ÆMMamba: An Efficient Medical Segmentation Model With Edge Enhancement

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is vital for clinical applications but challenged by image complexity.
• Existing methods like CNNs and Vision Transformers have limitations in modeling long-range dependencies efficiently.
• State Space Models offer potential for efficient long-range dependency modeling.

Purpose of the Study:

• To introduce ÆMMamba, a novel multi-scale feature extraction framework for enhanced medical image segmentation.
• To address the limitations of current models in capturing both local and global image features.
• To improve segmentation accuracy across diverse medical imaging modalities and datasets.

Main Methods:

• Developed ÆMMamba based on the Mamba backbone, incorporating an Efficient Fusion Bridge (EFB) for multi-scale feature fusion.
• Integrated an Edge-Aware Module (EAM) using Sobel-based edge extraction to enhance low-level features.
• Utilized a Boundary Sensitive Decoder (BSD) with inverse attention and residual convolutions for complex boundary handling.

Main Results:

• ÆMMamba achieved state-of-the-art performance on 8 medical segmentation datasets.
• Demonstrated superior performance in polyp segmentation (e.g., 72.22 mDice on ETIS) compared to MADGNet and Swin-UMamba.
• Outperformed H2Former and SwinUnet in lung and breast segmentation (e.g., 84.24 Dice on BUSI, 79.83 on COVID-19 Lung).
• Achieved high scores on LGG brain MRI segmentation (87.25 mDice, 79.31 mIoU).

Conclusions:

• ÆMMamba represents a significant advancement in medical image segmentation, outperforming existing methods.
• The proposed framework effectively integrates multi-scale features and boundary information for robust segmentation.
• ÆMMamba shows great potential for various clinical applications requiring precise image segmentation.