ECM-TransUNet: Edge-enhanced multi-scale attention and convolutional Mamba for medical image segmentation

The segmentation of CT and MRI images faces challenges such as detail loss and the inability to ensure consistency in physiological tissue representation. To address these issues, we propose a Edge-enhanced multi-scale attention and Convolutional Mamba Transformer UNet (ECM-TransUNet). ECM-TransUNet integrates the ECMBlock into the skip connections, incorporating the Edge-Enhanced Multi-Scale Transposed Attention (E-MTA) and the Multi-Scale Convolutional State-Space Module (MS-CSM) to improve feature extraction and spatial consistency modeling. Specifically, E-MTA enhances sensitivity to subtle grayscale variations, enabling accurate modeling of both local and global structural details in complex regions. Unlike traditional attention mechanisms, E-MTA integrates multi-scale depthwise convolutions to strengthen local feature representation, while the Sobel edge detection module further refines the extraction of critical edges and local detail features. MS-CSM combines state-space modeling with multi-scale feature extraction to improve the accuracy of local detail representation and global feature integration, while significantly reducing computational complexity. Compared to traditional convolution-based methods and earlier state-space models, it demonstrates superior performance and efficiency. Additionally, to achieve end-to-end feature balance within skip connections, we introduce the Cross-Region Multi-Scale Attention (CR-MSA) mechanism into the Transformer-based encoder architecture. CR-MSA effectively harmonizes multi-scale and spatial feature fusion, establishes cross-regional feature relationships, and enhances the model's ability to capture both local and global information, thereby further improving segmentation accuracy and stability. Our method effectively addresses the limitations of existing medical image segmentation techniques. Experimental results on large-scale annotated CT and MRI datasets demonstrate that our approach achieves an optimal balance between segmentation accuracy and computational efficiency. Specifically, on the Synapse dataset, ECM-TransUNet achieved a DSC of 84.68 %, with a computational cost of 50.68G FLOPs and a parameter count of 66.47 M. These findings underscore the reliability and efficiency of our method, offering a robust solution for complex medical image segmentation tasks. is available at: https://github.com/ lvchunjie/ECM-TransUNet.