Global and Local Feature Reconstruction for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning for Computer Vision

Background:

• Encoder-decoder networks are fundamental for medical image segmentation, but U-Net based methods struggle with capturing long-range dependencies and restoring spatial information from down-sampled features.
• Existing U-Net architectures exhibit limitations in global feature extraction and spatial information recovery, hindering segmentation accuracy.

Purpose of the Study:

• To develop a novel network architecture that effectively addresses the limitations of U-Net in medical image segmentation.
• To enhance the capture of global context features and improve the dynamic up-sampling of feature maps.

Main Methods:

• Propose a Global Feature Reconstruction (GFR) module to extract and reconstruct global context features by connecting feature elements across the entire space.
• Introduce a Local Feature Reconstruction (LFR) module that utilizes low-level feature maps to guide the up-sampling of high-level feature maps, preserving spatial information.
• Integrate GFR modules as skip connections and LFR modules into the decoder path of an encoder-decoder architecture, forming the Global and Local Feature Reconstruction Network (GLFRNet).

Main Results:

• The proposed GLFRNet demonstrates superior performance in capturing long-range dependencies and restoring spatial information compared to standard U-Net architectures.
• Experiments on four diverse medical image segmentation tasks show that GLFRNet achieves state-of-the-art results.

Conclusions:

• The developed Global and Local Feature Reconstruction Network (GLFRNet) effectively overcomes the limitations of traditional U-Net models in medical image segmentation.
• GFR and LFR modules significantly enhance global context understanding and spatial detail preservation, leading to improved segmentation accuracy and state-of-the-art performance.