Related Experiment Video
Updated: Jan 16, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
DMM-UNet: dual-path multi-scale Mamba UNet for medical image segmentation
Liquan Zhao1, Mingxia Cao1, Yanfei Jia2
1Northeast Electric Power University, Key Laboratory of Modern Power System Simulation and Control and Renewable Energy Technology, Ministry of Education, Information and Communication Engineering, Jilin, China.
The dual-path multi-scale Mamba UNet (DMM-UNet) model enhances medical image segmentation by effectively integrating local and global features. This novel approach improves multiscale detail extraction and contextual information fusion for better segmentation performance.
Area of Science:
- Medical Image Analysis
- Deep Learning Architectures
- Computer Vision
Background:
- State space models offer linear complexity for long-range dependencies in medical image segmentation.
- Existing models struggle to capture local features, limiting multiscale detail extraction and global-local information integration.
Purpose of the Study:
- To introduce the dual-path multi-scale Mamba UNet (DMM-UNet) model.
- To address the limitations of state space models in capturing local features and integrating multiscale contextual information for medical image segmentation.
Main Methods:
- Developed a U-shaped encoder-decoder framework with deep fusion of local and global features.
- Introduced multi-scale channel attention selective scanning blocks in the encoder for simultaneous long-range and local dependency modeling.
- Designed spatial attention selective scanning blocks in the decoder for precise semantic feature aggregation.
- Implemented multi-dimensional collaborative attention layers for cross-space-channel feature interaction.
Main Results:
- Achieved a Dice similarity coefficient of 89.88% on ISIC17, 90.52% on ISIC18, 83.07% on Synapse, and 92.60% on ACDC datasets.
- Demonstrated strong performance across multiple evaluation metrics on benchmark datasets.
Conclusions:
- The DMM-UNet model successfully integrates local and global features, overcoming state space model limitations.
- The proposed architecture enhances multiscale feature fusion and improves segmentation performance in medical imaging.
- DMM-UNet offers a promising solution for advanced medical image segmentation tasks.
