Enhancing vision Mamba with two-dimensional position embedding and multiscale fusion for medical image segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnostics and research.
• Vision Transformers (ViTs) have advanced segmentation but face challenges with long-range dependencies and efficiency.
• There is a need for accurate and computationally efficient segmentation models.

Purpose of the Study:

• To design a deep learning model for accurate medical image segmentation.
• To ensure the model maintains computational efficiency for practical applications.

Main Methods:

• Developed vision Mamba, an efficient visual state space model (SSM).
• Integrated two-dimensional (2D) position embedding to enhance spatial information in patch embedding.
• Incorporated a multiscale feature fusion block (MB) to recover information lost during sequential processing.

Main Results:

• Vision Mamba demonstrated superior performance on three public datasets.
• Achieved average performance improvements of 1.29% (Dice), 1.18% (VOE), 5.52% (ASD), 1.18% (Jaccard), and 1.03% (Recall) compared to a suboptimal model.
• The model maintained computational efficiency alongside enhanced accuracy.

Conclusions:

• The proposed vision Mamba method significantly improves medical image segmentation accuracy.
• The model's computational efficiency makes it suitable for real-world medical applications.
• This approach offers a promising solution for advanced medical image analysis.