VMDU-net: a dual encoder multi-scale fusion network for polyp segmentation with Vision Mamba and Cross-Shape...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Colorectal cancer often arises from polyps, necessitating early detection and removal.
• Accurate polyp segmentation is vital for preventing cancer progression but faces challenges like diverse polyp morphology and unclear boundaries.
• Existing segmentation algorithms struggle with long-range dependencies and effective convergence.

Purpose of the Study:

• To introduce VMDU-Net, a novel Dual Encoder Multi-Scale Feature Fusion Network for enhanced polyp segmentation.
• To overcome limitations in current polyp segmentation methods, particularly in handling complex polyp characteristics and improving convergence.
• To advance early detection and prevention of colorectal cancer through improved polyp identification.

Main Methods:

• Developed VMDU-Net, featuring parallel encoders with Vision Mamba and Cross-Shape Transformer modules.
• Integrated a Mamba-Transformer-Merge (MTM) module for attention-weighted fusion of spatial and channel features.
• Utilized Depthwise Separable Convolutions for multi-scale feature extraction and improved convergence efficiency.

Main Results:

• VMDU-Net demonstrated superior performance over state-of-the-art methods on five polyp segmentation datasets.
• Achieved high Dice scores: 0.934 on Kvasir-SEG and 0.951 on CVC-ClinicDB, indicating excellent segmentation accuracy.
• Showcased significant improvements in preserving boundary details of polyps.

Conclusions:

• VMDU-Net effectively addresses key challenges in polyp segmentation by integrating Mamba and Transformer architectures.
• The network's robust performance across diverse datasets suggests strong potential for clinical applications in early colorectal cancer prevention.
• The proposed method enhances segmentation accuracy and boundary preservation, crucial for effective polyp removal.