Lightweight medical image segmentation network with multi-scale feature-guided fusion.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Artificial Intelligence

Background:

• Medical image segmentation is vital for computer-aided diagnosis but faces challenges with limited computing resources.
• Developing accurate, real-time segmentation models that are computationally efficient is a significant ongoing challenge.
• Existing lightweight models often compromise segmentation performance to meet resource constraints.

Purpose of the Study:

• To propose a novel lightweight network for medical image segmentation that optimizes computational efficiency without sacrificing performance.
• To address the trade-off between computational cost and segmentation accuracy in resource-constrained scenarios.
• To enable accurate and fast medical image segmentation on devices with limited computing power.

Main Methods:

• Introduction of a lightweight transformer architecture.
• Development of a simplified core feature extraction network for enhanced semantic information capture.
• Implementation of a multi-scale feature interaction guidance framework with an embedded fusion module to manage spatial and channel complexities.

Main Results:

• The proposed network effectively extracts semantic information from low-resolution maps and spatial information from high-resolution maps.
• Achieved 82.33% mIoU accuracy and 71.26 FPS on the ISIC2017 dataset (256x256 images).
• The network is exceptionally lightweight, containing only 0.524 million parameters, reducing computational load and memory requirements.

Conclusions:

• The proposed lightweight network demonstrates effectiveness and efficiency for medical image segmentation tasks.
• The model successfully balances segmentation performance with reduced computational cost, offering a practical solution for real-world applications.
• The availability of source code facilitates further research and development in lightweight medical image analysis.