FSOU-Net: Feature supplement and optimization U-Net for 2D medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation is crucial for clinical diagnosis and treatment.
• The U-Net architecture is widely used but lacks specialized processing for scale-specific semantic features.
• Existing U-Net models do not optimize feature representation across different scales.

Purpose of the Study:

• To enhance the feature expression and segmentation performance of the U-Net model.
• To introduce a novel architecture, the Feature Supplement and Optimization U-Net (FSOU-Net).
• To address limitations in U-Net's handling of multi-scale semantic features.

Main Methods:

• Classified encoder-extracted semantic features into shallow and deep categories.
• Proposed the Shallow Feature Supplement Module (SFSM) for fine-grained shallow feature enhancement.
• Developed the Deep Feature Optimization Module (DFOM) using expansive convolutions and multi-scale feature fusion.

Main Results:

• FSOU-Net demonstrated superior segmentation performance on three public medical image datasets.
• The model achieved higher Dice index scores compared to the baseline U-Net and other advanced models.
• Specific improvements included 0.75% on RITE, 2.3% on Kvasir-SEG, and 0.24% on GlaS datasets.

Conclusions:

• The proposed FSOU-Net significantly improves feature representation capabilities.
• The method enhances overall model performance in medical image segmentation tasks.
• Differentiated processing of semantic features is effective for improving segmentation accuracy.