A Stability-Aware Dual-Head Network with Prototype-Based Consistency for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised semantic segmentation for medical images leverages unlabeled data but faces challenges like intra-class variance and domain distribution misalignment.
• Existing methods often struggle with costly training and achieving both semantic consistency and spatial detail preservation.

Purpose of the Study:

• To propose a novel stability-aware dual-head architecture for semi-supervised medical image segmentation.
• To synergize prototype-based and Fully Convolutional Network (FCN) methodologies for improved segmentation performance.
• To mitigate intra-class variance and class-domain distribution shifts effectively.

Main Methods:

• A stability-aware dual-head architecture integrating prototype-based methods (for feature consistency) and FCN methods (for spatial detail).
• A sample-level stability-aware adaptive augmentation strategy to reduce variance and distribution shifts.
• A certainty-guided fusion process for dynamic pseudo-label refinement.

Main Results:

• Achieved State-Of-The-Art (SOTA) performance on BraTS2019 and LA Heart datasets.
• Demonstrated significant improvements over previous SOTA methods across multiple evaluation metrics.
• Effectively bridged domain gaps and enhanced pseudo-label reliability in medical image analysis.

Conclusions:

• The proposed framework offers a robust solution for semi-supervised medical image segmentation.
• The stability-aware dual-head architecture effectively combines semantic and spatial information for precise segmentation.
• This approach advances the reliability and performance of medical image analysis techniques.