LM-Net: A light-weight and multi-scale network for medical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Current medical image segmentation methods struggle with multi-scale information and integrating local textures with global context.
• This leads to common issues like over-segmentation, under-segmentation, and blurred boundaries.

Purpose of the Study:

• To propose a novel, lightweight, and multi-scale architecture (LM-Net) for enhanced medical image segmentation accuracy.
• To address limitations in capturing multi-scale features and combining local and global information.

Main Methods:

• Developed LM-Net, integrating Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).
• Employed a lightweight multi-branch module for same-level multi-scale feature capture.
• Introduced Local Feature Transformer (LFT) and Global Feature Transformer (GFT) modules for concurrent local texture and global semantic capture using multi-scale features.

Main Results:

• LM-Net achieved state-of-the-art results on three diverse medical image datasets.
• The model demonstrated superior performance compared to existing methods.
• Achieved high accuracy with a lightweight design (4.66G FLOPs, 5.4M parameters).

Conclusions:

• LM-Net effectively integrates local and global representations, mitigating blurred segmentation boundaries.
• The proposed architecture shows significant effectiveness and adaptability across various medical image segmentation tasks.
• LM-Net offers a promising solution for accurate and efficient medical image segmentation.