Asymmetric Co-Training With Decoder-Head Decoupling for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning (SSL) is crucial for medical image segmentation, reducing annotation burden by using unlabeled data.
• Existing co-training methods face challenges like intra-network and inter-network coupling, leading to reduced diversity and confirmation bias, especially for complex cases.
• These limitations hinder the reliability of medical image analysis in clinical practice.

Purpose of the Study:

• To introduce AsyCo, an asymmetric co-training framework designed to mitigate coupling issues in semi-supervised medical image segmentation.
• To improve prediction diversity and training stability by decoupling decoder-head interactions and enforcing hierarchical consistency.
• To enhance the accuracy and reliability of medical image segmentation with minimal annotation.

Main Methods:

• AsyCo employs Asymmetric Decoder Coupling to decouple decoder-head connections, enabling dynamic feature remapping for diverse prediction paths.
• Hierarchical Consistency Regularization is utilized, enforcing consistency across different levels: branch outputs, inter-head predictions, and intermediate representations.
• The framework breaks intra-network coupling and promotes inter-network diversity without requiring additional parameters.

Main Results:

• AsyCo significantly outperforms nine state-of-the-art semi-supervised learning methods on three clinical benchmarks.
• The proposed method demonstrates consistent improvements under limited-label conditions.
• AsyCo effectively reduces confirmation bias and enhances training stability for medical image segmentation.

Conclusions:

• AsyCo offers an effective solution for accurate and reliable medical image segmentation with reduced annotation requirements.
• The framework's ability to mitigate coupling issues enhances its applicability in real-world clinical settings.
• This approach contributes to more dependable medical image analysis by improving segmentation accuracy and robustness.