Common Pattern Prior-Driven Semi-Supervised Medical Image Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Semi-supervised learning (SSL) is crucial for medical image segmentation due to annotation scarcity.
• Existing SSL methods face challenges with feature exploration and pseudo-label noise, impacting training stability.
• High-quality and stable model learning is essential for reliable medical image segmentation.

Purpose of the Study:

• To propose a novel Common Pattern Prior-driven Network (CPP-Net) for robust semi-supervised medical image segmentation.
• To enhance feature learning quality by extracting core semantic information using a dynamically updated pattern bank.
• To improve training stability by adaptively modulating pseudo-label confidence.

Main Methods:

• Implemented a pattern learning mechanism with a dynamically updated Common Pattern Bank (CP-Bank) for class-specific pattern extraction and reuse.
• Introduced an information gain-driven update strategy for the CP-Bank to ensure alignment with historical pattern distributions.
• Developed a dynamic regulation function to adaptively control pseudo-label impact based on confidence levels.

Main Results:

• CPP-Net demonstrated superior performance in semi-supervised medical image segmentation across various 2D and 3D datasets.
• The pattern learning mechanism improved feature robustness and discriminability, reducing redundant learning.
• The dynamic regulation function effectively mitigated the negative effects of low-confidence pseudo-labels, enhancing training stability.
• Achieved a 7.5% mean Dice improvement over state-of-the-art methods.

Conclusions:

• CPP-Net offers a high-quality and stable approach for semi-supervised medical image segmentation.
• The proposed pattern learning and dynamic regulation strategies effectively address limitations of existing SSL methods.
• CPP-Net shows significant potential for advancing medical image analysis through improved segmentation accuracy and generalizability.