FDSS-Net: feature enhancement and dual-stream semantic mixture network for polyp segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation is crucial for colorectal cancer early detection.
• Existing methods struggle with polyp variability (shape, size) and low contrast.

Purpose of the Study:

• Introduce FDSS-Net, a novel network architecture to enhance polyp segmentation accuracy in colonoscopy images.
• Address limitations of current segmentation techniques.

Main Methods:

• Developed FDSS-Net with three key modules: Feature Enhancement and Propagation Module (FEPM), Dual-Stream Semantic Mixture (DSSM), and Hierarchical Multi-scale Aggregation and Prediction Module (HMAP).
• FEPM captures multi-scale context using depthwise separable convolutions.
• DSSM aligns features and blends semantics across levels using cross-attention and global context.
• HMAP aggregates features hierarchically with a learnable gate.

Main Results:

• FDSS-Net outperformed 12 state-of-the-art methods on five datasets.
• Achieved a Dice coefficient of 0.8302 and mIoU of 0.7587 on the ETIS-LaribPolypDB dataset.
• Demonstrated superior performance in segmenting polyps with challenging characteristics.

Conclusions:

• FDSS-Net significantly enhances polyp segmentation accuracy in colonoscopy images.
• The proposed architecture shows potential for improving clinical computer-aided diagnosis systems.
• Offers a promising direction for future research in medical image analysis.