Semi-supervised medical image segmentation with dual-branch mixup-decoupling confidence training.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised medical image segmentation reduces labeling dependency and costs.
• Current methods struggle with diverse regularization for unlabeled data and category imbalance.
• Existing algorithms lack robust semantic representations for comparative learning.

Purpose of the Study:

• To propose a novel semi-supervised medical image segmentation algorithm.
• To address limitations in exploiting unlabeled data and handling category imbalance.
• To improve discriminative semantic information learning for better segmentation.

Main Methods:

• Utilizes dual-branch mixup-decoupling confidence training for dual-stream semantic links.
• Employs bidirectional confidence contrast learning for pixel consistency and distinction.
• Designs methods to alleviate semantic ambiguity and learn key intra-class and inter-class features.

Main Results:

• The proposed algorithm achieves notable segmentation performance on 2D and 3D datasets.
• Demonstrates superior performance compared to recent state-of-the-art algorithms.
• Effectively exploits unlabeled data and mitigates category imbalance issues.

Conclusions:

• The novel semi-supervised approach significantly improves medical image segmentation.
• Dual-stream semantic links and bidirectional contrast learning are effective.
• The algorithm offers a promising direction for medical image analysis with reduced annotation effort.