Collaborative multi-feature extraction and scale-aware semantic information mining for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• U-shaped structures with skip connections are common in medical semantic segmentation.
• Limitations exist in feature map compatibility and information extraction in current models.
• These limitations are particularly problematic for segmenting non-obvious, small, or blurred-edge targets.

Purpose of the Study:

• To design a novel segmentation network addressing limitations in current U-shaped architectures.
• To improve the segmentation of challenging medical targets, including small and blurred-edge lesions.
• To enhance information integration and feature extraction within the segmentation network.

Main Methods:

• A novel Global Context-Aware Network (GCN) was developed.
• Key modules include Multi-feature Collaboration Adaptation (MCA), Scale-Aware Mining (SAM), and Edge-enhanced Pixel Intensity Mapping (Edge-PIM).
• These modules are integrated into a U-shaped structure to improve feature processing and segmentation refinement.

Main Results:

• The GCN achieved strong performance on a newly collected Magnetic Resonance Imaging of Colorectal Cancer Liver Metastases (MRI-CRLM) dataset.
• The method also demonstrated effectiveness on the public ISIC-2018 dataset.
• Comparative analysis showed superior performance over state-of-the-art methods like CPFNet on multiple metrics.

Conclusions:

• The proposed GCN effectively addresses limitations in feature integration and extraction for medical semantic segmentation.
• The network significantly improves segmentation accuracy for non-obvious, small, and blurred-edge targets.
• The Edge-PIM visualization method aids radiologists by enhancing target edge prominence.