Semi-Supervised Medical Image Segmentation Using Cross-Style Consistency With Shape-Aware and Local Context...

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Semi-supervised deep learning methods face challenges in medical image segmentation due to insufficient labeled data.
• Limited data hinders capturing anatomical complexity and variability, impacting clinical applications.

Purpose of the Study:

• To develop a novel semi-supervised segmentation framework for anatomically plausible medical image predictions.
• To effectively utilize unlabeled data for improved segmentation accuracy.

Main Methods:

• A framework with two parallel networks: shape-agnostic and shape-aware, enabling mutual learning.
• Shape-aware network introduces implicit shape guidance; shape-agnostic network uses uncertainty estimation for pseudo-labels.
• Cross-style consistency strategy enriches data and prevents overfitting; a novel loss term enhances local context learning.

Main Results:

• The proposed method outperforms existing semi-supervised segmentation techniques on three medical image datasets.
• The framework demonstrates superior performance in shape perception compared to other methods.
• Achieved anatomically plausible predictions even with limited labeled data.

Conclusions:

• The developed semi-supervised framework effectively addresses data scarcity in medical image segmentation.
• The combination of shape-awareness, uncertainty estimation, and cross-style consistency significantly improves segmentation accuracy and shape fidelity.
• This approach offers a promising solution for real-world clinical applications requiring precise anatomical segmentation.