MSM-UNet: A medical image segmentation method based on wavelet transform and multi-scale Mamba-UNet

In the field of medical image processing, combining global and local relationship modeling is an effective method for achieving precise image segmentation. Previous studies have demonstrated the remarkable performance of Convolutional Neural Networks (CNNs) in local relationship modeling, while Transformer can directly establish interactions between any two points in an image, thereby effectively capturing global contextual information. However, the application of Transformer to address the shortcomings of Convolutional Neural Networks (CNNs) in modeling global relationships is hindered by their substantial computational complexity and substantial memory demands, posing significant challenges in practice. To address this issue, this paper introduces the Mamba model, a State Space Model (SSM) that exhibits notable advantages in modeling long-range dependencies in sequential data. Inspired by the success of the Mamba model, a two-dimensional medical image segmentation model named MSM-UNet is designed. This model employs a Multi-Scale Mamba feature extraction block (MSMamba), a Wavelet Transform Feature Enhancement Attention Block (WTFEAB), a Feature Enhancement Merge Block (FEMB), and a Fusion Output Layer (FOL), aiming to accurately capture and integrate long-range and local dependencies among multi-scale features. Compared to Transformer-based methods, MSM-UNet exhibits superior performance in holistic feature modeling, significantly improving segmentation accuracy. Tests conducted on the Automatic Cardiac Diagnosis Challenge (ACDC) dataset, the Synapse multi-organ CT abdominal segmentation dataset, and the Colorectal Cancer-Clinic (CVC-ClinicDB) dataset demonstrated that the MSM-UNet achieved Dice coefficients of 92.02, 83.10, and 94.03, respectively. These results comprehensively validate the efficacy and practicality of the proposed MSM-UNet architecture.