An Efficient and Rapid Medical Image Segmentation Network.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Accurate medical image segmentation is crucial for quantitative analysis.
• Current deep learning models like UNet achieve high performance but are computationally intensive.
• This limits their use on resource-constrained devices.

Purpose of the Study:

• To develop a lightweight yet effective model for medical image segmentation.
• To reduce the computational complexity and parameter count of existing segmentation networks.
• To maintain high segmentation accuracy for medical imaging applications.

Main Methods:

• Proposed SHFormer, a shallow hierarchical Transformer architecture.
• Introduced a spatial-channel connection module for focused attention.
• Developed an MLP-D module for lightweight multi-scale feature fusion.

Main Results:

• SHFormer demonstrated comparable performance to state-of-the-art networks on the ISIC-2018 dataset.
• Achieved 15x fewer parameters, 30x lower computational complexity, and 5x higher inference efficiency.
• Validated generalizability on a polyp dataset with similar results.

Conclusions:

• SHFormer provides an efficient alternative for medical image segmentation.
• The proposed lightweight modules enable high performance with reduced computational overhead.
• SHFormer is suitable for deployment on devices with limited computational resources.