CPFNet: Context Pyramid Fusion Network for Medical Image Segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning for medical imaging

Background:

• Accurate medical image segmentation is vital for diagnosis and analysis.
• Convolutional Neural Network (CNN) U-shape structures show promise but struggle with context extraction, class imbalance, and blurred boundaries.
• Existing methods have limitations in capturing comprehensive contextual information.

Purpose of the Study:

• To propose a novel Context Pyramid Fusion Network (CPFNet) for enhanced medical image segmentation.
• To address limitations in context information extraction within U-shape CNN architectures.
• To improve segmentation accuracy by effectively fusing global and multi-scale context.

Main Methods:

• Developed a novel Context Pyramid Fusion Network (CPFNet) integrating two pyramidal modules.
• Introduced multiple global pyramid guidance (GPG) modules to reconstruct skip-connections for varied global context.
• Designed a scale-aware pyramid fusion (SAPF) module for dynamic fusion of multi-scale context in high-level features.

Main Results:

• CPFNet demonstrated competitive performance against state-of-the-art methods.
• The network achieved strong results on four challenging segmentation tasks.
• Successfully improved segmentation in skin lesions, retinal lesions, and thoracic organs.

Conclusions:

• The proposed CPFNet effectively fuses global and multi-scale context information.
• CPFNet offers a significant advancement in medical image segmentation accuracy.
• The method shows broad applicability across diverse medical imaging challenges.