Meta semi-supervised medical image segmentation with label hierarchy.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) is crucial for medical image segmentation, leveraging unlabeled data.
• Current SSL methods often use perturbation-based consistency regularization, which can suffer from noisy targets and a knowledge gap.
• This gap exists between supervised guidance and unsupervised regularization objectives.

Purpose of the Study:

• To propose a meta-based semi-supervised segmentation framework to bridge the knowledge gap in medical imaging.
• To improve the utilization of unlabeled data in medical image segmentation tasks.
• To enhance the robustness and accuracy of segmentation models.

Main Methods:

• A novel meta-based semi-supervised segmentation framework is proposed.
• Key components include 'Divide and Generalize' and 'Label Hierarchy' exploitation.
• Domain generalization with a meta-optimization objective bridges the knowledge gap, and hierarchical consistencies distill noisy targets.

Main Results:

• The proposed framework effectively bridges the knowledge gap between supervised and unsupervised learning.
• Hierarchical consistencies are extracted to alleviate noise in self-predicted targets.
• Achieved new state-of-the-art results on two public medical segmentation benchmarks.

Conclusions:

• The meta-based framework offers a superior approach to semi-supervised medical image segmentation.
• Exploiting label hierarchy and domain generalization enhances model performance.
• The method demonstrates significant improvements over existing SSL techniques.