VMKLA-UNet: vision Mamba with KAN linear attention U-Net

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning Architectures

Background:

• Convolutional Neural Networks (CNNs) struggle with long-range dependencies in medical images.
• Transformer models face computational overhead challenges in medical image segmentation.
• Vision Mamba and KAN linear attention offer promising alternatives for efficient medical image analysis.

Purpose of the Study:

• To introduce VMKLA-UNet, a novel deep learning model for enhanced medical image segmentation.
• To leverage the strengths of Vision Mamba and KAN linear attention for improved feature extraction and context modeling.
• To address the limitations of existing CNN and Transformer architectures in medical image segmentation tasks.

Main Methods:

• The VMKLA-UNet encoder utilizes the VMamba framework for global visual context modeling and efficient feature extraction.
• The decoder incorporates the MKCSA architecture with KAN linear attention and channel-spatial attention mechanisms.
• KAN linear attention reduces computational complexity, while channel and spatial attention refine feature importance and segmentation accuracy.

Main Results:

• VMKLA-UNet demonstrates superior segmentation accuracy and robustness across diverse medical image datasets (Polyp, ISIC 2017/2018, PH², Synapse).
• The model effectively captures global context and enhances segmentation boundary details.
• The integration of attention mechanisms improves differentiation between tissue types and lesions.

Conclusions:

• VMKLA-UNet presents a highly effective and efficient solution for medical image segmentation.
• The proposed architecture overcomes computational and dependency limitations of prior models.
• VMKLA-UNet shows significant potential for clinical applications requiring precise medical image segmentation.