DPM-UNet: A Mamba-Based Network with Dynamic Perception Feature Enhancement for Medical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Effective medical image segmentation requires integrating both local and global features.
• Existing methods like CNNs struggle with long-range dependencies, while Transformers have high computational costs.
• State Space Models (SSMs) offer a solution with linear complexity for long-range dependency modeling.

Purpose of the Study:

• To propose DPM-UNet, a novel network for medical image segmentation.
• To effectively fuse local and global features using SSMs and other modules.
• To improve the modeling of long-range dependencies and multi-scale information.

Main Methods:

• Developed DPM-UNet incorporating a Dual-path Residual Fusion Module (DRFM) for local features.
• Utilized a DPMamba Module in deep layers for global semantic information and feature fusion.
• Integrated a Multi-scale Aggregation Attention Network (MAAN) to enhance multi-scale representations.

Main Results:

• DPM-UNet demonstrated superior performance in medical image segmentation across three public datasets.
• The method effectively captured local details, long-range dependencies, and multi-scale information.
• Outperformed existing state-of-the-art methods based on multiple evaluation metrics.

Conclusions:

• DPM-UNet offers an effective solution for medical image segmentation by balancing local and global feature extraction.
• The proposed architecture leverages SSMs to efficiently model long-range dependencies.
• The findings suggest DPM-UNet as a promising advancement for medical image analysis tasks.