integrating Frequency Representation and MLP-Based Methods

Medical image segmentation is a challenging and popular task in the field of medical image processing in recent decades. Most of the current mainstream segmentation networks are based on convolutional neural networks (CNNs) methods. Among them, encoding and decoding structures based on U-Net architecture and skip connection mechanism have made great progress in medical segmentation. However, these networks come with increased complexity and training difficulty as the accuracy of network segmentation continues to increase, and their ability remains to be improved for extracting feature information in specific information-intensive structure segmentation tasks, such as brain tumors. In addition, the high training cost raises the application threshold of medical image segmentation. To address these issues, we introduce a frequency representation approach that can effectively reduce the loss of feature during encoding and decoding of segmentation networks. Then a tokenized multi-layer perceptron (MLP) method is introduced to learn the space information. Frequency representation and tokenized MLP can greatly reduce the parameters and computational effort while achieving more accurate and efficient medical image segmentation. Therefore, a multi-level lightweight U-Net segmentation network named MLU-Net is proposed to perform segmentation tasks of medical images quickly. In brain tumor segmentation experiments under equivalent preprocessing conditions, our network achieves substantial efficiency gains with parameter and computational workload reductions to 1/39 and 1/61 of U-Net's, while simultaneously demonstrating superior performance, enhancing the Dice and Intersection over Union (IoU) metrics by 3.37% and 3.30%, respectively. In addition, we perform experiments on dermatologic data and still achieve segmentation performance that outperforms comparable networks. These experiments show that the proposed network is characterized by lightweight and high accuracy, which is contributing to the exploration of clinical medicine scenarios.